| 12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148114911501151115211531154115511561157115811591160116111621163116411651166116711681169117011711172117311741175117611771178117911801181118211831184118511861187118811891190119111921193119411951196119711981199120012011202120312041205120612071208120912101211121212131214121512161217121812191220122112221223122412251226122712281229123012311232123312341235123612371238123912401241124212431244124512461247124812491250125112521253125412551256125712581259126012611262126312641265126612671268126912701271127212731274127512761277127812791280128112821283128412851286128712881289129012911292129312941295129612971298129913001301130213031304130513061307130813091310131113121313131413151316131713181319132013211322132313241325132613271328132913301331133213331334133513361337133813391340134113421343134413451346134713481349135013511352135313541355135613571358135913601361136213631364136513661367136813691370137113721373137413751376137713781379138013811382138313841385138613871388138913901391139213931394139513961397139813991400140114021403140414051406140714081409141014111412141314141415141614171418141914201421142214231424142514261427142814291430143114321433143414351436143714381439144014411442144314441445144614471448144914501451145214531454145514561457145814591460146114621463146414651466146714681469147014711472147314741475147614771478147914801481148214831484148514861487148814891490149114921493149414951496149714981499150015011502150315041505150615071508150915101511151215131514151515161517151815191520152115221523152415251526152715281529153015311532153315341535153615371538153915401541154215431544154515461547154815491550155115521553155415551556155715581559156015611562156315641565156615671568156915701571157215731574157515761577157815791580158115821583158415851586158715881589159015911592159315941595159615971598159916001601160216031604160516061607160816091610161116121613161416151616161716181619162016211622162316241625162616271628162916301631163216331634163516361637163816391640164116421643164416451646164716481649165016511652165316541655165616571658165916601661166216631664166516661667166816691670167116721673167416751676167716781679168016811682168316841685168616871688168916901691169216931694169516961697169816991700170117021703170417051706170717081709171017111712171317141715171617171718171917201721172217231724172517261727172817291730173117321733173417351736173717381739174017411742174317441745174617471748174917501751175217531754175517561757175817591760176117621763176417651766176717681769177017711772177317741775177617771778177917801781178217831784178517861787178817891790179117921793179417951796179717981799180018011802180318041805180618071808180918101811181218131814181518161817181818191820182118221823182418251826182718281829183018311832183318341835183618371838183918401841184218431844184518461847184818491850185118521853185418551856185718581859186018611862186318641865186618671868186918701871187218731874187518761877187818791880188118821883188418851886188718881889189018911892189318941895189618971898189919001901190219031904190519061907190819091910191119121913191419151916191719181919192019211922192319241925192619271928192919301931193219331934193519361937193819391940194119421943194419451946194719481949195019511952195319541955195619571958195919601961196219631964196519661967196819691970197119721973197419751976197719781979198019811982198319841985198619871988198919901991199219931994199519961997199819992000200120022003200420052006200720082009201020112012201320142015201620172018201920202021202220232024202520262027202820292030203120322033203420352036203720382039204020412042204320442045204620472048204920502051205220532054205520562057205820592060206120622063206420652066206720682069207020712072207320742075207620772078207920802081208220832084208520862087208820892090209120922093209420952096209720982099210021012102210321042105210621072108210921102111211221132114211521162117211821192120212121222123212421252126212721282129213021312132213321342135213621372138213921402141214221432144214521462147214821492150215121522153215421552156215721582159216021612162216321642165216621672168216921702171217221732174217521762177217821792180218121822183218421852186218721882189219021912192219321942195219621972198219922002201220222032204220522062207220822092210221122122213221422152216221722182219222022212222222322242225222622272228222922302231223222332234223522362237223822392240224122422243224422452246224722482249225022512252225322542255225622572258225922602261226222632264226522662267226822692270227122722273227422752276227722782279228022812282228322842285228622872288228922902291229222932294229522962297229822992300230123022303230423052306230723082309231023112312231323142315231623172318231923202321232223232324232523262327232823292330233123322333233423352336233723382339234023412342234323442345234623472348234923502351235223532354235523562357235823592360236123622363236423652366236723682369237023712372237323742375237623772378237923802381238223832384238523862387238823892390239123922393239423952396239723982399240024012402240324042405240624072408240924102411241224132414241524162417241824192420242124222423242424252426242724282429243024312432243324342435243624372438 |
- # Hermes Agent CLI Configuration
- # Copy settings from this example into ~/.hermes/config.yaml, or use
- # `hermes config set <section.key> <value>` to update the active profile.
- # This file configures CLI behavior; only documented secret environment
- # variables in .env take precedence over their corresponding settings.
- # Schema version of this file. The installers copy it to seed config.yaml, and
- # `hermes update` uses it to know which one-time migrations the file already
- # has. Hermes manages it: do not copy it into another config.
- _config_version: 52
- # =============================================================================
- # Database Configuration
- # =============================================================================
- # WAL is the normal default. Hermes automatically falls back to DELETE when
- # SQLite reports that WAL is incompatible with the filesystem. Set this to
- # "delete" explicitly for deployments whose backing filesystem is not WAL
- # crash-safe, such as Linux containers bind-mounted through macOS virtiofs,
- # NFS, or SMB. Hermes will not live-downgrade a database already open in WAL.
- database:
- journal_mode: "wal" # Supported values: "wal", "delete"
- # Optional WAL sizing pragmas (integers). Unset = SQLite defaults.
- # wal_autocheckpoint: 1000 # pages between automatic checkpoints
- # journal_size_limit: 67108864 # cap the WAL/journal file size in bytes
- #
- # Durability level for every state.db connection: OFF, NORMAL, FULL, EXTRA
- # (or 0-3). Unset leaves SQLite's default, which is baked in at compile time
- # (SQLITE_DEFAULT_WAL_SYNCHRONOUS) and therefore differs between the bundled
- # interpreter, a distro python3 and a Homebrew one. Set it if you need to
- # know which one you are running rather than infer it. On macOS this is a
- # floor, not a pin: values below FULL are refused because Darwin's fsync()
- # does not guarantee write ordering, while EXTRA is honored normally.
- # synchronous: FULL
- # =============================================================================
- # Runtime Limits
- # =============================================================================
- # Long-running Hermes server processes raise their RLIMIT_NOFILE soft limit to
- # this value when the operating system permits it. The value is clamped to the
- # hard limit and never lowers an already higher soft limit. Set to 0, false, or
- # null to disable the adjustment. Default: 4096.
- #
- # malloc_arena_max caps glibc malloc arenas in the gateway / `hermes serve`
- # process (Linux/glibc only; child commands keep glibc's default). An explicit
- # MALLOC_ARENA_MAX (exported or in .env) wins. 0, false, or null = glibc default
- # (8 per CPU core). Default: 4.
- runtime:
- nofile_soft_limit: 4096
- malloc_arena_max: 4
- # =============================================================================
- # Attachments
- # =============================================================================
- # Where the TUI/Desktop gateway stages session file attachments (file uploads and
- # pasted text). Read per profile: a session bound to profile X follows X's own
- # config.yaml, so each agent's attachments stay with its workspace.
- # hermes-home (default) <profile home>/attachments — the directory container
- # backends bind-mount, so @file: refs resolve inside the sandbox.
- # workspace <session workspace>/.hermes/attachments — staging lands inside
- # the allowed reference root, so the profile's agent can always
- # read its own attachments back. A remote (ssh) workspace keeps
- # the profile home directory either way.
- attachments:
- storage: "hermes-home"
- # =============================================================================
- # Plugin Installation
- # =============================================================================
- plugins:
- # Deadline for each Git clone, fetch or checkout during plugin install. Default: 300
- # seconds. Subdirectory installs download only that folder.
- # Raise this for large repositories or slow connections. Maximum: 3600.
- clone_timeout_seconds: 300
- # =============================================================================
- # Model Configuration
- # =============================================================================
- model:
- # Default model to use (can be overridden with --model flag)
- # Both "default" and "model" work as the key name here.
- default: "anthropic/claude-opus-4.6"
-
- # Inference provider selection:
- # "auto" - Auto-detect from credentials (default)
- # "openrouter" - OpenRouter (requires: OPENROUTER_API_KEY or OPENAI_API_KEY)
- # "nous" - Nous Portal OAuth (requires: hermes auth add nous)
- # "nous-api" - Nous Portal API key (requires: NOUS_API_KEY)
- # "anthropic" - Direct Anthropic API (requires: ANTHROPIC_API_KEY)
- # "openai-codex" - OpenAI Codex (requires: hermes auth)
- # "gemini" - Use Google AI Studio direct (requires: GOOGLE_API_KEY or GEMINI_API_KEY)
- # "zai" - Use z.ai / ZhipuAI GLM models (requires: GLM_API_KEY)
- # "kimi-coding" - Kimi / Moonshot AI (requires: KIMI_API_KEY)
- # "minimax" - MiniMax global (requires: MINIMAX_API_KEY)
- # "minimax-cn" - MiniMax China (requires: MINIMAX_CN_API_KEY)
- # "huggingface" - Hugging Face Inference (requires: HF_TOKEN)
- # "nvidia" - NVIDIA NIM / build.nvidia.com (requires: NVIDIA_API_KEY)
- # "xiaomi" - Xiaomi MiMo (requires: XIAOMI_API_KEY)
- # "arcee" - Arcee AI Trinity models (requires: ARCEEAI_API_KEY)
- # "ollama-cloud" - Ollama Cloud (requires: OLLAMA_API_KEY — https://ollama.com/settings)
- # "deepinfra" - DeepInfra (requires: DEEPINFRA_API_KEY)
- # "kilocode" - KiloCode gateway (requires: KILOCODE_API_KEY)
- # "ai-gateway" - Vercel AI Gateway (requires: AI_GATEWAY_API_KEY)
- # "azure-foundry" - Microsoft Foundry / Azure OpenAI (API key or Entra ID)
- # "lmstudio" - LM Studio local server (optional: LM_API_KEY, defaults to http://127.0.0.1:1234/v1)
- #
- # Local servers (LM Studio, Ollama, vLLM, llama.cpp):
- # "custom" - Any other OpenAI-compatible endpoint. Set base_url below.
- # Aliases: "ollama", "vllm", "llamacpp" all map to "custom".
- # LM Studio is first-class and uses provider: "lmstudio".
- # It works with both no-auth and auth-enabled server modes.
- #
- # Can also be overridden for a single invocation with the --provider flag.
- provider: "auto"
-
- # API configuration (falls back to OPENROUTER_API_KEY env var)
- # api_key: "your-key-here" # Uncomment to set here instead of .env
- base_url: "https://openrouter.ai/api/v1"
- # Stream API responses from the provider (default: true). The agent core
- # prefers streaming for every turn — subagents included — for liveness
- # health-checking. Set false to force non-streaming requests for the whole
- # session (persists across mid-session model switches). Escape hatch for
- # self-hosted OpenAI-compatible servers whose streaming tool-call path is
- # broken (e.g. vLLM with --tool-call-parser qwen3_xml + a reasoning parser
- # can leak tool calls into plain text instead of returning tool_calls —
- # #72901). Orthogonal to display.streaming, which controls token rendering
- # only.
- # streaming: true
- # Azure Foundry keyless auth example:
- # provider: "azure-foundry"
- # base_url: "https://<resource>.openai.azure.com/openai/v1"
- # auth_mode: "entra_id" # DefaultAzureCredential: az login, managed identity, workload identity, etc.
- # default: "gpt-4o" # Deployment/model name
- # entra:
- # scope: "https://ai.azure.com/.default" # Optional; this is the default.
- # ── Token limits — two settings, easy to confuse ──────────────────────────
- #
- # context_length: TOTAL context window (input + output tokens combined).
- # Controls when Hermes compresses history and validates requests.
- # Leave unset — Hermes auto-detects the correct value from the provider.
- # Set manually only when auto-detection is wrong (e.g. a local server with
- # a custom num_ctx, or a proxy that doesn't expose /v1/models).
- #
- # context_length: 131072
- #
- # ollama_num_ctx: Ollama only — the num_ctx sent on every chat request.
- # Hermes auto-detects the model's window and sends it (Ollama otherwise
- # defaults to 2048); context_length, if set, caps the detected value.
- # An explicit ollama_num_ctx is sent as-is (never capped) — set it to
- # pin VRAM use.
- #
- # ollama_num_ctx: 32768
- #
- # Output-token limits are provider-owned, not user configuration. Native
- # protocols requiring a limit receive an internal value from Hermes.
- # ── Custom request headers (optional) ─────────────────────────────────────
- #
- # default_headers: extra HTTP headers sent on every request to an
- # OpenAI-compatible endpoint. User values take precedence over the
- # provider/SDK defaults, so this is the supported way to override the
- # OpenAI Python SDK's identifying headers (User-Agent: OpenAI/Python ...,
- # X-Stainless-*) when a custom provider sits behind a gateway/WAF that
- # rejects them — e.g. an upstream that returns "502 Upstream access
- # forbidden" for the SDK default User-Agent but accepts a plain one.
- # Applies on the OpenAI wire only (not native Anthropic / Bedrock).
- #
- # default_headers:
- # User-Agent: "curl/8.7.1"
- #
- # extra_headers: accepted as an alias of default_headers (merged, with
- # extra_headers winning when both are set) — matches the per-provider
- # extra_headers key below.
- #
- # Per-provider variant: named providers / custom_providers entries accept an
- # extra_headers dict scoped to that endpoint only — for reverse proxies,
- # gateways, or custom auth (e.g. Cloudflare Access service tokens).
- # Merged onto SDK/provider defaults with the entry's values winning.
- # Header values are treated as secrets and are never logged.
- #
- # providers:
- # my-proxy:
- # base_url: "https://llm.internal.example.com/v1"
- # key_env: "MY_PROXY_API_KEY"
- # extra_headers:
- # CF-Access-Client-Id: "xxxx.access"
- # CF-Access-Client-Secret: "${CF_ACCESS_SECRET}"
- # X-Client-Name: "hermes-agent"
- #
- # session_affinity_header: NAME of a header that carries Hermes' conversation
- # id on every request to that provider (all api_modes + auxiliary calls). Off
- # unless set. For session-aware proxies that otherwise treat each agent-loop
- # request as a new conversation and replay the whole history upstream.
- #
- # providers:
- # my-proxy:
- # base_url: "http://127.0.0.1:4000/v1"
- # session_affinity_header: x-litellm-session-id
- # providers:
- # meta:
- # base_url: https://api.meta.ai/v1
- # api_key: ${MODEL_API_KEY}
- # # api_mode auto-detected as codex_responses for api.meta.ai; no need to set
- # # (the bundled meta-ai provider covers this — a named custom provider is
- # # only needed for a non-default Meta-compatible endpoint)
- # Fast/priority tier behind an OpenAI-compatible gateway or proxy: agent.service_tier
- # (`/fast`) only reaches first-party endpoints; ask a gateway for its own tier via
- # extra_body, which is merged into every request routed to that provider.
