pipecat-docs-mcp

Search 317 pages of Pipecat docs + 1,028 GitHub Issues from Claude Desktop or Cursor — grounded answers, no hallucinated APIs.

License: MIT Python 3.10+ FastMCP

What it does

Indexes 8,817 chunks from the Pipecat docs and GitHub Issues, then exposes them through 4 MCP tools. Claude can answer questions about configuration, code examples, architecture concepts, and provider comparisons — with source-cited results pulled directly from the official docs and community issue threads.

Retrieval pipeline: BM25 + FAISS dense vectors → Reciprocal Rank Fusion → Cross-encoder reranker

Metric	Value
Pages indexed	317
GitHub Issues indexed	1,028
Total chunks	8,817
Mean Reciprocal Rank	0.863
Recall@5	100%
Avg query latency	35ms

Tools

Tool	Use it when...
`search_pipecat_docs`	General questions, config options, error messages
`get_example_code`	You need a runnable Python pipeline
`explain_concept`	"What is a Frame / Pipeline / VAD / Transport?"
`compare_services`	Choosing between STT, TTS, or transport providers

Quickstart

1. Clone and install

git clone https://github.com/DTufail/pipecat-docs-mcp.git
cd pipecat-docs-mcp
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

2. Build the search index

data/chunks.jsonl is included. This builds the BM25 and FAISS indexes (~2 min):

python indexer.py

3. (Optional) Index GitHub Issues

Adds 1,028 issues from pipecat-ai/pipecat for error-driven retrieval:

export GITHUB_TOKEN="your_token_here"   # optional but recommended (5k req/hr vs 60)
python github_indexer.py
python indexer.py                        # rebuild indexes with issues merged

4. Connect to Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "pipecat-docs": {
      "command": "/absolute/path/to/.venv/bin/python3",
      "args": ["/absolute/path/to/pipecat-docs-mcp/server.py"]
    }
  }
}

Use the Python binary from your virtual environment, not system Python. Restart Claude Desktop after saving.

5. Connect to Cursor (HTTP)

fastmcp run server.py:mcp --transport http --port 8000

Point your MCP client at http://localhost:8000/mcp.

Project structure

pipecat-docs-mcp/
├── server.py              # FastMCP server — 4 MCP tools
├── retrieval.py           # Hybrid search (BM25 + FAISS + RRF + reranker)
├── indexer.py             # Builds BM25, FAISS, and metadata indexes
├── github_indexer.py      # Fetches GitHub Issues → data/github_issues.jsonl
├── config.py              # Paths and model config
├── test_server.py         # In-process tests for all 4 tools
├── test_retrieval.py      # Retrieval eval — MRR, Recall@5, latency (17 queries)
├── requirements.txt
├── pipecat-scraper/
│   ├── scraper.py         # Sitemap-driven scraper (317 pages, resumable)
│   ├── utils.py           # HTML parsing and chunking
│   ├── audit.py           # Per-page extraction audit
│   └── config.py
└── data/
    ├── chunks.jsonl        # 6,936 scraped doc chunks (included)
    ├── chunk_metadata.json # Title, section, URL, content type per chunk
    ├── bm25_index.pkl      # Generated by indexer.py (gitignored)
    ├── faiss_index.bin     # Generated by indexer.py (gitignored)
    ├── embeddings.npy      # Generated by indexer.py (gitignored)
    └── github_issues.jsonl # Generated by github_indexer.py (gitignored)

Re-scraping the docs

To refresh against a newer Pipecat release:

cd pipecat-scraper
python scraper.py      # scrapes docs.pipecat.ai → data/chunks.jsonl
cd ..
python indexer.py      # rebuilds indexes

Roadmap

Auto-refresh on new Pipecat releases via GitHub webhook
Expanded coverage — LiveKit and Vapi docs

License

MIT