jev-mcp

Canonical version: jev-mcp.

jev-mcp is an open-source MCP server by jkudish that turns Jev into twelve judgment tools any MCP client can call (Claude Code, Codex, OpenCode, Amp...). Each call returns typed probabilities (and usually a confidence score) in roughly 150 to 500 ms, for a fraction of a cent.

The pitch is about the checks agents skip. A frontier model is too slow and too expensive to verify every claim, page or candidate list, so agents just... don't. At Jev prices, you can afford to check everything.

The tools:

• jev_verify: checks each claim in a report or PR description against its sources (verified, contradicted, unsupported)
• jev_screen: judges fetched text BEFORE it enters context: probability of prompt injection, substance, relevance. It recommends pass, review, block or skip, but never blocks on its own
• jev_noul: calibrated probability for up to 64 propositions per call
• jev_find and jev_rerank: rank candidates against a plain-language query, with no embeddings and no index to maintain
• jev_classify, jev_decide and jev_compare: batch labelling, choosing between alternatives (with an "ask the user" escape hatch), and judging how two passages relate
• jev_extract and jev_audit: pull values (prices, dates, IDs) with regex plus judgment, then audit them against the source (Candidate-Then-Select Extraction)
• jev_review and jev_gate: score a diff for correctness, spec match, test gaps and blast radius, and check "tests pass" claims against the evidence before a task is called done

What I like is the defensive design. Malformed answers "fail closed": the claim gets unknown and is flagged for review instead of being silently accepted. Thresholds are explicit parameters (auto_accept defaults to 0.8, block_at to 0.75). And jev_verify follows TypeSafe's own Atomic Question Decomposition advice: it asks a separate "is this evidence even about the same thing?" question, and a contradiction only stands if the answer is yes. That's one more question in the same request, not another API call.

It runs over stdio by default, or as a stateless HTTP server for a team (with a required bearer token, since every call spends your Jev key). It also ships an agent skill that teaches coding agents WHEN to reach for each tool, which addresses a real problem: MCP tools that sit registered but never get called.

The author calls it early software, and the accuracy figures in the repo come from contributors' private experiments, not independent benchmarks. Treat the thresholds as starting points and tune them on your own cases (Confidence-Gated Routing).

References

• jev-mcp on GitHub
• @jkudish/jev-mcp on npm

Related

• Jev
• Model Context Protocol (MCP)
• System One Models
• Confidence-Gated Routing
• Atomic Question Decomposition
• Candidate-Then-Select Extraction
• Prompt injection
• AI Guardrails

About Sébastien

Ready to get to the next level?