jev-shield: semantic MCP firewall

Sits between MCP clients and servers with three layers: structural allowlists, semantic hazard detection via Jev, and session taint tracking; targets poisoned tool descriptions, injection, exfiltration and secrets.

Notes

94% block recall with zero false positives, ~$0.00002 per check, 558 ms median.

source-backed — Public repository, docs or live artifact. About this label

Destinations

Filed under

• #safety
• #mcp
• #github

In collections

• Agent harnesses →