name: mediarag repo: https://github.com/dyiapanis/mediarag sha: 29d47c34f53340eb3a4267792c61f38c34ee8e03 description: "Multimodal local search over documents and media: pdf, office files, notes, images,\naudio and subtitles in one corpus, queried by meaning from any MCP-capable agent.\nDocuments and media sit level in the same index, so \"which file mentions the landing\"\nand \"where is the clip where I said it\" are the same kind of question. Retrieval is\nhybrid — semantic vectors fused with an FTS5/BM25 keyword lane — and results carry\nsnippets, timestamps and image thumbnails.\n\nEverything runs locally: the corpus is a SQLite file you own, the embedding weights\ndownload once (~285MB) and every search after that runs offline. Nothing is uploaded\nand there is no telemetry — the only network calls are ones you ask for: an http(s)\nsource you add, or the optional Raindrop adapter.\n\nShips as an Agent Plugins v1 package — a stdio MCP server for conforming clients, plus\nfive native Hermes tools so one gateway serves every agent with the model loaded once.\nLinux and Windows. Disclosure — downloads the ~285 MB EmbeddingGemma-2 ONNX weights from Hugging Face (pinned revision) on first setup; `mediarag_add_source` fetches user-supplied http(s) URLs (public hosts only, re-checked on every redirect) and runs packaged or user `sources.d` adapters as local subprocesses, including the optional Raindrop.io adapter, which sends RAINDROP_API_TOKEN to Raindrop's API; the standalone MCP launcher can build its own Python environment with `uv` from PyPI (~1.7 GB incl. torch) only when MEDIARAG_AUTO_BOOTSTRAP=1." maintainer: dyiapanis tier: community category: tools docs_url: https://github.com/dyiapanis/mediarag#readme version: "0.1.19" readme: true platforms: [linux, windows] capabilities: provides_tools: - mediarag_search - mediarag_add_note - mediarag_ingest_dir - mediarag_add_source - mediarag_status provides_hooks: [] provides_middleware: [] requires_env: []