- # providers:
- # my-gateway:
- # base_url: https://gateway.example.com/v1
- # api_key: ${MY_GATEWAY_KEY}
- # extra_body:
- # service_tier: priority # whatever tier value your gateway documents
- # providers:
- # router:
- # base_url: https://api.router.com/v1
- # api_key: ${RAMP_ROUTER_API_KEY}
- # # api_mode auto-detected as codex_responses for api.router.com — the
- # # Responses API is Router's native wire (chat/completions is only a
- # # compatibility shim). The bundled router provider covers this; a named
- # # custom provider is only needed for a non-default Router-compatible
- # # endpoint.
- # Command-minted credentials (optional): key_cmd
- # ------------------------------------------------------------------
- # Enterprise gateways often issue SHORT-LIVED bearers (SSO/OIDC brokers, cloud
- # IAM, internal auth proxies) rather than static API keys, so a value copied
- # into .env via `key_env` is stale within the hour and every later request 401s.
- # `key_cmd` names a command that PRINTS a token instead: Hermes runs it per
- # request (cached until shortly before expiry), so long sessions keep working
- # with no restart.
- #
- # Contract: print ONLY the token on stdout, either bare or as JSON with an
- # "access_token" field ("expires_in" is honoured). Same shape as OAuth 2.0
- # token endpoints, Claude Code's `apiKeyHelper`, `gcloud auth
- # print-access-token`, and `aws ecr get-login-password`.
- #
- # Precedence: an explicit --api-key still wins; otherwise key_cmd is preferred
- # over inline api_key / key_env on that entry.
- #
- # Applies to the main agent turn and to auxiliary tasks (title generation,
- # context compression, vision, embedding) alike.
- #
- # Not to be confused with `secrets.command`, which is a different mechanism:
- # that one runs a helper ONCE at startup to populate env vars for many secrets
- # at the process level. Use it for a vault or keychain helper that hands back a
- # KEY=VALUE blob. Use `key_cmd` when ONE provider needs a credential refreshed
- # DURING a session, because a startup-time env var cannot be re-minted after it
- # expires.
- #
- # providers:
- # my-gateway:
- # base_url: "https://gateway.internal.example.com/v1"
- # api_mode: chat_completions
- # key_cmd: "my-auth-cli print-token --profile prod"
- #
- # Worked example — an AI gateway that routes by model family, so one entry per
- # wire format shares the same credential helper:
- #
- # providers:
- # dbx: # OpenAI-compatible (MLflow) route
- # base_url: "https://<workspace>.cloud.databricks.com/ai-gateway/mlflow/v1"
- # api_mode: chat_completions
- # model: databricks-claude-sonnet-4-6
- # key_cmd: "databricks auth token -p MY-PROFILE"
- # dbx-gpt: # OpenAI Responses route
- # base_url: "https://<workspace>.cloud.databricks.com/ai-gateway/openai/v1"
- # api_mode: codex_responses
- # model: databricks-gpt-5-5
- # key_cmd: "databricks auth token -p MY-PROFILE"
- # dbx-claude: # Anthropic Messages route
- # base_url: "https://<workspace>.cloud.databricks.com/ai-gateway/anthropic"
- # api_mode: anthropic_messages
- # model: databricks-claude-fable-5
- # key_cmd: "databricks auth token -p MY-PROFILE"
- # Named provider overrides (optional)
- # Use this for per-provider request timeouts, non-stream stale timeouts,
- # and per-model exceptions.
- # Applies to the primary turn client on every api_mode (OpenAI-wire, native
- # Anthropic, and Anthropic-compatible providers), the fallback chain, and
- # client rebuilds during credential rotation. For OpenAI-wire chat
- # completions (streaming and non-streaming) the configured value is also
- # used as the per-request ``timeout=`` kwarg so it wins over the legacy
- # HERMES_API_TIMEOUT env var (which still applies when no config is set).
- # ``stale_timeout_seconds`` controls the non-streaming stale-call detector and
- # wins over the legacy HERMES_API_CALL_STALE_TIMEOUT env var. Leaving these
- # unset keeps the legacy defaults (HERMES_API_TIMEOUT=1800s,
- # HERMES_API_CALL_STALE_TIMEOUT=90s, native Anthropic 900s). The
- # implicit non-stream stale detector is auto-disabled for local endpoints
- # and can scale upward for very large contexts.
- #
- # AWS Bedrock (providers.bedrock): the resolved value is the socket read
- # timeout of the boto3 Converse client and of the AnthropicBedrock SDK
- # client, chosen per request for the active model. Connect stays fixed
- # (botocore 60s / SDK 10s); unset keeps botocore's 60s read on Converse.
- #
- # providers:
- # ollama-local:
- # request_timeout_seconds: 300 # Longer timeout for local cold-starts
- # stale_timeout_seconds: 900 # Explicitly re-enable stale detection on local endpoints
- # anthropic:
- # request_timeout_seconds: 30 # Fast-fail cloud requests
- # models:
- # claude-opus-4.6:
- # timeout_seconds: 600 # Longer timeout for extended-thinking Opus calls
- # openai-codex:
- # models:
- # gpt-5.4:
- # stale_timeout_seconds: 1800 # Longer non-stream stale timeout for slow large-context turns
- # =============================================================================
- # Unified Timeouts (operation deadlines)
- # =============================================================================
- # One place to override Hermes's internal operation deadlines (seconds).
- # Keys are dotted paths resolved by agent/deadline.py:resolve_timeout().
- # Precedence: this section > legacy HERMES_* env var (back-compat) > built-in
- # default. 0 or a negative value disables the bound (unbounded); very large
- # values are clamped to a platform-safe maximum automatically.
- #
- # Currently resolved keys (more paths migrate here over time — see issue #85125):
- #
- # timeouts:
- # tools:
- # concurrent_batch: 420 # Deadline for a parallel tool-call batch
- # # (legacy env: HERMES_CONCURRENT_TOOL_TIMEOUT_S)
- # sequential_call: 420 # Deadline for one sequentially-executed tool call.
- # # Defaults to concurrent_batch's value so the two
- # # executor paths stay in sync; human waits
- # # (approval prompts, clarify) never count against it.
- # =============================================================================
- # OpenRouter Provider Routing (only applies when using OpenRouter)
- # =============================================================================
- # Control how requests are routed across providers on OpenRouter.
- # See: https://openrouter.ai/docs/guides/routing/provider-selection
- #
- # provider_routing:
- # # Sort strategy: "price" (default), "throughput", or "latency"
- # # Append :nitro to model name for a shortcut to throughput sorting.
- # sort: "throughput"
- #
- # # Only allow these providers (provider slugs from OpenRouter)
- # # only: ["anthropic", "google"]
- #
- # # Skip these providers entirely
- # # ignore: ["deepinfra", "fireworks"]
- #
- # # Try providers in this order (overrides default load balancing)
- # # order: ["anthropic", "google", "together"]
- #
- # # Require providers to support all parameters in your request
- # # require_parameters: true
- #
- # # Data policy: "allow" (default) or "deny" to exclude providers that may store data
- # # data_collection: "deny"
- #
- # # Per-model overrides: same keys, applied only when the agent is on that model
- # # (spelling-tolerant match; unset keys fall through to the flat values above).
- # # models:
- # # "openai/gpt-6-astra":
- # # only: ["openai"]
- # # "anthropic/claude-fable-5.1":
- # # only: ["anthropic"]
- # =============================================================================
- # OpenRouter Response Caching (only applies when using OpenRouter)
- # =============================================================================
- # Cache identical API responses at the OpenRouter edge for free instant replays.
- # When enabled, identical requests (same model, messages, parameters) return
- # cached responses with zero billing. Separate from Anthropic prompt caching.
- # See: https://openrouter.ai/docs/guides/features/response-caching
- #
- # openrouter:
- # response_cache: true # Enable response caching (default: true)
- # response_cache_ttl: 300 # Cache TTL in seconds, 1-86400 (default: 300)
- # =============================================================================
- # Git Worktree Isolation
- # =============================================================================
- # When enabled, each CLI session creates an isolated git worktree so multiple
- # agents can work on the same repo concurrently without file collisions.
- # Equivalent to always passing --worktree / -w on the command line.
- #
- # worktree: true # Always create a worktree when in a git repo
- # worktree: false # Default — only create when -w flag is passed
- #
- # By default a new worktree branches from the freshly-fetched remote tip
- # (the current branch's upstream, else the remote's default branch) so it
- # starts current with the project instead of from the local clone's
- # (possibly stale) HEAD. Set worktree_sync: false to branch from local HEAD
- # instead — useful when offline or when you deliberately want the clone's
- # exact current state as the base.
- #
- # worktree_sync: true # Default — branch from the fetched remote tip
- # worktree_sync: false # Branch from local HEAD (offline / pinned base)
- # Stale worktree sweep (startup of `hermes -w` + every 6h from the cron ticker, over
- # every repo Hermes created a worktree in). Idle dirty/unpushed trees are archived to
- # refs/hermes/archive/<tree> and removed; `hermes worktree restore <name>` brings one back.
- # worktrees:
- # archive_dirty: true # false = never remove dirty/unpushed trees (old behaviour)
- # archive_max_file_mb: 50 # untracked files above this are not archived; their tree is kept
- # archive_retention_days: 30 # archive refs older than this are deleted (0 = keep)
- # max_total_gb: 0 # optional budget for idle trees; oldest evicted first
- # =============================================================================
- # Kanban Review Dispatch
- # =============================================================================
- # First-class review tasks are dispatched automatically by default. The worker
- # is spawned with the bundled sdlc-review skill and can approve, request changes
- # back to the original implementer, or escalate a genuine external blocker.
- # Disable this only when every review is performed manually from the dashboard.
- kanban:
- review_dispatch: true
- # =============================================================================
- # Terminal Tool Configuration
- # =============================================================================
- # Choose ONE of the following terminal configurations by uncommenting it.
- # The terminal tool executes commands in the specified environment.
- # -----------------------------------------------------------------------------
- # OPTION 1: Local execution (default)
- # Commands run directly on your machine in the current directory
- # -----------------------------------------------------------------------------
- # Working directory behavior:
- # - CLI (`hermes` command): Uses "." (current directory where you run hermes)
- # - Gateway/messaging/cron: Uses terminal.cwd here; legacy .env cwd values are deprecated
- cron:
- catch_up_missed: true # False skips past-grace recurring misses after planned downtime.
- terminal:
- backend: "local"
- cwd: "." # For local backend: "." = current directory. Ignored for remote backends unless a backend documents otherwise.
- # Desktop xterm font. Install the font locally; Nerd Fonts render Powerlevel10k glyphs.
- # font_family: "MesloLGS NF" # Also accepts a CSS stack; blank uses bundled JetBrains Mono.
- timeout: 180
- # HOME policy for tool subprocesses:
- # auto - default: host uses your real HOME; containers use HERMES_HOME/home
- # real - force your real OS-user HOME
- # profile - force HERMES_HOME/home for strict per-profile CLI config isolation
- home_mode: "auto"
- docker_mount_cwd_to_workspace: false # SECURITY: off by default. Opt in to mount the launch cwd into Docker /workspace.
- lifetime_seconds: 300
- # Two process_manage wait/poll calls in a row with no new output on a running background process
- # turn on notify_on_complete for it and tell the agent to stop checking; false = poll as before.
- process_wait_auto_notify: true
- # A `hermes` / `hermes --tui` session's local tools (terminal, background processes, execute_code)
- # start with the launching shell's toolchain: PATH, an activated venv and CONDA_* / NODE_* / NVM_* /
- # PYENV_* (plus GOPATH, GOROOT, CARGO_HOME, RUSTUP_HOME, JAVA_HOME), for that session only, when
- # the client runs on the gateway's host (a remote client's session keeps the gateway's toolchain).
- # As in the classic CLI, the venv reaches `terminal` through PATH (`which python`): VIRTUAL_ENV and
- # CONDA_PREFIX themselves are unset there; execute_code sees them.
- # A toolchain name the client did not export is unset for its session's tools (PATH excepted).
- # List more non-secret names here (default: none); names execute_code treats as credentials (…TOKEN,
- # …KEY, …SECRET, …PASSWORD, …AUTH, …DSN, …WEBHOOK, …CREDS, …BEARER, …COOKIE) and HERMES_* names
- # are never sent: pass secrets with `env_passthrough` instead.
- client_env_passthrough: [] # e.g. ["GEM_HOME", "POETRY_HOME"]
- # sudo_password: "hunter2" # Optional: pipe a sudo password via sudo -S. SECURITY WARNING: plaintext.
- # sudo_password: "" # Explicit empty password: try empty and never open the interactive sudo prompt.
- # -----------------------------------------------------------------------------
- # OPTION 2: SSH remote execution
- # Commands run on a remote server - agent code stays local (sandboxed)
- # Great for: keeping agent isolated from its own code, using powerful remote hardware
- # -----------------------------------------------------------------------------
- # terminal:
- # backend: "ssh"
- # cwd: "/home/myuser/project" # Path on the REMOTE server
- # timeout: 180
- # lifetime_seconds: 300
- # ssh_host: "my-server.example.com"
- # ssh_user: "myuser"
- # ssh_port: 22
- # ssh_key: "~/.ssh/id_rsa" # Optional - uses ssh-agent if not specified
- # -----------------------------------------------------------------------------
- # OPTION 3: Docker container
- # Commands run in an isolated Docker container
- # Great for: reproducible environments, testing, isolation
- # -----------------------------------------------------------------------------
- # terminal:
- # backend: "docker"
- # cwd: "/workspace" # Path INSIDE the container (default: /)
- # timeout: 180
- # lifetime_seconds: 300
- # docker_image: "nousresearch/hermes-sandbox:desktop"
- # docker_mount_cwd_to_workspace: true # Explicit opt-in: mount your launch cwd into /workspace
- # # Optional: run the container as your host user's uid:gid so files written
- # # into bind-mounted dirs are owned by you, not root. Drops SETUID/SETGID
- # # caps too since no gosu privilege drop is needed. Leave off if your
- # # chosen docker_image expects to start as root.
- # docker_run_as_host_user: true
- # # Optional: explicitly forward selected env vars into Docker.
- # # These values come from your current shell first, then ~/.hermes/.env.
- # # Warning: anything forwarded here is visible to commands run in the container.
- # docker_forward_env:
- # - "GITHUB_TOKEN"
- # - "NPM_TOKEN"
- # # Optional: extra flags passed verbatim to docker run (appended after security defaults).
- # # Useful for adding capabilities (e.g. apt installs needing SETUID) or custom options.
