atproto-mcp

MCP server providing a searchable knowledge base for the AT Protocol ecosystem — protocol documentation, lexicon schemas, Bluesky developer API docs, and cookbook examples — powered by txtai semantic search.

Data Sources

Related MCP server: mcp-docs

Tools

Prompts

Installation

Prerequisites

• Python 3.12+

• uv (recommended) or pip

• Git (for cloning source repositories)

Install from source

git clone https://github.com/Ashex/atproto-mcp.git
cd atproto-mcp
uv sync

Run with uvx

uvx atproto-mcp

Configuration

Claude Desktop (MCPB bundle) — recommended

Download atproto-mcp-<version>.mcpb from the latest release and open it. Claude Desktop shows an install dialog with optional fields for the cache directory, refresh interval, and embedding model; leave them blank for the defaults. No Python or JSON editing required — the host installs dependencies with uv.

Bundles are self-signed, so the installer shows an "unknown publisher" warning. See Releases for what that does and does not guarantee.

Every other host below uses the PyPI package and is unaffected by the bundle.

VS Code / Copilot

Add to .vscode/mcp.json in your workspace:

{
  "mcpServers": {
    "atproto": {
      "command": "uvx",
      "args": [
        "atproto-mcp"
        ]
    }
  }
}

Kiro Power

1. Open Kiro → Powers

2. Select Import power from GitHub

3. Enter https://github.com/Ashex/atproto-mcp

Claude Desktop (manual)

Prefer the MCPB bundle above. To configure by hand instead, add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "atproto": {
      "command": "uvx",
      "args": [
         "atproto-mcp"
      ]
    }
  }
}

MCPHub

Add to ~/.config/mcphub/servers.json:

{
  "mcpServers": {
    "atproto": {
      "command": "uvx",
      "args": ["atproto-mcp"]
    }
  }
}

OpenCode

Add to your opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "atproto": {
      "type": "local",
      "command": ["uvx", "atproto-mcp"]
    }
  }
}

Environment Variables

How It Works

The server connects immediately and warms up in the background, so a slow first run never blocks the host's startup handshake.

On first launch, warmup:

1. Shallow clones the repos into ~/.cache/atproto-mcp/repos/

2. Parses MDX docs, lexicon schemas, and cookbook examples into text chunks

3. Indexes the chunks using txtai hybrid search (BM25 keyword + dense vectors from the bge-small-en-v1.5 sentence-transformer, ~130MB, runs locally) — exact identifiers like NSIDs match reliably alongside semantic queries

4. Index is persisted in ~/.cache/atproto-mcp/index/ for subsequent starts

This takes a few minutes. Until it finishes, tools return a short message describing what the server is doing rather than failing. If warmup fails — no network yet, for example — the server stays up, says so, and retries on its own with exponential backoff until an index is in service. refresh_sources forces a retry immediately.

On subsequent launches the cached index is loaded and put into service before any network access, so queries work within seconds even offline. Repos older than 24 hours are then refreshed with a shallow fetch and a hard reset to the tracked branch. If the refreshed repos differ from what the index was built from, the existing index keeps serving queries while a fresh one is rebuilt in the background and swapped in when ready.

Development

# Install in development mode
uv sync

# Run the server locally (stdio)
uv run atproto-mcp

# Test with the MCP Inspector
uv run mcp dev src/atproto_mcp/server.py

# Run with debug logging
ATPROTO_MCP_CACHE_DIR=/tmp/atproto-mcp uv run atproto-mcp

# Run the test suite (what CI runs)
uv run python -m unittest discover -s tests -p 'test*.py'

# Build the MCPB bundle
npm install -g @anthropic-ai/mcpb
mcpb validate .
mcpb pack . /tmp/atproto-mcp.mcpb

License

MIT