magpie-search — Claude Code MCP server: search agent history and files
A local search engine for your AI agent. One call fans out across transcripts, project files, a knowledge graph, vectors and the web, then returns ranked hits.
claude mcp add magpie-search -- uvx magpie-searchWhat magpie-search does
Magpie keeps a local SQLite index of what your agent has worked through: an FTS5 full-text index ranked by BM25 sits next to a sqlite-vec vector index built from 384-dim embeddings of the all-MiniLM-L6-v2 model, which runs on your machine. A scrubber strips secrets at ingest — keys, tokens, private keys, connection strings — before anything lands in the index. A crash mid-session stops being amnesia.
Five modes cover different ways of looking: grep for exact strings and regexes, lexical for keywords, semantic for meaning via vector K-NN, hybrid mixing both through RRF, and rerank where jina-reranker re-scores the candidates. The federation layer plugs in five providers — transcripts, files, knowledge graph, vector store, web — each tagging hits as fact, reference, lead or stale. Queries fan out concurrently and a slow backend fails open, so it adds nothing instead of stalling the call. Trust-weighted RRF then merges everything, duplicates collapse by content hash, and a token budget keeps the output from flooding the model's context.
Agents reach it over MCP through search, recent, session, list_sessions, stats and reindex. Nothing in that list writes an answer. Installation goes through uvx, so you need uv present before adding the server.
What it helps with
- find an exact error string across past agent sessions
- recover context after a reboot or a crash
- see which sources back up the same fact
- search by meaning when you don't recall the wording
- trim results to fit a model's context window
Example prompts for the agent
- Search my session history for where we discussed the database connection timeout
- Grep the transcripts and project files for
ECONNREFUSED - Run a deep web search on SQLite migrations and tell me how many domains agree
How to install magpie-search
Platform CLI
claude mcp add magpie-search -- uvx magpie-searchuvx runs the command named after the package; if the README names another one, use uvx --from magpie-search <command>
Platform CLI
codex mcp add magpie-search -- uvx magpie-searchuvx runs the command named after the package; if the README names another one, use uvx --from magpie-search <command>
Config file
{
"mcpServers": {
"magpie-search": {
"command": "uvx",
"args": [
"magpie-search"
]
}
}
}uvx runs the command named after the package; if the README names another one, use uvx --from magpie-search <command>
Config file
{
"servers": {
"magpie-search": {
"type": "stdio",
"command": "uvx",
"args": [
"magpie-search"
]
}
}
}uvx runs the command named after the package; if the README names another one, use uvx --from magpie-search <command>
Platform CLI
gemini mcp add magpie-search uvx magpie-searchuvx runs the command named after the package; if the README names another one, use uvx --from magpie-search <command>
Step by step, with commands for each platform: How to add an MCP server to Claude Code, Codex, Cursor, VS Code, Gemini CLI and more
Questions about magpie-search
How do I install magpie-search in Claude Code?
Run claude mcp add magpie-search -- uvx magpie-search. The uvx launcher has to be on your system already.
What does magpie-search do?
It takes one query, sends it to conversation transcripts, local files, a knowledge graph, a vector store and the web at the same time, then returns a single ranked list with trust tiers attached.
Is it a RAG tool?
Only the retrieval half. There is no generator in it — it scores and orders hits and leaves any prose answer to the model you connect it to.
