Darwin RAG

Local-first RAG engine. Ingest documents, index them with BM25 and embeddings, and search via hybrid/semantic/keyword retrieval. Runs as an MCP server for AI agent integration or via CLI. All local, all offline-capable.

Built by BrightDotDev.

Setup

Internet is required for first-time setup to download and validate models. After that, everything runs offline.

Interactive (guided)

Detects your hardware, shows available models, lets you pick. Takes about 2 minutes.

pip install darwin-rag
darwin-admin setup interactive

One-shot: required (embedding only)

Downloads and configures only the embedding model. No prompts, no choices. Fastest path to a working install.

darwin-admin setup --preset required

One-shot: recommended (embedding + reranker + OCR)

Also downloads a reranker and OCR model. Better search quality out of the box.

darwin-admin setup --preset recommended

After setup, start the MCP server:

darwin mcp --http --port 8765

For a full setup walkthrough including Docker, from source, and API key configuration, see docs/setup.md.

Connecting your agent

# Generate config for Claude Desktop, Cursor, etc.
darwin config claude

# Generate config for all clients
darwin config all --copy

Or point any MCP client directly:

• SSE: http://localhost:8765/sse
• HTTP: http://localhost:8765/mcp

Commands

darwin — user CLI (mcp, config)

darwin-admin — power-user CLI (setup, models, store, pipeline, search, logs, system, uninstall, status)

See docs/admin.md for the full command reference.

Documentation

License

MIT with Attribution — see LICENSE.

Core architecture and implementation by BrightDotDev.