devdocs-mcp AGENTS.md

devdocs-mcp Development Guidelines

Project Structure

src/devdocs_mcp/
├── __init__.py      # Entry point
├── __main__.py      # CLI entry point with error handling
├── config.py        # XDG paths (platformdirs), settings, sources
├── catalog.py       # Fetches 819 docs from devdocs.io API
├── download.py      # Downloads/extracts .tar.gz bundles
├── embedder.py      # HTML→text extraction + SQLite metadata storage
├── chunking.py      # Document chunking strategies (sentence/paragraph)
├── http_utils.py    # Shared HTTP retry logic with exponential backoff
├── faiss_index.py   # FAISS vector index with IndexIDMap
└── mcp_server.py    # 10 MCP tools via FastMCP

Setup & Dependencies

• Python >=3.13
• uv sync for core deps (mcp, httpx, platformdirs)
• uv sync --extra ml for ML deps (faiss-cpu, sentence-transformers, numpy)
• ML deps are optional — server runs without them but semantic search requires them

Building New Tools

1. Add tool function with @mcp.tool() decorator in mcp_server.py
2. Import dependencies at top of file
3. Keep tool descriptions clear and parameter types explicit
4. Return strings for simple responses; use structured formatting for multi-field output

Code Style

• Type hints on all public functions
• Docstrings with Args/Returns sections
• No print() in library code — use logging.getLogger(name)
• Import order: stdlib → third-party → local (with noqa: E402 if needed)

Testing

uv run python -c "from devdocs_mcp.mcp_server import ML_AVAILABLE, list_docs; ..."

Adding New Doc Scrapers

devdocs.io bundles are served from https://downloads.devdocs.io/{slug}.tar.gz containing:

• index.json — entry index with path/type info
• db.json — page content map (path → HTML)
• meta.json — metadata

New slugs can be added by downloading the tarball and extracting. The embedder handles all formats automatically.

Configuration

Config stored at XDG paths:

• Config: ~/.config/devdocs-mcp/config.json
• Cache: ~/.cache/devdocs-mcp/ (docs + embeddings)