Magpie-search cuts AI agent token costs to 1/2000th

What happened

A developer has released Magpie-search, a federated search engine that runs entirely on a user's machine and searches across five sources—AI conversation transcripts, local files, a knowledge graph, a vector store, and the live web—using keyword, semantic, and exact-match modes. Results are ranked by trust tier (fact, reference, lead, stale) and deduplicated across sources.

Why it matters

For AI agents, losing context after a crash or power outage is costly. Magpie indexes everything an agent has worked through locally, so recovery is possible. The tool also reduces token cost dramatically: a deep research question that would cost a multi-agent swarm ~2,000,000 tokens costs Magpie's deepweb mode ~1,050 tokens—about 1/2000th the cost—because searching and page-reading are pure retrieval with zero model tokens.

What to watch

Magpie is installed via pip install magpie-search and runs on Python 3.10+ on Windows, macOS, and Linux. It connects to any MCP-capable agent (a standard protocol for AI tools) and strips ~30 classes of secrets at ingest before indexing. The optional session summarizer requires Ollama, which is free and runs locally.