jev-shield
Semantic MCP firewall powered by Jev, screens every tool call, tool result, and tool description with calibrated System One verification.
What it does
Semantic MCP firewall powered by Jev, screens every tool call, tool result, and tool description with calibrated System One verification. 94% block recall, 0 false positives, ~$0.00002/check. Agent plumbing where probabilities beat regex on model output.
Fork the pattern, not the prose. Tags in the wild: agent tooling, mcp, guardrails. The repo has install steps, env vars, and recent commits if you are deciding whether to depend on it.
How Jev is used
Wraps any stdio MCP server, running System One verification on exfiltration, destructive args, injection in results, tool poisoning, and secrets before the agent sees poisoned content.
Setup: npx jev-shield -- <upstream command> with AI_GATEWAY_API_KEY or TypeSafe keys. Optional JSON policies and install-hooks for Claude Code, opencode, and Codex MCP paths.
Caveats: Each check adds roughly half a second to a second plus per-event cost documented in README demo (~$0.00003/event).
Probabilities drive policy
jev-shield returns the same semantic scores regardless of policy; you change thresholds in code review. That beats rewriting prompts when compliance wants stricter blocks.
FAQ
- Does jev-shield replace MCP server auth?
No. It is an inline semantic filter. You still scope credentials and upstream servers correctly.
- Which hosts are supported?
MCP wrap works anywhere you can swap the server command. Optional hooks cover Claude Code and opencode; Codex uses the MCP firewall path per README table.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.