- # # Example: add a Linux capability not included by default
- # # docker_extra_args:
- # # - "--cap-add"
- # # - "SETUID"
- # -----------------------------------------------------------------------------
- # OPTION 4: Singularity/Apptainer container (the official `singularity` plugin)
- # Commands run in a Singularity container (common in HPC environments)
- # Great for: HPC clusters, shared compute environments
- # Install it first: hermes plugins install singularity
- # -----------------------------------------------------------------------------
- # terminal:
- # backend: "singularity"
- # cwd: "/workspace" # Path INSIDE the container (default: /root)
- # timeout: 180
- # lifetime_seconds: 300
- # singularity_image: "docker://nousresearch/hermes-sandbox:desktop"
- # -----------------------------------------------------------------------------
- # OPTION 5: Modal cloud execution
- # Commands run on Modal's cloud infrastructure
- # Great for: GPU access, scalable compute, serverless execution
- # -----------------------------------------------------------------------------
- # terminal:
- # backend: "modal"
- # cwd: "/workspace" # Path INSIDE the sandbox (default: /root)
- # timeout: 180
- # lifetime_seconds: 300
- # modal_image: "nousresearch/hermes-sandbox:desktop"
- # -----------------------------------------------------------------------------
- # OPTION 6: Daytona cloud execution
- # Commands run in Daytona cloud sandboxes
- # Great for: Cloud dev environments, persistent workspaces, team collaboration
- # Requires: pip install daytona, DAYTONA_API_KEY env var
- # -----------------------------------------------------------------------------
- # terminal:
- # backend: "daytona"
- # cwd: "~"
- # timeout: 180
- # lifetime_seconds: 300
- # daytona_image: "nousresearch/hermes-sandbox:desktop"
- # container_disk: 10240 # Daytona max is 10GB per sandbox
- #
- # --- Container resource limits (docker, modal, daytona, container plugin backends -- ignored for local/ssh) ---
- # These settings apply to all container backends. They control the resources
- # allocated to the sandbox and whether its filesystem persists across sessions.
- container_cpu: 1 # CPU cores
- container_memory: 5120 # Memory in MB (5120 = 5GB)
- container_disk: 51200 # Disk in MB (51200 = 50GB)
- container_persistent: true # Persist filesystem across sessions (false = ephemeral)
- # -----------------------------------------------------------------------------
- # SUDO SUPPORT (works with ALL backends above)
- # -----------------------------------------------------------------------------
- # Add sudo_password to any terminal config above to enable sudo commands.
- # The password is piped via `sudo -S`. Works with local, ssh, docker, etc.
- #
- # SECURITY WARNING: Password stored in plaintext!
- #
- # INTERACTIVE PROMPT: If sudo_password is unset and the CLI is running,
- # you'll be prompted to enter your password when sudo is needed:
- # - 45-second timeout (auto-skips if no input)
- # - Press Enter to skip (command fails gracefully)
- # - Password is hidden while typing
- # - Password is cached for the session
- #
- # EMPTY PASSWORDS: Setting sudo_password to an explicit empty string is different
- # from leaving it unset. Hermes will try an empty password via `sudo -S` and
- # will not open the interactive prompt. This is useful for passwordless sudo,
- # Touch ID sudo setups, and environments where prompting is just noise.
- #
- # ALTERNATIVES:
- # - SSH backend: Configure passwordless sudo on the remote server
- # - Containers: Run as root inside the container (no sudo needed)
- # - Local: Configure /etc/sudoers for specific commands
- #
- # Example (add to your terminal section):
- # sudo_password: "your-password-here"
- # =============================================================================
- # Security
- # =============================================================================
- # security:
- # # Private/internal addresses (localhost, 192.168.x.x, CGNAT) for web_extract,
- # # vision/media URL fetches and cloud/CDP browsers. The local Chromium and the
- # # hybrid sidecar (browser.auto_local_for_private_urls) are not covered.
- # # Cloud-metadata endpoints stay blocked even when true.
- # allow_private_urls: false
- # approval:
- # transport: builtin # Or an explicitly enabled plugin transport name
- # transport_fallback: deny # Set builtin to opt into fallback on transport failure
- # # Agent writes to instruction files that steer the agent itself
- # # (AGENTS.md, CLAUDE.md, SOUL.md, ...) always ask for human approval,
- # # even under --yolo, and are refused when nobody can answer. Set false to
- # # disable the gate; add fnmatch patterns on the basename to protect more files.
- # protected_instruction_files: true
- # protected_instruction_extra_patterns: []
- # =============================================================================
- # LSP Diagnostics (post-write lint for write_file / patch)
- # =============================================================================
- # lsp:
- # enabled: true
- # # Language servers only load code a project ships (its .venv interpreter,
- # # node_modules TypeScript, svelte.config.js, cargo/Gradle/mix build files)
- # # in the git worktree of the directory you launched Hermes in or opened
- # # the session in (hermes -w, a Desktop project, terminal.cwd), or under a
- # # directory listed here. In other checkouts, such as a repository the agent
- # # cloned, only servers that run no project code start (pinned to
- # # Hermes-side tools) and the npx tsc / rustfmt lint fallbacks are skipped.
- # trusted_workspaces:
- # - ~/code/my-app
- # =============================================================================
- # Browser Tool Configuration
- # =============================================================================
- browser:
- # Inactivity timeout in seconds - browser sessions are automatically closed
- # after this period of no activity between agent loops (default: 120 = 2 minutes)
- inactivity_timeout: 120
- # Local Chromium for browser_exec. Empty executable_path = pm's chromium
- # (`hermes pm install chromium`), the Docker image's build, or a system
- # chromium where pm builds none (Termux). user_data_dir keeps one persistent
- # profile for every local session (empty = a throwaway profile per session).
- # executable_path: ""
- # user_data_dir: ""
- # chromium_args: ["--lang=de"]
- # With a cloud provider (Browserbase, Browser Use, ...), LAN/localhost URLs
- # open in a local Chromium sidecar on this host instead. The sidecar ignores
- # security.allow_private_urls (like the plain local Chromium, it reaches what
- # the terminal reaches); set false to refuse private URLs instead.
- # auto_local_for_private_urls: true
- # =============================================================================
- # Tool Loop Guardrails
- # =============================================================================
- # Soft warnings are enabled by default. They append guidance to repeated failed
- # or non-progressing tool results but still let the tool execute. Hard stops stay
- # opt-in for interactive CLI/TUI/Desktop/ACP sessions, but default on for unattended
- # gateway/cron sessions where nobody is present to interrupt a model that
- # ignores loop warnings.
- tool_loop_guardrails:
- warnings_enabled: true
- hard_stop_enabled: false
- non_interactive_hard_stop_enabled: true
- warn_after:
- exact_failure: 2
- same_tool_failure: 3
- idempotent_no_progress: 2
- hard_stop_after:
- exact_failure: 5
- same_tool_failure: 8
- idempotent_no_progress: 5
- # =============================================================================
- # Context Compression (Auto-shrinks long conversations)
- # =============================================================================
- # When conversation approaches model's context limit, middle turns are
- # automatically summarized to free up space while preserving important context.
- #
- # HOW IT WORKS:
- # 1. Tracks actual token usage from API responses (not estimates)
- # 2. When prompt_tokens >= threshold% of model's context_length, triggers compression
- # 3. Protects first 3 turns (system prompt, initial request, first response)
- # 4. Protects last N turns (default 20 messages = ~10 full turns of recent context)
- # 5. Summarizes middle turns using a fast/cheap model
- # 6. Inserts summary as a user message, continues conversation seamlessly
- #
- # Post-compression tail budget is target_ratio × threshold × context_length:
- # 200K context, threshold 0.50, ratio 0.20 → 20K tokens of recent tail preserved
- # 1M context, threshold 0.50, ratio 0.20 → 100K tokens of recent tail preserved
- #
- compression:
- # Enable automatic context compression (default: true)
- # Set to false if you prefer to manage context manually or want errors on overflow
- enabled: true
- # Fail closed before lossy compaction unless an active memory provider that
- # implements the pre-compress checkpoint contract (API v2) confirms its
- # durable checkpoint (default: false). With this on and no checkpoint, the
- # compaction attempt errors with BLOCKED_MISSING_PREREQUISITE and the
- # uncompressed transcript is preserved for a later retry. The gate binds to
- # every compaction authority: server-side native compaction
- # (codex_responses_native) is suppressed while armed, post-turn
- # micro-compaction is forced off (no checkpoint hook in its path), and the
- # codex_app_server API mode is refused at agent init (the codex agent
- # compacts its own thread — no checkpoint can be guaranteed there). Only
- # enable it with a checkpoint-capable provider configured — see
- # website/docs/developer-guide/memory-provider-plugin.md
- # ("Pre-Compress Checkpoints").
- checkpoint_required: false
- # Opt-in compression progress notices on chat platforms (default: false).
- # By design, routine automatic compression is SILENT on human-facing chat
- # gateways (Telegram, Discord, Slack, ...) — it happens in the background
- # with server-side logging only. Set true to also deliver the routine
- # progress statuses (compacting started, preflight/pre-API compression,
- # idle compaction, retry progress, and the compaction-complete notice) to
- # chat platforms. Unrelated operational noise (auxiliary model failures,
- # provider retry chatter) stays suppressed either way, and compression
- # FAILURE notices + manual /compress feedback are always visible
- # regardless of this setting. (#52995)
- progress_notices: false
-
- # Trigger compression at this % of model's context limit (default: 0.50 = 50%)
- # Lower values = more aggressive compression, higher values = compress later
- # Models with context windows below 512K are floored at 0.75 (raise-only) so
- # compaction doesn't fire with half the window still free; set above 0.75 to override.
- threshold: 0.50
- # Per-model threshold overrides: keys are substring-matched against the model
- # name (longest match wins). Useful when some models need different compaction
- # points — e.g. a 1M-context model can compress later (0.30) while a 128K
- # model needs to compress earlier (0.60). The small-context floor (75% for
- # <512K models) still applies on top of per-model overrides.
- # model_thresholds:
- # "glm-5.2": 0.40
- # "claude-sonnet": 0.35
- # "gpt-5": 0.30
- # Optional absolute token cap for the compression trigger (default: null = disabled).
- # When set, compression fires at the LOWER of the ratio-based threshold and this
- # absolute token count. Clamped to the model's context length at apply-time, so a
- # cap above the window is a no-op. Survives model switches and fallback activations.
- # threshold_tokens: 200000
- # Existing Codex gpt-5.5 behavior: raise Hermes' compaction trigger to 85%
- # for the ChatGPT Codex OAuth route. Set false to opt back down to threshold.
- codex_gpt55_autoraise: true
- # Fraction of the threshold to preserve as recent tail (default: 0.20 = 20%)
- # e.g. 20% of 50% threshold = 10% of total context kept as recent messages.
- # Summary output is separately capped at 12K tokens (Gemini output limit).
- # Range: 0.10 - 0.80
- target_ratio: 0.20
- # Number of most-recent messages to always preserve (default: 20 ≈ 10 full turns)
- # Higher values keep more recent conversation intact at the cost of more aggressive
- # compression of older turns.
- protect_last_n: 20
- # Minimum number of REAL (actionable) user messages guaranteed to survive in
- # the uncompressed tail (default: 1 = the existing single last-user anchor,
- # behavior-preserving). Raise to e.g. 3 to keep the last 3 real user turns
- # verbatim even when bulky tool outputs fill the tail token budget — blank
- # platform echoes, compaction handoffs, and synthetic continuation rows never
- # count toward N. The tail can exceed the token budget when this pulls the
- # cut back; the guarantee wins over the budget by design.
- min_tail_user_messages: 1
- # Compression retry rounds before a turn gives up with "max compression
- # attempts reached" (default: 3, same as the previous hardcoded value).
- # Raise (e.g. 6) for tool-schema-heavy sessions where 3 rounds cannot bring
- # the request estimate under the threshold. Validated >= 1, hard cap 10.
- max_attempts: 3
- # Codex app-server (codex CLI runtime) thread-compaction mode. The codex
- # agent owns the real thread context on this runtime, so Hermes' summarizer
- # cannot shrink it — compaction goes through the app server instead.
- # native = let Codex decide when to compact its own thread (default)
- # hermes = let Hermes threshold trigger Codex thread/compact/start
- # off = Hermes will not auto-trigger compaction; Codex may still compact natively
- codex_app_server_auto: native
- # Native OpenAI Responses server-side compaction (default: false). When true,
- # gpt-5.6-family models on the DIRECT OpenAI API (api.openai.com) or a ChatGPT
- # Codex subscription, plus gpt-6-astra (and its -900k alias) on official Codex OAuth, compact
- # server-side: OpenAI prunes older context into an
- # encrypted checkpoint that Hermes replays on later turns. No other provider,
- # route, or model is affected. Hermes' local compression stays armed as the
- # fallback and still handles every non-eligible session.
- codex_responses_native: false
- # Optional absolute server compaction trigger in input tokens. The default
- # null value follows the resolved local compression trigger with an 8192 token
- # safety margin. For example, a local trigger of 765000 selects 756808.
- # A positive integer stays absolute and only clamps downward when needed so
- # the server compacts first. Invalid values use this automatic behavior. If
- # the local trigger is unavailable, automatic mode uses 200000.
- codex_responses_compact_threshold: null
- # Number of non-system messages to protect at the head of the transcript, in
- # ADDITION to the system prompt (which is always implicitly protected).
- # Head messages are NEVER summarized — they survive every compression
- # indefinitely. This gives stable early context for short/medium sessions,
- # but in long-running sessions that rely on rolling compaction the pinned
- # opening turns may not match how you want the session framed over time.
- # Set to 0 to preserve ONLY the system prompt (plus the rolling summary
- # and recent tail) — the cleanest configuration for long-running sessions.
- # Default 3 preserves the system prompt plus the first three non-system
- # head messages, matching the pre-feature behaviour.
- protect_first_n: 3
- # Idle compaction (default: 0 = disabled). When > 0, a session that resumes
- # after at least this many seconds of inactivity compacts its accumulated
- # history up front, before the first reply, so a long-lived thread you come
- # back to later doesn't re-read its full stale context on every turn.
- # Time-based, so it complements (does not replace) the size-based `threshold`
- # above. It is skipped when the context is already small (at or below the
- # post-compression target = threshold × target_ratio), so it never wastes a
- # summarization on a short idle thread. Example: 1800 = compact after 30 min idle.
- idle_compact_after_seconds: 0
- # Gateway session-hygiene turn-hold budget (default: 10). Max seconds an
- # arriving user turn is held while a still-streaming hygiene summary
- # finishes. Distinct from hygiene_timeout_seconds (compressor inactivity
- # budget): this bounds user-visible latency so chat transports (Telegram
- # ~30s) do not drop a silent connection. On expiry the turn proceeds
- # uncompressed; the detached worker keeps its commit admission (when the
- # commit is watermark-fenced) and the summary is adopted at the next safe
- # boundary. Thinking-model summarizers often need longer than 10s to emit
- # the first content token — raise to 300 (or >= your summarizer's real
- # time-to-first-content) only if you want THIS turn to wait for the
- # compression instead of adopting it one turn late.
- hygiene_max_turn_hold_seconds: 10
- # Proactive tool-result prune (default: 0 = disabled). Opt-in token trigger
- # for a deterministic, no-LLM prune of OLD tool-result payloads, run
- # independently of `threshold` above. On large-window models (512K/1M) the
- # ratio threshold rarely fires, so bulky tool outputs (terminal dumps, file
- # reads, web extracts) ride along in history and get re-billed every turn.
- # When re-sent history exceeds this many tokens, the prune dedupes identical
- # results, summarizes older oversized ones, and truncates large tool-call
- # arguments — protecting the most recent `protect_last_n` messages and never
- # calling the model. Try 48000 to enable. Built-in compressor engine only;
- # other context engines inherit a safe no-op.
- # NOTE: a committed prune rewrites already-sent history, which invalidates
- # the provider's prompt-cache prefix — the min_reclaim gate below keeps
- # those cache breaks episodic (like a compression boundary) instead of
- # per-turn.
