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

GitHub

Jev 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

  1. Confirm jev-rw-systemone is connected. If not, say so and stop — do not invent probabilities.
  2. Build compact state with only the fields the questions need.
  3. Prefer one jev_system_one over many single-tool calls.
  4. Read ok / error envelopes. On missing_credential, tell the user to set JEV_TOKEN (or TYPESAFE_API_KEY); never invent keys.
  5. Return the selected answer fields (choice / score / noul), confidence / probabilities when present, and the model used. 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).