jev-mcp
Use when evaluating typed semantic judgments with TypeSafe Jev / System One — Choice, Score, or Noul questions against text or structured state; also when reranking candidates, verifying evidence supp
SKILL.md.claude/skills/jev-mcp/SKILL.md
GitHubJev MCP (System One)
Prefer MCP tools from jev-rw-systemone when connected. Each call hits the paid TypeSafe API —
batch related questions in one jev_system_one request when possible.
Jev is not a generative LLM and not a vision model. state must be a string, JSON object,
or array of strings. Preprocess images/audio into text fields before calling.
Tools
| Tool | Use when |
|---|---|
jev_choice |
Pick one label from a closed set |
jev_score |
Score along ordered criteria levels |
jev_noul |
Yes/no probability for one proposition |
jev_system_one |
Multiple typed questions against the same state |
Default model pin is jev-1.13.0 (override only when the user asks). Keep workflow, side effects,
and deterministic rules in normal code; give Jev narrow semantic judgments only.
How to call
- Confirm
jev-rw-systemoneis connected. If not, say so and stop — do not invent probabilities. - Build compact
statewith only the fields the questions need. - Prefer one
jev_system_oneover many single-tool calls. - Read
ok/errorenvelopes. Onmissing_credential, tell the user to setJEV_TOKEN(orTYPESAFE_API_KEY); never invent keys. - Return the selected answer fields (
choice/score/noul),confidence/probabilitieswhen present, and themodelused. Do not dump secrets or raw SDK internals.
Harness decision points (automatic)
This repo's Claude Code hooks call the sibling backend harness
(../image-scoring-pipeline/scripts/agent_harness), which asks Jev per-turn questions through the
backend's modules/typesafe client: which rule packs a request sees, whether a script may run, whether
files may go to an external reviewer, whether a subtask can leave the main model, and whether a subgoal
is a duplicate. Modes live in .agent/jev_harness.json. See
JEV_AGENT_HARNESS.md.
Use the jev-rw-systemone MCP tools for ad-hoc judgments; leave those decision points to the harness.
Boundaries
- Do not treat Jev as a chat model, coder, or image judge.
- Do not bypass MCP by pasting API keys into prompts or tool args.
- Canonical server lives at
D:\Projects\jev-mcp(AGENTS.md,README.md).