- proactive_prune_tokens: 0
- # The prune's summarize pass only touches tool results larger than this many
- # characters (clamped to >= 200 so a generated summary can't be
- # re-summarized). Default 8000.
- proactive_prune_min_result_chars: 8000
- # A proactive prune only COMMITS when it reclaims at least this many tokens
- # (measured on the pruned output). This is the prompt-cache hysteresis gate:
- # one meaningful, amortized cache break per batch of stale tool output
- # instead of a tiny break on every tool iteration. 0 = commit any non-zero
- # prune. Default 4096.
- proactive_prune_min_reclaim_tokens: 4096
- # To pin a specific model/provider for compression summaries, use the
- # auxiliary section below (auxiliary.compression.provider / model).
- # =============================================================================
- # Tool-result budget (optional)
- # =============================================================================
- # Controls when a large tool result is spilled to disk (full output saved to
- # $HERMES_HOME/cache/spillover, preview + path kept in context). MCP tool
- # results (tools named mcp_*) spill at a tighter 50,000-char threshold than
- # the generic 100K default: MCP servers routinely return un-paginated 20-50K
- # payloads that bloat context and slow every subsequent turn. Nothing is
- # lost — the full result is on disk and readable with read_file.
- #
- # tool_budget:
- # mcp_result_size_chars: 50000 # per-result spillover threshold for mcp_* tools
- # =============================================================================
- # Anthropic prompt caching TTL
- # =============================================================================
- # When prompt caching is active (Claude via OpenRouter or native Anthropic),
- # Anthropic supports two TTL tiers for cached prefixes: "5m" (default) and
- # "1h". Other values are ignored and "5m" is used.
- #
- prompt_caching:
- cache_ttl: "5m" # use "1h" for long sessions with pauses between turns
- # =============================================================================
- # Auxiliary Models (Advanced — Experimental)
- # =============================================================================
- # Hermes uses lightweight "auxiliary" models for side tasks: image analysis,
- # browser screenshot analysis, web page summarization, TTS audio-tag insertion,
- # and context compression.
- #
- # By default these use Gemini Flash via OpenRouter or Nous Portal and are
- # auto-detected from your credentials. You do NOT need to change anything
- # here for normal usage.
- #
- # WARNING: Overriding these with providers other than OpenRouter or Nous Portal
- # is EXPERIMENTAL and may not work. Not all models/providers support vision,
- # produce usable summaries, or accept the same API format. Change at your own
- # risk — if things break, reset to "auto" / empty values.
- #
- # Each task has its own provider + model pair so you can mix providers.
- # For example: OpenRouter for vision (needs multimodal), but your main
- # local endpoint for compression (just needs text).
- #
- # Provider options:
- # "auto" - Best available: OpenRouter → Nous Portal → main endpoint (default)
- # "openrouter" - Force OpenRouter (requires OPENROUTER_API_KEY)
- # "nous" - Force Nous Portal (requires: hermes auth add nous)
- # "gemini" - Force Google AI Studio direct (requires: GOOGLE_API_KEY or GEMINI_API_KEY)
- # "ollama-cloud" - Ollama Cloud (requires: OLLAMA_API_KEY)
- # "codex" - Force Codex OAuth (requires: hermes model → Codex).
- # Uses gpt-5.3-codex which supports vision.
- # "main" - Use your custom endpoint (OPENAI_BASE_URL + OPENAI_API_KEY).
- # Works with OpenAI API, local models, or any OpenAI-compatible
- # endpoint. Also falls back to Codex OAuth and API-key providers.
- #
- # Model: leave empty to use the provider's default. When empty, OpenRouter
- # uses "google/gemini-3-flash-preview" and Nous uses "gemini-3-flash".
- # Other providers pick a sensible default automatically.
- #
- # Native vision embeds (vision_analyze / browser screenshots when the MAIN model
- # is vision-capable) ride conversation history and are re-sent on every later
- # API call, so both knobs below bound recurring cost:
- #
- # vision:
- # # Byte budget for one embedded image (clamped 64 KiB..4 MiB). Raise it for
- # # dense phone screenshots of tables that read as "unreadable" at 256 KB.
- # embed_target_bytes: 262144
- # # How often vision_analyze may embed the SAME image (region crops included)
- # # per session. Unset = 3 inside delegated subagents, unlimited for the main
- # # agent; a number applies everywhere; 0 = unlimited.
- # max_calls_per_image: 3
- # # Newest user images re-sent on turns rebuilt from the session DB (API server,
- # # --resume) when the model takes images natively; 0 = off. Older images drop
- # # in whole batches of this size, so the cached prefix changes once per batch.
- # replay_recent_images: 3
- #
- # auxiliary:
- # # Image analysis: vision_analyze tool + browser screenshots
- # vision:
- # provider: "auto"
- # model: "" # e.g. "google/gemini-2.5-flash", "openai/gpt-4o"
- # timeout: 30 # LLM API call timeout (seconds)
- # download_timeout: 30 # Image HTTP download timeout (seconds)
- # # Increase for slow connections or self-hosted image servers
- # reasoning_effort: "" # Per-task thinking level: none, minimal, low, medium,
- # # high, xhigh, max, ultra. Empty = provider default.
- # # Works on every auxiliary task block (vision,
- # # web_extract, compression, title_generation, curator,
- # # background_review, moa_reference, ...). Example: run
- # # compression at "low" and vision at "none" to cut
- # # side-task latency/cost on reasoning models.
- #
- # # Web page scraping / summarization + browser page text extraction
- # web_extract:
- # provider: "auto"
- # model: ""
- # reasoning_effort: "low"
- #
- # # Gemini 3.1 TTS hidden audio-tag insertion
- # tts_audio_tags:
- # provider: "auto" # empty model = your main chat model
- # model: ""
- # timeout: 30
- #
- # # Automatic session title generation after the first exchange
- # title_generation:
- # enabled: true # set false to disable auto-title generation
- # model_upgrade_enabled: true # set false to keep the instant derived title and skip the model call
- # provider: "auto"
- # model: ""
- # timeout: 30
- # language: "" # empty = match the user's language; or e.g. "English"
- #
- # # Session search — summarizes matching past sessions
- # session_search:
- # provider: "auto"
- # model: ""
- # timeout: 30
- # max_concurrency: 3 # Limit parallel summaries to reduce request-burst 429s
- # extra_body: {} # Provider-specific OpenAI-compatible request fields
- # # Example for providers that support request-body
- # # reasoning controls:
- # # extra_body:
- # # enable_thinking: false
- # # Some vLLM/Qwen deployments expect this nested:
- # # extra_body:
- # # chat_template_kwargs:
- # # enable_thinking: false
- #
- # # Auto-generated short session titles after the first exchange.
- # # Each active Discord/Telegram channel can spawn a background title
- # # call. Cap concurrency to keep retries during provider incidents
- # # from amplifying the request burst. Leave unset for legacy behavior
- # # (unlimited).
- # title_generation:
- # provider: "auto"
- # model: ""
- # # max_concurrency: 2 # Optional: cap simultaneous title calls
- #
- # # Context compression — summarizes long sessions to shrink the prompt.
- # # Heavy and often hits the slowest provider chain. Setting a small
- # # cap prevents many sessions from compressing simultaneously during
- # # provider degradation. Leave unset for legacy behavior (unlimited).
- # compression:
- # provider: "auto"
- # model: ""
- # # max_concurrency: 2 # Optional: cap simultaneous compression calls
- # # no_progress_timeout: 60 # Codex/Responses-only: seconds a stream may go without a
- # # substantive event before it fails fast (default: 60).
- # # Independent of "timeout" above (the overall request
- # # budget) — raising "timeout" alone does not widen this
- # # window. Each substantive event re-arms it; a slow but
- # # progressing reasoning/summary stream is not affected.
- #
- # # Post-turn memory/skill self-improvement review fork. Runs after a turn
- # # when the nudge intervals fire; writes skills/memories in a daemon thread.
- # # Usage is recorded under session_model_usage task='background_review'.
- # background_review:
- # enabled: true # false = skip automatic forks (/refine still works)
- # provider: "auto" # or pin a cheaper model (see memory.md)
- # model: ""
- # =============================================================================
- # Persistent Memory
- # =============================================================================
- # Bounded curated memory injected into the system prompt every session.
- # Two stores: MEMORY.md (agent's notes) and USER.md (user profile).
- # Character limits keep the memory small and focused. The agent manages
- # pruning -- when at the limit, it must consolidate or replace entries.
- # Disabled by default in batch_runner.
- #
- memory:
- # Agent's personal notes: environment facts, conventions, things learned
- memory_enabled: true
-
- # User profile: preferences, communication style, expectations
- user_profile_enabled: true
- # External prefetch only: opt in to head/tail previews using hooks.output_spill.
- # Defaults to keeping the full provider-ranked recall. Restart to apply changes.
- prefetch_spill_enabled: false
-
- # Character limits (~2.75 chars per token, model-independent)
- memory_char_limit: 2200 # ~800 tokens
- user_char_limit: 1375 # ~500 tokens
- # Periodic memory nudge: remind the agent to consider saving memories
- # every N user turns. Set to 0 to disable. Only active when memory is enabled.
- nudge_interval: 10 # Nudge every 10 user turns (0 = disabled)
- # =============================================================================
- # Session Continuity (Messaging Platforms)
- # =============================================================================
- # Conversations persist across inactivity and daily boundaries. Use /new or
- # /reset explicitly; context compression manages long conversations.
- # Maximum number of simultaneously active chat sessions across CLI, TUI,
- # dashboard chat, and messaging gateway. Set to null, 0, or omit to allow
- # unlimited concurrent sessions. When the limit is reached, new sessions get a
- # clean error while existing active sessions keep their normal behavior. This
- # top-level key takes precedence over gateway.max_concurrent_sessions. The cap
- # is a best-effort single-host/profile runtime guard; Hermes fails open if the
- # local runtime lease registry cannot be read or locked.
- max_concurrent_sessions: null
- # When true, group/channel chats use one session per participant when the platform
- # provides a user ID. This is the secure default and prevents users in the same
- # room from sharing context, interrupts, and token costs. Set false only if you
- # explicitly want one shared "room brain" per group/channel.
- group_sessions_per_user: true
- # =============================================================================
- # Native Hosted-Room Discussion Limits
- # =============================================================================
- # Shared ceilings per discussion across all members, applied to all native hosted
- # rooms (not Desktop renderer/mirrored rooms). Positive integers only: no 0,
- # unlimited, null, booleans, strings, or floats; an invalid value is logged and
- # the defaults apply. Omitted keys keep these defaults.
- # Set in the root owner's config.yaml beside shared-state.db, not a member's
- # profile config. Read once at startup; refresh every server process sharing the
- # room database to apply (not just the service the client connects to).
- # For a larger discussion budget, use max_rounds: 6 and max_messages: 30.
- hosted_rooms:
- discussion:
- max_rounds: 3
- max_messages: 10 # Total visible bot replies, not per member
- # =============================================================================
- # Session Storage Retention (state.db)
- # =============================================================================
- # ~/.hermes/state.db keeps every session, message, and tool call, plus the
- # FTS5 search indexes. Auto-pruning is OFF by default; when enabled, at
- # CLI/gateway/cron startup (at most once per min_interval_hours) Hermes deletes
- # ENDED sessions whose last activity is older than retention_days. While
- # auto_prune is unset, Hermes asks once when the disk holding state.db runs
- # low (/autoprune on|off|later answers it). Open, pinned, and in-progress sessions are never
- # deleted. Stale automation sessions (cron/kanban/subagent/one-shot CLI) whose
- # process died without closing them are first *closed*, then aged through a
- # further full retention window before removal.
- #
- # After a prune that removed rows, VACUUM reclaims disk space only when both
- # the time throttle (min_vacuum_interval_days) has elapsed AND more than 25%
- # of the file's pages are reclaimable — a dense database never pays for a full
- # rewrite to reclaim a few MB.
- #
- # A CLI/TUI/Desktop session left with no viewer (a killed terminal, a dropped
- # connection, a gateway restart) is ended by the gateway orphan_end_grace_s
- # seconds later, once its running turn settles. It stays resumable, and a viewer
- # that comes back within the grace keeps it open. 0 never ends it. After a
- # restart it also waits out the rest of 6h since the session's last activity.
- # It replaces dashboard.ws_orphan_reap_grace_s / startup_orphan_sweep for these
- # sessions; while it is unset, their opt-outs (0 / false) still turn it off.
- #
- # Uncomment to change the defaults shown; auto_prune: true bounds the file,
- # auto_prune: false keeps every ended session and silences the low-disk prompt.
- # sessions:
- # auto_prune: false
- # retention_days: 90
- # vacuum_after_prune: true
- # min_vacuum_interval_days: 30
- # min_interval_hours: 24
- # orphan_end_grace_s: 300
- # Optional direct endpoint for autonomous Bot Mode rooms spanning gateways.
- # Leave unset for the safe default: Desktop coordinates cross-gateway rooms and
- # same-gateway rooms can still continue on their own. Set this only to the
- # public HTTPS base URL that another trusted Hermes gateway can reach. The API
- # server key authorizes the initial invitation; scoped grants are signed by a
- # separate installation-private secret and survive API-key rotation. Disbanding
- # a Group Chat revokes its known peer routes; API-key rotation alone does not.
- # Network access should still be limited to the peer gateways (for example with
- # your VPN or firewall).
- # Plain HTTP is accepted only for loopback testing; every reachable peer URL
- # must use HTTPS.
- #
- # gateway:
- # room_link_url: "https://hermes.example.net"
- #
- # The equivalent environment override is HERMES_ROOM_LINK_URL. After changing
- # it, restart the gateway and reopen New Group Chat > Advanced; Hermes will
- # re-check the selected gateways automatically.
- # Startup sweep of session rows orphaned by a dead gateway process.
- # The normal disconnect cleanup runs on an in-process grace timer, so a
- # gateway restart (update, crash, systemd) leaves those rows permanently
- # "active". On every gateway boot — stdio TUI *and* the desktop/dashboard
- # WS sidecar — tui/desktop/subagent rows whose start time AND newest
- # message are both older than the session TTL (HERMES_TUI_SESSION_TTL_S,
- # default 6h) are closed with end_reason "startup_orphan_reap".
- # Messaging-platform sessions (Telegram, Discord, ...) are never touched;
- # live in-memory sessions are excluded; swept sessions stay resumable.
- #
- # dashboard:
- # startup_orphan_sweep: true
- # ─────────────────────────────────────────────────────────────────────────────
- # API Server — per-client model routing
- # ─────────────────────────────────────────────────────────────────────────────
- # Route different API clients to different models/providers on a single
- # Hermes deployment. Clients choose a backend by sending a specific string
- # as the OpenAI ``model`` field. Unmapped model values fall back to the
- # global model configured in the ``model:`` section above, and an explicit
- # session /model override always wins over a route.
