Overview & result
jevwire wires Jev into an agent harness in three forms: seven MCP tools (rank, pick, verify, evaluate, gate an action, choose the next step, list models), an embeddable JevDecisionModel library so mandatory checks can live in the harness instead of in a tool the agent may decline to call, and a Claude Code plugin whose hooks place judgments at the harness boundary.
From the original post
Jev decision layer for agents: MCP server, embeddable DecisionModel library, and an escalate-only Claude Code plugin (TypeSafe AI's Jev)
How Jev fits in the loop
- Agents can call seven MCP tools to rank, pick, verify, evaluate, gate an action or choose a next step.
- Harness authors can embed the JevDecisionModel library so mandatory checks run in code.
- The Claude Code plugin adds hooks that judge tool calls at the harness boundary.
- A prefilter keeps most calls from needing a judgment at all.
How to reproduce
- Inspect the original showcase and source material at https://github.com/Brainwires/jevwire
- Verify the bounded prompt and input candidate schema configured for Jev
- Benchmark decision latency and classification accuracy against baseline models
Why this build matters
It separates checks an agent may choose to call from checks the harness always runs.
Limitations
- A judged call adds roughly half a second; the author measured 447–480 ms per hook judgment.
- The README calls it advisory, not a security boundary.
- It says Jev is not injection-hardened: text inside a tool input or fetched page can move its probabilities.