QMD

An on-device document search engine combines keyword and semantic search with local GGUF models and gives AI agents access through MCP.

QMD searches Markdown notes, meeting transcripts, documentation and knowledge bases on your own machine. It's for people who want to find material by its wording or meaning, and make that search available to an AI agent.

Hybrid search: Keyword matching and semantic search work together, with local models expanding queries and reranking results.
Agent access: Structured command-line output and an MCP server let coding agents search your collections; a shared HTTP server avoids repeated model loading.
Multilingual search: The default embeddinggemma model focuses on English, while optional Qwen3-Embedding models support multilingual material.
Document structure: Indexing respects Markdown sections and code blocks, with optional structure-aware handling for TypeScript, JavaScript, Python, Go and Rust files.

Search and model inference run locally through node-llama-cpp using GGUF models. QMD stores its index in SQLite and downloads models from Hugging Face on first use, then caches them on your machine. Collection filters let you limit a search to particular document sets.

QMD is open source under the MIT license. Its requirements include Node.js and Bun; macOS also requires Homebrew SQLite for extension support.