Open-source MCP server compresses agent context: 93.8% recall vs 51.9% baseline

Beginning_Note_5385 · reddit · 2026-09-09

A developer released cap-shield-mcp (MIT), an MCP server for selective context selection and compression aimed at agent-to-agent traffic (context, memory, identity handoffs). Measured Top-10 recall on LongMemEval-S: 93.8% vs 51.9% lexical baseline, with all measured and unmeasured metrics published honestly — including one package where 28% of messages came out larger after compression. It offers account-free online testing (512KB max), versioned memory tools, and short-lived signed passes so receiving agents can verify identity. Setup via Claude custom connector is documented; the author is soliciting real-traffic test feedback.

Original post →

More from coding & agent

coding & agent channel →