- #
- # Configure via the ``platforms.api_server.extra.model_routes`` gateway
- # config block:
- #
- # platforms:
- # api_server:
- # enabled: true
- # extra:
- # key: "your-api-server-secret"
- # model_routes:
- # # Xiaozhi clients send model="minimax-m2" → routed to MiniMax via OpenRouter
- # minimax-m2:
- # model: "minimax/minimax-m1"
- # provider: "openrouter" # optional — overrides global provider
- # # api_key: "sk-..." # optional — per-route UPSTREAM provider
- # # key (NOT caller auth; never logged)
- # # base_url: "https://..." # optional — per-route base URL
- # # GPT clients keep their own alias
- # gpt-5:
- # model: "openai/gpt-5"
- # provider: "openrouter"
- #
- # Configured aliases are automatically listed by GET /v1/models so clients
- # can discover them without manual coordination. Caller authentication is
- # unchanged: every request still authenticates with the global API server
- # key (``extra.key`` / API_SERVER_KEY).
- # ─────────────────────────────────────────────────────────────────────────────
- # Gateway Streaming
- # ─────────────────────────────────────────────────────────────────────────────
- # Stream tokens to messaging platforms in real-time. The bot sends a message
- # on first token, then progressively edits it as more tokens arrive.
- # Disabled by default — enable to try the streaming UX on Telegram/Discord/Slack.
- # For Telegram, partial edits are sent as plain text and only the final edit uses MarkdownV2.
- streaming:
- enabled: false
- # transport: edit # "edit" = progressive editMessageText
- # edit_interval: 0.3 # seconds between message edits
- # buffer_threshold: 40 # chars before forcing an edit flush
- # cursor: " ▉" # cursor shown during streaming
- # =============================================================================
- # Skills Configuration
- # =============================================================================
- # Skills are reusable procedures the agent can load and follow. The agent can
- # also create new skills after completing complex tasks.
- #
- skills:
- # Nudge the agent to create skills after complex tasks.
- # Every N tool-calling iterations, remind the model to consider saving a skill.
- # Set to 0 to disable.
- creation_nudge_interval: 15
- # External skill directories — share skills across tools/agents without
- # copying them into ~/.hermes/skills/. Each path is expanded (~ and ${VAR})
- # and resolved to an absolute path. External dirs are read-only: skill
- # creation always writes to ~/.hermes/skills/. Local skills take precedence
- # when names collide.
- # external_dirs:
- # - ~/.agents/skills
- # - /home/shared/team-skills
- # =============================================================================
- # Agent Behavior
- # =============================================================================
- agent:
- # Maximum tool-calling iterations per conversation (default: 500)
- # Higher = more room for complex tasks, but costs more tokens
- # Recommended: 20-30 for focused tasks, 50-100 for open exploration
- max_turns: 500
- # Inactivity timeout for gateway agent runs (seconds, 0 = unlimited).
- # The agent can run indefinitely when actively calling tools or receiving
- # API responses. Only fires after the agent has been idle for this duration.
- # gateway_timeout: 1800
- # Maximum time an alias routing key waits for an active turn holding the same
- # resolved session lease. On expiry Hermes rejects this inbound message and
- # asks the user to resend rather than running it without serialization.
- # Non-positive values fall back to the 5-second default.
- # gateway_turn_lease_timeout: 5
- # Staged warning: send a warning before escalating to full timeout.
- # Fires once per run when inactivity reaches this threshold (seconds).
- # Set to 0 to disable the warning.
- # gateway_timeout_warning: 900
- # Session stall watchdog (seconds). When a busy session has a pending
- # inbound follow-up and the agent activity clock is idle this long, the
- # gateway logs a WARNING and notifies the user to try /new. Does not kill
- # the turn (see gateway_timeout). 0 = disable. Default 300.
- # session_stall_timeout: 300
- # Transcript-sanitiser repeated-heal escalation (#96870). The pre-send
- # sanitiser silently repairs empty non-final turns on the wire copy; after
- # this many heal passes within a 10-minute session window Hermes logs one
- # ERROR (with session id + heal pattern) and sends a ONE-TIME out-of-band
- # notice to the user suggesting /debug share or `hermes doctor`. The notice
- # goes through the status channel only — conversation context and prompt
- # caching are untouched. Set 0 to disable escalation (per-window WARNINGs
- # still fire). Default 3.
- # sanitizer_heal_escalation_threshold: 3
- # Related in-agent compression timeouts (they live under the top-level
- # compression: block, shown here for discoverability next to the stall
- # watchdog they complement — a hung compression is a common stall cause):
- # compression:
- # context_timeout_seconds: 120 # inactivity budget for in-agent
- # # compress_context (0 = disable)
- # context_total_ceiling_seconds: 600 # absolute cap on the pre-commit
- # # wait even while tokens stream
- # Graceful drain timeout for gateway stop/restart (seconds).
- # Default 0 = no drain: a restart interrupts in-flight agents immediately,
- # cleans up, and exits. Set a positive value only if you want a grace
- # window on /restart, and keep it well under systemd's TimeoutStopSec.
- # restart_drain_timeout: 0
- # Floor under the same drain for cron jobs and api_server (/v1) runs
- # (seconds). Default 30.
- # restart_drain_timeout above is written for chat turns, which are cheap to
- # interrupt: the user is told the gateway is restarting and the session
- # resumes on their next message. A cron run has no such safety net — it is
- # recorded in jobs.json as a permanent failure, nobody is waiting on it, and
- # a recurring job simply skips to its next schedule. So in-flight cron work
- # gets its own grace window instead of inheriting the 0 above, and so do
- # api_server (/v1) runs: an interrupted /v1 run fails its waiting caller.
- # Clamped at runtime to the shutdown-watchdog leash (restart_drain_timeout
- # + 60s) minus teardown headroom, so values past ~50s need a matching
- # TimeoutStopSec bump to take effect. Set 0 to opt out and drain cron and
- # /v1 runs on restart_drain_timeout like before.
- # cron_drain_timeout: 30
- # In-band restart wait (seconds) for active turns to finish BEFORE stop()
- # begins. /restart and SIGUSR1 refuse new work, then wait up to this cap for
- # in-flight agent, cron and API runs to complete so the requesting turn is
- # not cut off by restart_drain_timeout. 0 = enter stop()/drain immediately.
- # Excluded from the wait: wedged runs, and cron runs already handed to a
- # worker in its own restart-safe systemd scope (they outlive the restart).
- # The default (30 min) is a safety valve for wedged agents, not a target
- # latency; raise it for long unattended turns. Env: HERMES_RESTART_AFTER_TURN_TIMEOUT.
- # restart_after_turn_timeout: 1800
- # Upper bound (seconds) a submitted prompt waits for the deferred agent
- # build (MCP discovery, model metadata, skills scan) before failing with a
- # visible error. The wait is patient — the message is delivered as soon as
- # the build completes, and a progress notice is shown past 30s — so this cap
- # only fires on a genuinely hung build. Raise it for deployments with many
- # slow or unreachable MCP servers. Default 600.
- # build_wait_timeout: 600
- # Max app-level retry attempts for API errors (connection drops, provider
- # timeouts, 5xx, etc.) before the agent surfaces the failure. Lower this
- # to 1 if you use fallback providers and want fast failover on flaky
- # primaries (default 3). The OpenAI SDK does its own low-level retries
- # underneath this wrapper — this is the Hermes-level loop.
- # api_max_retries: 3
- # Once api_max_retries AND the fallback chain are spent on a transient outage
- # (5xx, overloaded/529, connect/read timeouts) and nothing has been delivered
- # yet, Hermes waits and retries this many more cycles (jittered 15/30/60/60/60s;
- # a provider Retry-After wins, up to 120s) showing "Provider temporarily
- # unavailable — retrying automatically in Ns (cycle k/5); press Esc to stop"
- # instead of ending the turn. Auth/format/billing/policy errors never enter.
- # Set 0 to disable (default 5).
- # auto_recovery_cycles: 5
- # After the agent edits code without fresh passing verification, nudge it to
- # verify before finishing. Off by default (false). Set true to force it on
- # everywhere, or "auto" for the surface-aware mode: on for interactive
- # coding surfaces (CLI, TUI, desktop) and programmatic callers, off on
- # conversational messaging surfaces (Telegram, Discord, etc.) where the
- # verification summary would reach a human as chat noise. The
- # HERMES_VERIFY_ON_STOP env var (1/0) takes precedence.
- # verify_on_stop: false
- # Standing operator instructions for the coding posture (when Hermes is in a
- # code workspace). Appended to the coding brief as an extra system block, so
- # you can pin project-wide workflow rules without editing the shipped brief.
- # Accepts a string or a list of strings. Takes effect next session.
- # coding_instructions:
- # - "For UI work, don't run tsc/lint until I approve the look."
- # - "Clean the diff before you commit and push."
- # When verify-on-stop finds edited code without fresh verification evidence,
- # append guidance for creative UI work (avoid broad tsc/lint/test before visual
- # approval) and clean-diff expectations. Set false to keep that nudge terse.
- # verify_guidance: true
- # A `pre_verify` hook (plugin or shell, see Event Hooks docs) can keep the
- # agent going one more turn to verify/clean before finishing. This caps how
- # many times one turn may be nudged to continue, so a hook can't trap the loop.
- # Default 3.
- # max_verify_nudges: 3
- # Enable verbose logging
- verbose: false
-
- # Reasoning effort level (OpenRouter and Nous Portal)
- # Controls how much "thinking" the model does before responding.
- # Options: "xhigh" (max), "high", "medium", "low", "minimal", "none" (disable)
- # Providers with bespoke thinking tiers (a relay serving fast/thinking): use the
- # dict form and the level is sent verbatim. Bare strings stay strict (typo guard).
- # reasoning_effort:
- # enabled: true
- # effort: thinking
- reasoning_effort: "medium"
-
- # Per-model reasoning effort overrides (optional dict)
- # Key: any sensible model spelling works (exact, dots↔dashes interchangeable,
- # provider prefix optional). First match wins.
- # Value: reasoning effort level (same options as reasoning_effort, dict form included)
- # Override the global reasoning_effort for that specific model.
- # NOTE: no `hermes config set` support for this key -- edit YAML directly.
- # reasoning_overrides:
- # "openrouter/anthropic/claude-opus-4.5": "xhigh"
- # "openai/gpt-5": "low"
- # "claude-opus-4.6": "high" # bare model name also works
- # "deepseek/deepseek-v4-pro": "xhigh" # dots and dashes are interchangeable
- reasoning_overrides: {}
- # Fast mode (OpenAI Priority Processing / xAI Grok 4.6 / Anthropic Fast Mode
- # on Opus 4.8+). Premium pricing; only sent to first-party endpoints.
- # "" / "normal" - off (default)
- # "fast" - every request
- # "auto" - only the first fast_auto_seconds of every turn
- # "cold" - that window on the first turn of a session only
- # Also: /fast normal|fast|auto|cold [--global]
- service_tier: ""
- fast_auto_seconds: 60
- # OpenAI Daybreak (Trusted Access for Cyber) on the ChatGPT/Codex
- # subscription. Sent only on models your account's Codex catalog lists for
- # it; Desktop shows the switch on those models. Grants no access itself.
- daybreak: false
- # Custom personalities (use with /personality command).
- # Built-ins (helpful, concise, technical, creative, teacher, kawaii, catgirl,
- # pirate, shakespeare, surfer, noir, uwu, philosopher, hype) are always
- # available on every surface — defined once in hermes_cli/personality.py.
- # Entries here ADD new personalities or OVERRIDE a built-in by name.
- # The active selection is stored in display.personality (never here, and
- # never in agent.system_prompt — that field is your manual system prompt).
- personalities: {}
- # mentor: "You are a supportive mentor. Guide, don't lecture."
- # reviewer:
- # system_prompt: "You are a meticulous code reviewer."
- # tone: "direct"
- # style: "terse"
- # =============================================================================
- # Gateway Lifecycle
- # =============================================================================
- gateway:
- # Post-interrupt grace for running agents during an unexpected SIGTERM.
- # Adapter, bridge, and database teardown starts after this many seconds even
- # if an agent has not unwound. Keep it below the service-manager stop budget.
- # signal_interrupt_grace_timeout: 1
- # Let platform adapters honor HTTP_PROXY / HTTPS_PROXY / NO_PROXY (and
- # SSL_CERT_FILE) from the process environment, plus macOS system-proxy
- # auto-detection. Set to false when the gateway inherits a proxy it must not
- # use (e.g. a Windows Scheduled Task picking up a local Clash/V2Ray proxy that
- # isn't running). Explicit per-platform vars like DISCORD_PROXY still apply.
- # trust_env: true
- # =============================================================================
- # Toolsets
- # =============================================================================
- # Control which tools the agent has access to.
- # Use `hermes tools` to interactively enable/disable tools per platform.
- # =============================================================================
- # Platform Toolsets (per-platform tool configuration)
- # =============================================================================
- # Override which toolsets are available on each platform.
- # If a platform isn't listed here, its built-in default is used.
- #
- # You can use EITHER:
- # - A preset like "hermes-cli" or "hermes-telegram" (curated tool set)
- # - A list of individual toolsets to compose your own (see list below)
- #
- # Supported platform keys: cli, telegram, discord, whatsapp, slack, qqbot, teams, google_chat
- #
- # Examples:
- #
- # # Use presets (same as defaults):
- # platform_toolsets:
- # cli: [hermes-cli]
- # telegram: [hermes-telegram]
- #
- # # Custom: give Telegram only web + terminal + file + planning:
- # platform_toolsets:
- # telegram: [web, terminal, file, todo]
- #
- # # Custom: CLI without browser or image gen:
- # platform_toolsets:
- # cli: [web, terminal, file, skills, todo, tts, cronjob]
- #
- # # Restrictive: Discord gets read-only tools only:
- # platform_toolsets:
- # discord: [web, vision, skills, todo]
- #
- # If not set, defaults are:
- # cli: hermes-cli (everything + cronjob management)
- # telegram: hermes-telegram (terminal, file, web, vision, image, tts, browser, skills, todo, cronjob, messaging)
- # discord: hermes-discord (same as telegram)
- # whatsapp: hermes-whatsapp (same as telegram)
- # slack: hermes-slack (same as telegram)
- # signal: hermes-signal (same as telegram)
- # qqbot: hermes-qqbot (same as telegram)
- # teams: hermes-teams (same as telegram)
- # google_chat: hermes-google_chat (same as telegram)
- #
- platform_toolsets:
- cli: [hermes-cli]
- telegram: [hermes-telegram]
- discord: [hermes-discord]
- whatsapp: [hermes-whatsapp]
- slack: [hermes-slack]
- signal: [hermes-signal]
- qqbot: [hermes-qqbot]
- yuanbao: [hermes-yuanbao]
- # Plugin platforms (teams, google_chat, ...) need no row: they default to hermes-<platform> once loaded.
- # =============================================================================
- # Gateway Platform Settings
- # =============================================================================
- # Optional per-platform messaging settings.
- # Platform-specific knobs live under `extra`.
- #
- # platforms:
- # telegram:
- # reply_to_mode: "first" # off | first | all
- # # guest_mode lets explicit @mentions from non-allowlisted groups through.
- # # Default false; ordinary messages, replies, and regex wake words stay blocked.
