gptme-wisdom-mcp

RAG-as-MCP Knowledge Server for gptme agents — search foundational CS books and AI session history via the Model Context Protocol (MCP).

Prior art and consolidation

This package is related to the existing gptme/gptme-rag project, which already provides ChromaDB-backed semantic document RAG, file watching, CLI commands, and an MCP server. This contrib package is a narrower MCP surface for the gptme-wisdom book index and Bob/gptme session search.

If gptme-rag is upstreamed into gptme-contrib, this package should be reconciled with that repo instead of replacing or reimplementing its existing indexing stack.

What it does

gptme-wisdom-mcp exposes two knowledge planes as MCP tools:

Tool	What it searches
`search_wisdom(query)`	Indexed CS books (SICP, OSTEP, Pro Git, …) via BM25
`list_wisdom_sources()`	Lists which books are currently indexed
`search_sessions(query)`	gptme journals, Claude Code transcripts, conversation logs

Install

uv tool install gptme-wisdom-mcp
# or
pip install gptme-wisdom-mcp

Quick start

# Run on stdio (default MCP transport)
gptme-wisdom-mcp

# Custom DB paths
gptme-wisdom-mcp --wisdom-db ~/books/wisdom.db --sessions-db ~/sessions/index.db

Claude Desktop config

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "wisdom-rag": {
      "command": "gptme-wisdom-mcp",
      "args": [
        "--wisdom-db", "/path/to/wisdom.db",
        "--sessions-db", "/path/to/session-index.db"
      ]
    }
  }
}

Claude Code config

Add to .claude/settings.json in your project:

{
  "mcpServers": {
    "wisdom-rag": {
      "command": "gptme-wisdom-mcp",
      "args": ["--wisdom-db", "/path/to/wisdom.db"]
    }
  }
}

Or via claude mcp add:

claude mcp add wisdom-rag -- gptme-wisdom-mcp --wisdom-db /path/to/wisdom.db

Building the wisdom index

The wisdom DB is populated separately by indexing CC-licensed books. Use gptme-wisdom to ingest supported books:

gptme-wisdom ingest --source thinkpython --file /tmp/thinkpython.txt
gptme-wisdom search "recursion base case"

Building the session index

The session index is built from gptme conversation logs, Claude Code transcripts, and journal markdown files. Point the server at your existing index:

gptme-wisdom-mcp --sessions-db ~/.local/share/gptme/session-index.db

Acknowledgments

Built on top of the gptme agent framework. Session indexing uses SQLite FTS5 (BM25) — zero external dependencies beyond the mcp package.