magpie-search-mcp

npx launcher for Magpie Search — a federated search engine for AI agents, exposed as an MCP server.

npx -y magpie-search-mcp

What Magpie is

A normal search engine looks in one place. Magpie takes one query and fans it across five sources at once — your AI's conversation history, your local files, a knowledge graph, a vector store (all on your machine), plus the live web when you need outside facts — and pulls the answer back from wherever it lives. It searches each the right way: exact/regex, keyword (BM25), or by meaning (local embeddings), or all of them fused.

Every hit is tagged with a trust tier (fact > reference > lead > stale), fused into one ranked set, deduped across sources, and trimmed to a token budget so it never floods the context window.

How it helps

• Never lose context. Your transcripts, files, and knowledge graph are searched locally, so your agent recovers what it was working on after a crash, reboot, or new session instead of you re-explaining it.
• Saves tokens. Retrieval happens outside the model, so you stop burning tokens re-feeding history into the prompt. The deepweb mode does a full multi-angle research sweep for ~1,050 tokens — searching and page-reading are pure retrieval (zero model tokens), so your model only pays for one synthesis pass instead of the ~millions an agent-swarm spends.
• It's not RAG — it's a better R. No generator in the path. Magpie returns ranked, trust-tagged hits and stops; the AI you bring is the "G". A stronger retriever, not a RAG.
• Fully local. No server, no account, telemetry off unless you opt in.

Use it in an MCP client

Add it to your client config (e.g. Claude Desktop claude_desktop_config.json):

{
  "mcpServers": {
    "magpie-search": {
      "command": "npx",
      "args": ["-y", "magpie-search-mcp"]
    }
  }
}

The agent then has search, recent, session, list_sessions, stats, and reindex available — federated, trust-ranked, context-budgeted.

Requirements

• Node.js ≥ 18 (for npx)

• uv — the launcher uses uvx to fetch and run the Python server in an isolated environment (no manual pip install needed).

  Install uv once:

  # macOS / Linux
  curl -LsSf https://astral.sh/uv/install.sh | sh
  # Windows
  powershell -c "irm https://astral.sh/uv/install.ps1 | iex"

Prefer Python directly? uvx --from magpie-search magpie-search-mcp, or pip install magpie-search.

Full docs, CLI reference, and the deep-dive on the trust-weighted fusion and the not-RAG argument live in the main repo.

License

Apache-2.0 — Copyright © 2026 VektorGeist LLC.