- # guest_mode: false
- # # allowed_chats: ["-1001234567890"]
- # extra:
- # drop_pending_on_cold_boot: true # Drop Telegram's queued updates on a cold boot (default). Set false on hosts that power off, so messages sent while offline are delivered on the next start; watcher reconnects always preserve them
- # disable_link_previews: false # Set true to suppress Telegram URL previews in bot messages
- # rich_messages: false # Bot API 10.1 rich messages (tables/task lists/details/math); default false for copyable legacy MarkdownV2, set true to opt in
- # rich_drafts: false # Experimental rich draft previews during Telegram DM streaming; default false because Telegram Desktop/macOS can visually overlay draft frames
- # allow_cjk_rich_messages: false # Keep CJK content on the safe MarkdownV2 path unless an unaffected client explicitly opts in to native rich rendering
- # command_menu:
- # # Telegram allows up to 100 BotCommands; Hermes defaults to 60 so
- # # all built-in commands plus common skill commands stay visible
- # # while remaining under Telegram's payload-size limit. Clamped 1..100.
- # max_commands: 60
- # # prepend = user priority first, then Hermes defaults
- # # append = Hermes defaults first, then user priority
- # # replace = only the list below defines priority
- # priority_mode: prepend
- # # Priority is applied across core + plugin + skill commands before
- # # the cap, so a listed skill command always keeps a menu slot.
- # priority:
- # - my_plugin_command
- # - my-important-skill
- # matrix:
- # channel_overrides:
- # "!exampleTeamRoom:example.org":
- # # Successful explicit NO_REPLY/[SILENT] finals stay silent, even when addressed.
- # # Omitted/null/false keeps the current reply policy; failures remain visible.
- # # Restart the gateway after changing room overrides.
- # allow_silence: true
- # slack:
- # extra:
- # # Render live tool calls as Slack-native plan/task cards. This explicit
- # # opt-in works even though Slack text tool_progress defaults to off.
- # native_task_cards: false
- # # Slack user IDs whose Web-API posts (user token, e.g. your own
- # # dashboard/mobile front-end) count as human instead of being dropped
- # # as app traffic. Narrower than allow_bots: all. Users only — never apps.
- # api_human_users: ["U0AAAAAAA", "U0BBBBBBB"]
- # # Suppress automatic link-preview cards without removing clickable links.
- # # Omit either key to preserve Slack's default for that preview type.
- # unfurl_links: false
- # unfurl_media: false
- # webhook:
- # extra:
- # # Route scripts default to a 30 second timeout. Scripts must live under
- # # the active profile's scripts directory and receive webhook JSON on stdin.
- # script_timeout_seconds: 30
- # # Public HTTPS origin in front of this listener; MCP servers with
- # # `events: true` POST event deliveries to <public_url>/webhooks/<route>.
- # public_url: "https://hermes.example.com"
- #
- # Discord-specific settings (config.yaml top-level, not under platforms:):
- #
- # discord:
- # require_mention: true # Require @mention in server channels (default: true)
- # bots_require_inline_mention: true # Bot authors must type a literal @mention (default: true)
- # auto_thread: true # Auto-create thread on @mention (default: true)
- # free_response_auto_thread: false # Free-response channels also auto-thread (default: reply inline)
- # free_response_channels: "" # Channel IDs where no mention is needed
- # reactions: true # Show processing reactions (default: true)
- # history_backfill: true # Recover missed channel messages on mention (default: true)
- # history_backfill_limit: 50 # Max messages to scan backwards (default: 50)
- # ─────────────────────────────────────────────────────────────────────────────
- # Available toolsets (use these names in platform_toolsets or the toolsets list)
- #
- # Run `hermes chat --list-toolsets` to see all toolsets and their tools.
- # Run `hermes chat --list-tools` to see every individual tool with descriptions.
- # ─────────────────────────────────────────────────────────────────────────────
- #
- # INDIVIDUAL TOOLSETS (compose your own):
- # web - web_search, web_extract
- # search - web_search only (no scraping)
- # terminal - terminal, process
- # file - read_file, write_file, patch, search
- # browser - browser_exec, browser_cdp, browser_dialog, browser_vault
- # (local Chromium, or a cloud browser with its API key)
- # vision - vision_analyze (requires OPENROUTER_API_KEY)
- # image_gen - image_generate (requires FAL_KEY)
- # skills - skills_list, skill_view
- # skills_hub - skill_hub (search/install/manage from online registries — user-driven only)
- # todo - todo (in-memory task planning, no deps)
- # tts - text_to_speech (Edge TTS free, or ELEVENLABS/OPENAI/MINIMAX/MISTRAL key)
- # cronjob - cronjob (create/list/update/pause/resume/run/remove scheduled tasks)
- #
- # PRESETS (curated bundles):
- # hermes-cli - All of the above except send_message
- # hermes-telegram - terminal, file, web, vision, image_gen, tts, browser,
- # skills, todo, cronjob, send_message
- # hermes-discord - Same as hermes-telegram
- # hermes-whatsapp - Same as hermes-telegram
- # hermes-slack - Same as hermes-telegram
- #
- # COMPOSITE:
- # debugging - terminal + web + file
- # safe - web + vision (no terminal access)
- # all - Everything available
- #
- # web - Web search and content extraction (web_search, web_extract)
- # search - Web search only, no scraping (web_search)
- # terminal - Command execution and process management (terminal, process)
- # file - File operations: read, write, patch, search
- # browser - Browser automation over CDP (browser_exec drives a real Chromium)
- # vision - Image analysis (vision_analyze)
- # image_gen - Image generation with FLUX (image_generate)
- # skills - Load skill documents (skills_list, skill_view)
- # todo - Task planning and tracking for multi-step work
- # memory - Persistent memory across sessions (personal notes + user profile)
- # session_search - Search and recall past conversations (FTS5 + Gemini Flash summarization)
- # tts - Text-to-speech (Edge TTS free, ElevenLabs, OpenAI, MiniMax, Mistral)
- # cronjob - Schedule and manage automated tasks (CLI-only)
- #
- # Composite toolsets:
- # debugging - terminal + web + file (for troubleshooting)
- # safe - web + vision (no terminal access)
- # NOTE: The top-level "toolsets" key is deprecated and ignored.
- # Tool configuration is managed per-platform via platform_toolsets above.
- # Use `hermes tools` to configure interactively, or edit platform_toolsets directly.
- #
- # CLI override: hermes chat --toolsets terminal,web,file
- # =============================================================================
- # MCP (Model Context Protocol) Servers
- # =============================================================================
- # Connect to external MCP servers to add tools from the MCP ecosystem.
- # Each server's tools are automatically discovered and registered.
- # See website/docs/user-guide/features/mcp.md for full documentation.
- #
- # Stdio servers (spawn a subprocess):
- # command: the executable to run
- # args: command-line arguments
- # env: environment variables (only these + safe defaults passed to subprocess)
- #
- # HTTP servers (connect to a URL):
- # url: the MCP server endpoint
- # headers: HTTP headers (e.g., for authentication)
- #
- # Optional per-server settings:
- # timeout: tool call timeout in seconds (default: 120)
- # connect_timeout: initial connection timeout (default: 60)
- # keepalive_interval: liveness ping cadence in seconds (floored at 5s).
- # HTTP servers default to 180s; lower it below the server's session TTL for
- # servers that expire idle sessions quickly (e.g. Unreal Engine editor MCP,
- # ~15s), otherwise idle tool calls hit an expired session and pay a slow
- # reconnect. Stdio servers disable keepalive by default; set this explicitly
- # to opt a stdio server in.
- # lazy: register the server's tools from the on-disk schema cache at startup
- # and only spawn/connect it on the first tool call (default: false). Saves
- # one child process per server in processes that rarely call tools. Needs
- # one prior live connect to populate the cache; a missing or stale cache
- # entry falls back to the normal eager connect. The banner and the TUI show
- # such a server as "lazy" with its cached tool count until it is first used.
- #
- # mcp_servers:
- # time:
- # command: uvx
- # args: ["mcp-server-time"]
- # filesystem:
- # command: npx
- # args: ["-y", "@modelcontextprotocol/server-filesystem", "/home/user"]
- # notion:
- # url: https://mcp.notion.com/mcp
- # github:
- # command: npx
- # args: ["-y", "@modelcontextprotocol/server-github"]
- # env:
- # GITHUB_PERSONAL_ACCESS_TOKEN: "ghp_..."
- #
- # Sampling (server-initiated LLM requests) — enabled by default.
- # Per-server config under the 'sampling' key:
- # analysis:
- # command: npx
- # args: ["-y", "analysis-server"]
- # sampling:
- # enabled: true # default: true
- # model: "gemini-3-flash" # override model (optional)
- # max_tokens_cap: 4096 # max tokens per request
- # timeout: 30 # LLM call timeout (seconds)
- # max_rpm: 10 # max requests per minute
- # allowed_models: [] # model whitelist (empty = all)
- # max_tool_rounds: 5 # tool loop limit (0 = disable)
- # expose_client_tools: false # advertise extended sampling.tools capability
- # log_level: "info" # audit verbosity
- #
- # MCP events (experimental, webhook delivery) — opt-in per server. Adds the
- # list_event_types / subscribe_event / unsubscribe_event tools; needs
- # platforms.webhook.extra.public_url and a running gateway:
- # feeds:
- # url: https://feeds.example.com/mcp
- # events: true
- # =============================================================================
- # Text-to-Speech
- # =============================================================================
- # TTS defaults to Edge TTS unless changed in ~/.hermes/config.yaml.
- # Gemini TTS supports persona/director prompt files, and Gemini 3.1 Flash TTS
- # can use a hidden auxiliary rewrite pass to insert expressive square-bracket
- # audio tags into the TTS script without showing tags in chat.
- #
- # tts:
- # provider: "gemini"
- # speed: 1.0 # global speed multiplier (provider-specific overrides this)
- # # Per-platform audio upload limits. Built-in defaults: discord 10 MiB,
- # # telegram 50 MiB, everything else 10 MiB, each with safety_ratio 0.85
- # # (the fraction of max_file_bytes actually targeted). Override per platform:
- # delivery_profiles:
- # telegram:
- # max_file_bytes: 52428800
- # safety_ratio: 0.85
- # gemini:
- # model: "gemini-3.1-flash-tts-preview"
- # voice: "Kore"
- # audio_tags: false
- # persona_prompt_file: "" # e.g. ~/.hermes/tts/radio-host.md
- # xai:
- # voice_id: "eve" # built-in or custom voice ID from xAI Console
- # language: "en" # BCP-47 code ("en", "pt-BR") or "auto"
- # speed: 1.0 # 0.7-1.5 playback speed
- # auto_speech_tags: false # insert expressive audio tags via LLM rewrite
- # text_normalization: false # normalize numbers/abbreviations/symbols
- # optimize_streaming_latency: 0 # 0-2, trades quality for lower latency
- # sample_rate: 24000 # 22050 / 24000 / 44100 / 48000
- # bit_rate: 128000 # MP3 bitrate (codec=mp3 only)
- # =============================================================================
- # Voice Transcription (Speech-to-Text)
- # =============================================================================
- # Automatically transcribe voice messages on messaging platforms.
- # Providers: local (free, faster-whisper) | groq (free tier) | openai (Whisper API) | mistral (Voxtral Transcribe)
- # Set the corresponding API key in .env: GROQ_API_KEY, OPENAI_API_KEY, or MISTRAL_API_KEY.
- stt:
- enabled: true
- # provider: "local" # auto-detected if omitted
- # --- Cloud pre-upload silence trim (groq/openai/mistral/xai/elevenlabs/deepinfra) ---
- # Local whisper gets Silero VAD; cloud endpoints otherwise receive raw audio —
- # silence inflates upload time, per-audio-minute billing, and hallucination risk.
- # Collapses pauses with ffmpeg client-side; clips under 12s skip the trim, and
- # on any failure the original uploads untouched.
- # cloud_trim_silence: true # set false to always upload the original audio
- # cloud_trim_threshold_db: -40 # audio quieter than this counts as silence
- # cloud_trim_keep_ms: 300 # how much of each pause survives (keeps natural pacing)
- # Optional static transcription prompt: vocabulary/context hints for
- # prompt-capable backends. Composition is deterministic: this config value
- # is the base, then `pre_transcription` hooks run in registration order with
- # last-writer-wins semantics per field.
- #
- # Privacy: the final prompt is sent to the configured STT provider alongside
- # the audio. Do not include secrets or session context you would not send to
- # that provider.
- #
- # Length / truncation contract: Whisper-family backends (local, openai,
- # groq, deepinfra) only use the final ~224 prompt tokens, so Hermes
- # truncates longer prompts client-side (keeping the tail) with a WARNING
- # log — never an error. Other backends own their own validation.
- #
- # Provider behavior:
- # local (faster-whisper) -> `initial_prompt`, forwarded unchanged
- # openai / groq -> `prompt`, forwarded unchanged
- # mistral -> `prompt`, forwarded unchanged
- # deepinfra -> OpenAI-compatible `prompt`, unchanged
- # local_command/xai/
- # elevenlabs -> unsupported; DEBUG log, then continue
- # plugin providers -> `prompt` in `**extra`; plugin owns limits
- # prompt: "Hermes, Teknium, Nous Research, kanban"
- local:
- model: "base" # tiny | base | small | medium | large-v3 | turbo
- # language: "" # blank = inherit stt.language; set "en", "es", "fr", etc. to force
- # initial_prompt: "" # Optional faster-whisper prompt, e.g. bias Chinese output to simplified Chinese
- # --- Anti-hallucination hardening (whisper decodes junk from silence without these) ---
- # vad: true # Silero VAD filter (default on) — silence never reaches whisper.
- # # Set false to restore raw behavior (e.g. transcribing music/ambient audio).
- # vad_min_silence_ms: 500 # min silence (ms) that splits speech chunks when vad is on
- # no_speech_prob_threshold: 0.6 # drop a segment only if no_speech_prob > this...
- # logprob_threshold: -1.0 # ...AND avg_logprob < this (both must hit — quiet real speech survives)
- # unload_after_idle_seconds: 0 # 0=never unload (default); e.g. 300 releases the model after 5min idle (frees VRAM on GPU; next voice message reloads it)
- language: "en" # Global hint; a nonempty provider language wins. Blank falls back to HERMES_LOCAL_STT_LANGUAGE, then auto-detect.
- # groq:
- # model: "whisper-large-v3-turbo"
- # language: "" # blank inherits stt.language, then HERMES_LOCAL_STT_LANGUAGE
- openai:
- model: "whisper-1" # whisper-1 | gpt-4o-mini-transcribe | gpt-4o-transcribe | gpt-transcribe
- language: "" # blank = inherit stt.language; set "en", "es", "fr", etc. to force
- timeout: 60 # request timeout in seconds; increase for self-hosted model cold starts
- max_retries: 1 # OpenAI SDK transport retries; set 0 to disable
- # mistral:
- # model: "voxtral-mini-latest" # voxtral-mini-latest | voxtral-mini-2602
- # xai:
- # # Sent as the multipart ``model`` field. Hermes uses 2.0 by default;
- # # pin 1.0 only if a rollback is needed.
- # model: "grok-voice-transcribe-2.0" # or grok-voice-transcribe-1.0
- # # language: "" # blank = stt.language > HERMES_LOCAL_STT_LANGUAGE > auto-detect
- # deepinfra:
- # # Model id is discovered live from the DeepInfra catalog filtered
- # # by the `stt` surface tag — leave `model` blank to take the first
- # # live result. Pin only when you need a specific Whisper variant.
