Verify, review & gate
Check whether a diff, document answer, or proposed action actually satisfies every stated acceptance criterion before committing or shipping.
Developer tooling · Open-source MCP server on PyPI Active system
Jev Judge MCP is an open-source Model Context Protocol server built by PyModel LLC and powered by TypeSafe AI’s Jev model. Published on GitHub and PyPI, it gives coding and automation agents typed verify, classify, rerank, review, score, and gate tools that map calibrated probabilities to clear policy actions.
Built for: Engineers building MCP agent workflows, CI review gates, and applications requiring calibrated natural-language judgment.
Calibrated judgment instead of self-grading
Autonomous coding and operations agents routinely fail at two boundaries: rigid regex rules break on legitimate natural-language variation, while asking a frontier LLM to self-grade its own diff wastes tokens and rubber-stamps subtle omissions.
PyModel LLC built Jev Judge MCP (github.com/PyModel/jev-judge-mcp) to expose TypeSafe AI’s Jev judgment model as clean, typed MCP tools—verify, screen, find, classify, rerank, decide, compare, extract, review, gate, and score. Each call evaluates explicit evidence against a stated criterion and maps calibrated probabilities into deterministic policy decisions (auto, review, or escalate), giving Claude Code, Codex, Cursor, and custom pipelines an independent gate before acting.
Open-source MCP server installable from PyPI (jev-judge-mcp) and GitHub.
Check whether a diff, document answer, or proposed action actually satisfies every stated acceptance criterion before committing or shipping.
Route tickets, triage inputs, and moderate natural-language payloads using typed labels and calibrated confidence thresholds.
Score retrieved passages for true relevance and extract structured fields without spinning up heavy frontier reasoning calls.
Translate calibrated probabilities into repeatable operational boundaries so uncertain edge cases escalate cleanly to a human owner.
Verification pattern comparison
When to use calibrated small-model MCP judgment tools in agent and application pipelines.
| Evaluation Property | Jev Judge MCP (TypeSafe AI Jev) | Frontier LLM Self-Grading Prompts | Regex / Keyword Heuristics |
|---|---|---|---|
| Independence | Separate evaluator model checking evidence against explicit criteria | Same authoring context grading its own reasoning and assumptions | Independent of LLM, but blind to semantic meaning |
| Output contract | Typed labels, calibrated probabilities, and auto / review / escalate policy actions | Verbose prose or uncalibrated "looks good" self-assessments | Binary pattern match with high false-positive/negative rates on prose |
| Latency & token profile | Purpose-built judgment calls designed for fast, repeatable gating loops | High token cost and multi-second latency per verification turn | Sub-millisecond, suitable only for strict syntactic formats |
| MCP host integration | Drop-in stdio MCP server on PyPI for Claude Code, Codex, and Cursor | Ad-hoc custom prompt templates duplicated across repos | Custom glue scripts per repository |
Direct answers
Architecture write-ups, adjacent products, and delivery services from PyModel LLC.
Share your current architecture, data boundaries, and target outcome with our engineering team.