- # model: ""
- # Custom command providers (stt.providers.<name>: type: command) can
- # normalize browser WebM/Opus and other input formats to 16 kHz mono m4a
- # before substituting {input_path}. This is opt-in so CLIs that already
- # accept arbitrary containers (e.g. whisper.cpp) remain unchanged.
- # providers:
- # custom-asr:
- # type: "command"
- # command: "custom-asr --input {input_path} --output {output_path}"
- # normalize: true # requires ffmpeg; default false
- # Text-to-speech. Only the deepinfra block is documented here — the
- # remaining providers (edge, openai, xai, minimax, mistral, gemini,
- # elevenlabs, neutts, kittentts, piper) inherit sensible defaults from
- # DEFAULT_CONFIG in hermes_cli/config.py.
- # tts:
- # provider: "deepinfra"
- # deepinfra:
- # # Model id is discovered live from the DeepInfra catalog filtered
- # # by the `tts` surface tag — leave `model` blank to take the first
- # # live result.
- # model: ""
- # voice: "default"
- # "Hey Hermes" hands-free wake word (off by default; toggle with /wake).
- # Full engine/provider options inherit from DEFAULT_CONFIG in
- # hermes_cli/config_defaults.py — only capture placement is documented here.
- # wake_word:
- # enabled: false
- # capture: "auto" # auto | local | client
- # # auto: backend PortAudio mic when one exists; a remote
- # # desktop on a mic-less (headless/VPS) backend
- # # streams its own mic via the wake.feed RPC.
- # # local: always the backend mic (wake_word.input_device).
- # # client: always desktop-streamed PCM (detection stays
- # # on the backend).
- # Image generation. Each provider plugin reads its own ``image_gen.<name>``
- # block; deepinfra discovers models live from
- # api.deepinfra.com/v1/openai/models filtered by the ``image-gen`` tag —
- # no model id is hardcoded, so retired models disappear automatically.
- # image_gen:
- # provider: "deepinfra"
- # deepinfra:
- # # Leave `model` blank for the first live `image-gen`-tagged result.
- # model: ""
- # =============================================================================
- # Response Pacing (Messaging Platforms)
- # =============================================================================
- # Add human-like delays between message chunks.
- # human_delay:
- # mode: "off" # "off" | "natural" | "custom"
- # min_ms: 800 # Min delay (custom mode only)
- # max_ms: 2500 # Max delay (custom mode only)
- # =============================================================================
- # Session Logging
- # =============================================================================
- # Session trajectories are automatically saved to logs/ directory.
- # Each session creates: logs/session_YYYYMMDD_HHMMSS_UUID.json
- #
- # The session ID is displayed in the welcome banner for easy reference.
- # Logs contain full conversation history in trajectory format:
- # - System prompt, user messages, assistant responses
- # - Tool calls with inputs/outputs
- # - Timestamps for debugging
- #
- # No configuration needed - logging is always enabled.
- # To disable, you would need to modify the source code.
- # =============================================================================
- # Code Execution Sandbox (Programmatic Tool Calling)
- # =============================================================================
- # The execute_code tool runs Python scripts that call Hermes tools via RPC.
- # Intermediate tool results stay out of the LLM's context window.
- code_execution:
- timeout: 300 # Max seconds per script before kill (default: 300 = 5 min)
- max_tool_calls: 50 # Max RPC tool calls per execution (default: 50)
- # Local execution uses a persistent session kernel: variables, imports, and
- # loaded data survive across execute_code calls in one conversation, so
- # multi-step data work stops re-loading its inputs every call. Pass
- # reset=true on a call to discard state; a timed-out/interrupted cell kills
- # the kernel (next call starts fresh). Remote terminal backends currently
- # run per-call (remote kernel host is tracked follow-up work). Security
- # scrubbing, tool whitelist, and output redaction are identical either way.
- # kernel_idle_timeout: 1800 # Reap kernels idle longer than this (seconds)
- # max_session_kernels: 4 # Process-wide LRU cap on live kernels
- # =============================================================================
- # Subagent Delegation
- # =============================================================================
- # The delegate_task tool spawns child agents with isolated context.
- # Supports single tasks and batch mode (default 3 parallel, configurable).
- delegation:
- # backend: "default" # Strategy routing each delegate_task task: "default" = the keys below,
- # else the parent's route. model_map / profile_routes come from the
- # advanced-delegation plugin; a backend that is missing, disabled or
- # unavailable fails closed (tool error naming it; /delegation default or
- # `hermes config set delegation.backend default` restores the default).
- # backends: # Each plugin backend's own settings, keyed by its backend name.
- # model_map: # backend: model_map (advanced-delegation)
- # options:
- # - name: fast
- # description: "Cheap and quick; lookups, summaries, simple edits."
- # provider: "openrouter"
- # model: "google/gemini-3-flash-preview"
- # - name: deep
- # description: "Slow and careful; hard debugging and design work."
- # provider: "anthropic"
- # model: "claude-opus-5"
- # reasoning_effort: "high"
- # service_tier: "fast" # optional: that task is refused when the route has no such tier
- # # fallback_providers: [...] # optional; absent = delegation.fallback_providers, [] = none
- # default: fast # optional: option for tasks that set none
- # profile_routes: # backend: profile_routes (advanced-delegation). A listed
- # profiles: # profile's subagents use ITS credentials, pool, SOUL.md and terminal policy.
- # - name: coder
- # description: "Writes and refactors code; runs tests."
- max_iterations: 250 # Max tool-calling turns per child (default: 250)
- # max_concurrent_children: 10 # Max parallel child agents per batch (default: 10, floor: 1, no ceiling).
- # WARNING: values above 10 multiply API cost linearly.
- # max_spawn_depth: 1 # Delegation tree depth cap (range: 1-3, default: 1 = flat).
- # Raise to 2 to allow workers to spawn their own subagents.
- # Requires role="orchestrator" on intermediate agents.
- # orchestrator_enabled: true # Kill switch for role="orchestrator" children (default: true).
- # subagent_auto_approve: false # When a subagent hits a dangerous-command approval prompt, auto-deny (default: false)
- # or auto-approve "once" (true) instead of blocking on stdin.
- # The parent TUI owns stdin, so blocking would deadlock; non-interactive resolution is required.
- # Both choices emit a logger.warning audit line. Flip to true only for cron/batch pipelines.
- # inherit_mcp_toolsets: true # When explicit child toolsets are narrowed, also keep the parent's MCP toolsets (default: true). Set false for strict intersection.
- # model: "google/gemini-3-flash-preview" # Override model for subagents (empty = inherit parent)
- # provider: "openrouter" # Override provider for subagents (empty = inherit parent)
- # # Resolves full credentials (base_url, api_key) automatically.
- # # Supported: openrouter, nous, zai, kimi-coding, minimax
- # # Cost tip: keep the parent on a frontier model and pin
- # # delegation.model to an inexpensive one — children carry the
- # # vast majority of tokens, so this is where spend is cut while
- # # planning quality stays with the frontier parent.
- # fallback_providers: # Fallback chain for delegated children (same entry format as the
- # - provider: "openrouter" # top-level list). Pinned children use only a declared child chain;
- # model: "deepseek/deepseek-chat" # unpinned children inherit when unset. [] disables fallback (#65038).
- # =============================================================================
- # Honcho Integration (Cross-Session User Modeling)
- # =============================================================================
- # AI-native persistent memory via Honcho (https://honcho.dev/).
- # Builds a deeper understanding of the user across sessions and tools.
- # Runs alongside USER.md — additive, not a replacement.
- #
- # Requires: hermes plugins install honcho (Plastic Labs' plugin from the catalog)
- # Config: ~/.honcho/config.json (shared with Claude Code, Cursor, etc.)
- # API key: HONCHO_API_KEY in ~/.hermes/.env or ~/.honcho/config.json
- #
- # Hermes-specific overrides (optional — most config comes from ~/.honcho/config.json):
- # honcho: {}
- # =============================================================================
- # Display
- # =============================================================================
- display:
- # Use compact banner mode (hides the ASCII-art banner, shows a single line).
- # true: Compact single-line banner
- # false: Full ASCII banner with tool/skill summary (default)
- compact: false
- # Tool progress display level (CLI, gateway, and the TUI/Desktop tool feed)
- # off: Silent — no tool activity shown, just the final response
- # new: Show a tool indicator only when the tool changes (skip repeats)
- # all: Show every tool call with a short preview (default)
- # verbose: Full args, results, and debug logs (same as /verbose)
- # log: Silent in chat; append every tool call to ~/.hermes/logs/tool_calls.log (gateway only)
- # Toggle at runtime with /verbose in the CLI
- # Independent of show_reasoning: hiding reasoning never hides tool rows.
- # Answer-only (just the final response) is show_reasoning: false plus
- # tool_progress: off.
- # tool_progress: all
- # tool_progress, interim_assistant_messages, long_running_notifications,
- # busy_ack_detail and show_reasoning are left unset here on purpose: a value
- # set under display: applies to EVERY messaging platform and replaces its
- # built-in default. To change one platform, set it under
- # display.platforms.<platform> instead.
- #
- # Per-platform defaults can be quieter than the CLI. Telegram
- # tunes for mobile: tool_progress and busy_ack_detail default off (no
- # per-tool breadcrumb stream, no "iteration 21/60" debug detail in busy
- # acks or heartbeats), but interim_assistant_messages and
- # long_running_notifications STAY ON so the user has real signal between
- # turn start and final answer (mid-turn assistant commentary + a single
- # edit-in-place "⏳ Working — N min" heartbeat). Override under
- # display.platforms.telegram.
- # Auto-cleanup of temporary progress bubbles after the final response lands.
- # On platforms that support message deletion (currently Telegram), this
- # removes the tool-progress bubble, "⏳ Still working..." notices, and
- # context-pressure status messages once the final reply has been delivered —
- # keeping long-running turns visible live, then tidy afterward. Failed runs
- # leave the bubbles in place as breadcrumbs. Off by default.
- # Per-platform override: display.platforms.telegram.cleanup_progress
- # true: Delete tracked progress/status bubbles on successful turn
- # false: Leave everything in place (default)
- # Example:
- # display:
- # platforms:
- # telegram:
- # cleanup_progress: true
- cleanup_progress: false
- # Natural mid-turn assistant updates.
- # On gateway platforms, when true, completed assistant status messages are
- # sent as separate chat messages. On the Desktop app, when true, mid-turn
- # assistant narration streamed between tool calls is kept in the transcript
- # instead of the bubble collapsing to only the final message on completion.
- # Independent of tool_progress and gateway streaming.
- # true: Keep/send mid-turn assistant updates (default; off on platforms
- # that cannot edit messages)
- # false: Only keep/send the final response
- # interim_assistant_messages: true
- # Opt-in: hide automatic warning/diagnostic notifications (compression, retry
- # and fallback notices, credit/subagent failure lines, inactivity and watchdog
- # stalls, media-delivery fallbacks, cron failure alerts) on this surface.
- # Requested results, approvals, clarifications and command outcomes are never
- # hidden, and nothing is removed from the logs. Per-platform override under
- # display.platforms.<platform>.suppress_warning_notifications.
- suppress_warning_notifications: false
- # Gateway-only long-running status heartbeats.
- # When false, the platform does not receive periodic "⏳ Working — N min"
- # notifications even if agent.gateway_notify_interval is non-zero. The
- # heartbeat edits a single message in place (where the adapter supports
- # editing) instead of posting a new bubble each interval.
- # Default: true, including Telegram (silent agents are worse than a single
- # edit-in-place heartbeat); false on Slack and on platforms that cannot edit.
- # long_running_notifications: true
- # Include detailed iteration/tool/status context in busy acknowledgments
- # and long-running heartbeats. When true, busy acks show "iteration 21/60,
- # terminal, 10 min" and the heartbeat shows "⏳ Working — 12 min,
- # iteration 21/60, terminal". When false (Telegram, Slack and no-edit
- # platform default), both stay terse: "Interrupting current task" and
- # "⏳ Working — 12 min, terminal".
- # busy_ack_detail: true
- # What Enter does when Hermes is already busy (CLI and gateway platforms).
- # interrupt: Interrupt the current run and redirect Hermes (default)
- # queue: Queue your message for the next turn
- # steer: Inject your message mid-run via /steer, arriving at the agent
- # after the next tool call — no interrupt, no role violation.
- # Falls back to 'queue' if the agent isn't running yet or if
- # images are attached (steer only carries text).
- # Ctrl+C (or /stop in gateway) always interrupts regardless of this setting.
- # Toggle at runtime with /busy <interrupt|queue|steer>.
- busy_input_mode: interrupt
- # Background process notifications (gateway/messaging only).
- # Controls how chatty the process watcher is when you use
- # terminal(background=true, notify_on_complete=true) from Telegram/Discord/etc.
- # concise: One-line status message on completion; failures include a short
- # output tail (default)
- # off: No watcher messages at all
- # result: Only the final raw-output completion message
- # error: Only the final raw-output message when exit code != 0
- # all: Running output updates + final raw-output message
- background_process_notifications: concise
- # Play terminal bell when agent finishes a response.
- # Useful for long-running tasks — your terminal will ding when the agent is done.
- # Works over SSH. Most terminals can be configured to flash the taskbar or play a sound.
- # true: Ring the terminal bell on each response
- # false: Silent (default)
- bell_on_complete: false
- # Play terminal bell when a blocking prompt opens and waits on you:
- # clarify questions, dangerous-command approvals, sudo password, secret
- # capture. Same mechanism as bell_on_complete (\a) — works over SSH.
- # true: Ring whenever the agent is waiting for your input
- # false: Silent (default)
- bell_on_prompt: false
- # Show model reasoning/thinking before each response.
- # When enabled, a dim box shows the model's thought process above the response.
- # Toggle at runtime with /reasoning show or /reasoning hide.
- # true: Show the reasoning box (CLI default)
- # false: Hide reasoning (messaging platform default)
- # show_reasoning: true
- # Stream tokens to the terminal as they arrive instead of waiting for the
- # full response. The response box opens on first token and text appears
- # line-by-line. Tool calls are still captured silently.
- # true: Stream tokens as they arrive (default)
- # false: Wait for the full response before rendering
- streaming: true
- # Show [HH:MM] timestamps on user input and assistant response labels.
- # timestamps: false
- # ───────────────────────────────────────────────────────────────────────────
- # Skin / Theme
- # ───────────────────────────────────────────────────────────────────────────
- # Customize CLI visual appearance — banner colors, spinner faces, tool prefix,
- # response box label, and branding text. Change at runtime with /skin <name>.
- #
- # Built-in skins:
- # default — Classic Hermes gold/kawaii
- # ares — Crimson/bronze war-god theme with spinner wings
- # mono — Clean grayscale monochrome
- # slate — Cool blue developer-focused
- # daylight — Bright light-mode theme
- # warm-lightmode — Warm paper-tone light-mode theme
- # poseidon — Sea-green/teal Olympian theme
- # sisyphus — Earthy stone-and-moss theme
- # charizard — Fiery orange dragon theme
- #
- # Custom skins: drop a YAML file in ~/.hermes/skins/<name>.yaml
- # Schema (all fields optional, missing values inherit from default):
- #
- # name: my-theme
- # description: Short description
- # colors:
- # banner_border: "#HEX" # Panel border
- # banner_title: "#HEX" # Panel title
- # banner_accent: "#HEX" # Section headers (Available Tools, etc.)
- # banner_dim: "#HEX" # Dim/muted text
- # banner_text: "#HEX" # Body text (tool names, skill names)
- # ui_accent: "#HEX" # UI accent color
- # response_border: "#HEX" # Response box border color
- # spinner:
- # waiting_faces: ["(⚔)", "(⛨)"] # Faces shown while waiting
- # thinking_faces: ["(⚔)", "(⌁)"] # Faces shown while thinking
- # thinking_verbs: ["forging", "plotting"] # Verbs for spinner messages
- # wings: # Optional left/right spinner decorations
- # - ["⟪⚔", "⚔⟫"]
- # - ["⟪▲", "▲⟫"]
- # branding:
- # agent_name: "My Agent" # Banner title and branding
- # welcome: "Welcome message" # Shown at CLI startup
- # response_label: " ⚔ Agent " # Response box header label
- # prompt_symbol: "⚔" # Prompt symbol (bare token; renderers add trailing space)
- # tool_prefix: "╎" # Tool output line prefix (default: ┊)
- #
- skin: default
- # =============================================================================
- # Model Aliases — short names for /model command
- # =============================================================================
- # Map short aliases to exact (model, provider, base_url) tuples.
- # Used by /model tab completion and resolve_alias().
- # Aliases are checked BEFORE the models.dev catalog, so they can route
- # to endpoints not in the catalog (e.g. Ollama Cloud, local servers).
- #
- # An alias pointing at its own endpoint can carry that endpoint's
- # credential with `api_key` (a literal, or a "${VAR}" reference) or
- # `key_env` (an env var name). `api_key` wins if both are set. When
- # neither is set the key is resolved from the alias HOST — never from
- # whatever provider was active before the switch, which would send that
- # provider's secret to an unrelated third party.
- #
- # model_aliases:
- # opus:
- # model: claude-opus-4-6
- # provider: anthropic
- # qwen:
- # model: "qwen3.5:397b"
- # provider: custom
- # base_url: "https://ollama.com/v1"
- # glm:
- # model: glm-4.7
- # provider: custom
- # base_url: "https://ollama.com/v1"
- # theta:
- # model: theta-1
- # provider: custom
- # base_url: "https://theta.example.com/v1"
- # key_env: THETA_API_KEY # or: api_key: "${THETA_API_KEY}"
- # =============================================================================
- # Privacy
- # =============================================================================
- # privacy:
- # # Redact PII from the LLM context prompt.
- # # When true, phone numbers are stripped and user/chat IDs are replaced
- # # with deterministic hashes before being sent to the model.
- # # Names and usernames are NOT affected (user-chosen, publicly visible).
- # # Routing/delivery still uses the original values internally.
- # redact_pii: false
- # =============================================================================
- # Shell-script hooks
- # =============================================================================
- # Register shell scripts as plugin-hook callbacks. Each entry is executed as
- # a subprocess (shell=False, shlex.split) with a JSON payload on stdin. On
- # stdout the script may return JSON that either blocks the tool call or
- # injects context into the next LLM call.
- #
- # Valid events (mirror hermes_cli.plugins.VALID_HOOKS):
- # pre_tool_call, post_tool_call, pre_llm_call, post_llm_call,
- # pre_api_request, post_api_request, on_session_start, on_session_end,
- # on_session_finalize, on_session_reset, subagent_stop
- #
- # First-use consent: each (event, command) pair prompts once on a TTY, then
- # is persisted to ~/.hermes/shell-hooks-allowlist.json. Non-interactive
- # runs (gateway, cron) need --accept-hooks, HERMES_ACCEPT_HOOKS=1, or the
- # hooks_auto_accept key below.
- #
- # See website/docs/user-guide/features/hooks.md for the full JSON wire
- # protocol and worked examples.
- #
- # hooks:
- # pre_tool_call:
- # - matcher: "terminal"
- # command: "~/.hermes/agent-hooks/block-rm-rf.sh"
- # timeout: 10
- # post_tool_call:
- # - matcher: "write_file|patch"
- # command: "~/.hermes/agent-hooks/auto-format.sh"
- # pre_llm_call:
- # - command: "~/.hermes/agent-hooks/inject-cwd-context.sh"
- # subagent_stop:
- # - command: "~/.hermes/agent-hooks/log-orchestration.sh"
- #
- # hooks_auto_accept: false
- # =============================================================================
- # Telemetry
- # =============================================================================
- # Shared metrics are disabled by default. When enabled, Hermes writes only
- # allowlisted aggregate counters and immutable JSON
- # packages under $HERMES_HOME/telemetry/shared_metrics.
- # Packages include a random profile-scoped ID that stays stable until this
- # directory is deleted. It is not derived from hardware, account, or host data.
- # Successfully exported local history is retained for 30 days; pending deltas
- # are retained until they can be exported.
- # This profile-owned choice is not overridden by managed-scope configuration.
- #
- # Nothing is uploaded unless you also set `send: true`. That is a separate
- # opt-in and requires `enabled`; it never turns collection on by itself.
- # When sending is on:
- # * only packages whose entire collection period falls within one
- # continuous recorded consent window are ever sent. Consent windows open
- # when you enable sending and close when you disable it, so data
- # collected before you opted in — or during any gap between opt-ins —
- # stays on this machine;
- # * each package carries the profile-scoped ID as-is. It is a random UUID
- # with no hardware, account, or host-derived content, and deleting the
- # shared-metrics directory resets it.
- # See docs/observability/relay-shared-metrics.md (Appendix A) for the full
- # consent, identity, retention, and deletion decisions.
- telemetry:
- shared_metrics:
- enabled: false
- send: false
- # endpoint: https://telemetry.nousresearch.com/v1/telemetry
- # =============================================================================
- # Login Policy (credentials themselves live in auth.json / .env)
- # =============================================================================
- auth:
- # Borrow and refresh the Codex CLI (~/.codex/auth.json) and Claude Code logins when Hermes has
- # no usable login of its own. Their refresh tokens are single-use, so two programs on one login
- # can log each other out; set false to make Hermes use only its own logins.
- adopt_external_logins: true
- # How `hermes auth add openai-codex` / `hermes model` sign in to OpenAI Codex.
- # device_code (default) — open a URL and enter a code.
- # browser — authorization-code + PKCE on http://localhost:1455/auth/callback (the redirect
- # OpenAI registered for the Codex client) for organizations that disable the
- # device-code grant; falls back to device code when that port is busy.
- # `hermes auth add openai-codex --browser` opts in for a single login without changing this.
- codex_login_flow: device_code
- # =============================================================================
- # Update Behavior
- # =============================================================================
- updates:
- # Disable passive CLI/banner update checks with `hermes config set updates.check false`.
- # Explicit `hermes update --check` and `hermes update` remain available.
- check: true
- # Create a full HERMES_HOME zip before every `hermes update`.
- # Backups land in ~/.hermes/backups/ and can be restored with `hermes import`.
- # Off by default because large homes can add minutes to every update.
- pre_update_backup: false
- # Number of pre-update backup zips to retain.
- backup_keep: 5
- # What non-interactive updates do with local source edits in the Hermes repo.
- # Interactive terminal updates always prompt before restoring the autostash.
- #
- # stash - auto-stash before pull, then auto-restore after success (default)
- # discard - drop the update-created stash after success; use only on managed
- # installs where local source edits should not persist
- non_interactive_local_changes: "stash"
- # =============================================================================
- # Web Dashboard
- # =============================================================================
- # OAuth gate configuration for `hermes dashboard --host <non-loopback>`.
- # The bundled Nous Portal plugin reads these on startup; settings here are
- # the canonical surface. Each can be overridden by an environment variable:
- #
- # dashboard.oauth.client_id <- HERMES_DASHBOARD_OAUTH_CLIENT_ID
- # dashboard.oauth.portal_url <- HERMES_DASHBOARD_PORTAL_URL
- # dashboard.public_url <- HERMES_DASHBOARD_PUBLIC_URL
- #
- # Env wins when set to a non-empty value. This is what Fly.io's platform-
- # secret injection uses to push per-deploy client_ids without needing to
- # bake a config.yaml into the image. Empty env values are treated as unset
- # so a provisioned-but-not-populated secret can't shadow a valid entry here.
- #
- # Local dev / on-prem deploys should typically set these via config.yaml
- # (the ~/.hermes/.env file is reserved for API keys and secrets).
- #
- # dashboard:
- # oauth:
- # client_id: "" # agent:{instance_id}; Portal provisions this at deploy
- # portal_url: "" # blank → default https://portal.nousresearch.com
- #
- # # Force the absolute base URL the OAuth callback (and any other public
- # # URL the dashboard hands to external systems) is built from. Set this
- # # for deploys behind reverse proxies that don't reliably forward
- # # X-Forwarded-Host / X-Forwarded-Proto / X-Forwarded-Prefix (manual
- # # nginx setups, on-prem ingresses, custom-domain Fly deploys without
- # # full proxy header chains).
- # #
- # # When set, the value is the complete authority: scheme + host +
- # # optional path prefix (e.g. "https://example.com/hermes"). The OAuth
- # # callback URL becomes "<public_url>/auth/callback" — X-Forwarded-Prefix
- # # is IGNORED on this code path because the operator has explicitly
- # # declared the public URL and we no longer need to guess.
- # #
- # # Leave empty to use the existing proxy-header reconstruction (the
- # # default — works on Fly.io out of the box).
- # #
- # # public_url: "https://example.com/hermes"
- # #
- # # Reverse proxies connecting from another container or host are not
- # # trusted by default. Add only the proxy's exact IP address, or a bounded
- # # CIDR for a dedicated proxy network, so X-Forwarded-Proto and
- # # X-Forwarded-For can be honored. Loopback is always trusted. Wildcards
- # # and /0 networks are rejected.
- # #
- # # trusted_proxies:
- # # - "172.20.0.5"
- # # # - "172.20.0.0/24" # dedicated network, if the IP is dynamic
- #
- # -----------------------------------------------------------------------------
- # Self-hosted OIDC dashboard auth (generic OpenID Connect — Authentik,
- # Keycloak, Zitadel, Authelia, Auth0, Okta, Google, …). Use this INSTEAD of the
- # nous block above when gating the dashboard with your own identity provider.
- # Each setting can be overridden by an environment variable:
- #
- # dashboard.oauth.self_hosted.issuer <- HERMES_DASHBOARD_OIDC_ISSUER
- # dashboard.oauth.self_hosted.client_id <- HERMES_DASHBOARD_OIDC_CLIENT_ID
- # dashboard.oauth.self_hosted.scopes <- HERMES_DASHBOARD_OIDC_SCOPES
- # dashboard.oauth.self_hosted.client_secret <- HERMES_DASHBOARD_OIDC_CLIENT_SECRET
- #
- # dashboard:
- # oauth:
- # provider: self-hosted
- # self_hosted:
- # issuer: "https://auth.example.com/application/o/hermes/" # required
- # client_id: "hermes-dashboard" # required
- # scopes: "openid profile email" # optional
- #
- # # OPTIONAL — set ONLY if your IDP registered the client as
- # # *confidential* (Authentik / Keycloak often default to this). When
- # # set, Hermes authenticates the client at the token endpoint
- # # (client_secret_basic or client_secret_post, auto-selected from the
- # # IDP's discovery doc) IN ADDITION to PKCE. Leave unset for a public
- # # (PKCE-only) client — the common case.
- # #
- # # This is a CREDENTIAL: prefer setting HERMES_DASHBOARD_OIDC_CLIENT_SECRET
- # # in ~/.hermes/.env over putting it here in config.yaml.
- # # client_secret: ""
- # =============================================================================
- # External secret sources
- # =============================================================================
- # Pull provider credentials from external secret managers at process startup
- # instead of storing them in ~/.hermes/.env. Only the manager's bootstrap
- # credential (e.g. BWS_ACCESS_TOKEN / OP_SERVICE_ACCOUNT_TOKEN) lives in .env
- # (or your shell / desktop session); everything else rotates centrally.
- # Failures never block startup — Hermes warns once and continues with
- # whatever .env already had.
- #
- # Multiple sources can be enabled at once:
- # - "mapped" sources (explicit VAR -> ref bindings, e.g. 1Password's env:
- # map) beat "bulk" sources (whole-project dumps like Bitwarden BSM)
- # - within a shape, the first source to claim a var wins; later claims
- # are skipped with a startup warning (never a silent clobber)
- # - a source's override_existing lets it beat .env/shell values, but
- # never another secret source's claim
- # Docs: https://hermes-agent.nousresearch.com/docs/user-guide/secrets/
- #
- # secrets:
- # # Optional explicit ordering of enabled sources.
- # # sources: [onepassword, bitwarden]
- #
- # # ---- Bitwarden Secrets Manager (bws CLI) --------------------------------
- # bitwarden:
- # enabled: false
- # access_token_env: BWS_ACCESS_TOKEN # bootstrap token, sourced from .env
- # project_id: "" # UUID of the BSM project to sync
- # server_url: "" # "" = US Cloud; EU/self-hosted URL otherwise
- # cache_ttl_seconds: 300 # 0 disables fresh caching
- # encrypted_cache: # optional encrypted stale fallback
- # enabled: false
- # max_stale_seconds: 0 # 0 disables stale fallback
- # override_existing: true # BSM values win over existing env
- # auto_install: true # lazy-download bws into ~/.hermes/bin
- #
- # # ---- 1Password (op CLI) -------------------------------------------------
- # onepassword:
- # enabled: false
- # # Map env-var names to op:// secret references. Each is resolved with a
- # # single `op read` at startup.
- # env:
- # OPENAI_API_KEY: "op://Private/OpenAI/api key"
- # ANTHROPIC_API_KEY: "op://Private/Anthropic/credential"
- # account: "" # op --account shorthand; "" = default
- # service_account_token_env: OP_SERVICE_ACCOUNT_TOKEN # headless auth; unset = desktop session
- # binary_path: "" # "" = resolve op via PATH; else absolute path
- # cache_ttl_seconds: 300 # 0 disables BOTH cache layers
- # override_existing: true # resolved values win over existing env
- #
- # # ---- Command helper (any CLI vault) --------------------------------------
- # # Run a user-configured helper that prints KEY=VALUE lines on stdout —
- # # works with any secret store that has a CLI: keepassxc-cli, secret-tool,
- # # pass, gpg, or a script that cats a tmpfs env file. Composes with the
- # # sources above (enable any combination). POSIX-only (needs /bin/sh).
- # # The helper must be fast and NON-interactive (hard timeout, 1 MiB cap);
- # # its stderr is discarded so diagnostics can't leak secret material.
- # command:
- # enabled: false
- # command: "cat /run/user/1000/hermes-secrets.env"
- # helper_timeout_seconds: 3
- # override_existing: false # .env/shell win by default
- # # This runs ONCE per process at startup, so it cannot replace a credential
- # # that expires mid-session. For a single provider whose token needs
- # # re-minting during a session, use `providers.<name>.key_cmd` instead.
|