Jev for agents: remote MCP server and agent skill

Remote MCP server: four typed tools

jevmodel.org runs a remote MCP server at https://jevmodel.org/mcp over Streamable HTTP. It exposes jev_decide (the full contract — state plus up to 8 typed questions in one call) and three convenience tools for the common cases: jev_noul for yes/no probabilities, jev_choice for labelled options, jev_score for rubric scores.

Authentication is your normal jevmodel.org API key as a Bearer header — no separate credential, same token balance, same 120 requests/minute limit.

# Claude Code
claude mcp add --transport http jev https://jevmodel.org/mcp \
  --header "Authorization: Bearer $JEVMODEL_API_KEY"

# Cursor / other clients — server config
{
  "mcpServers": {
    "jev": {
      "url": "https://jevmodel.org/mcp",
      "headers": { "Authorization": "Bearer sk-..." }
    }
  }
}

The agent skill

Read the raw SKILL.md ↗ — endpoint, request shape, playbook and limits.

MCP hands your agent the tools; the SKILL.md teaches it the judgment patterns: when a typed decision beats an LLM call, how to batch questions into one request, how to threshold a noul probability, and how to write criteria that survive production. Drop the file into your agent’s skills directory and it is picked up automatically.

# Claude Code / agents
curl -o .claude/skills/jev/SKILL.md --create-dirs \
  https://jevmodel.org/skills/jev/SKILL.md

# or Devin-style layouts
curl -o .devin/skills/jev/SKILL.md --create-dirs \
  https://jevmodel.org/skills/jev/SKILL.md

A worked example: a guardrail before a tool call

The pattern agents use most: before executing something risky, ask Jev for a typed judgment on the proposed action and threshold the probability.

{
  "state": {
    "action": "bash",
    "command": "rm -rf ./dist && pnpm build",
    "cwd": "/srv/app",
    "recent_errors": 0
  },
  "questions": {
    "allow": {
      "type": "noul",
      "instructions": "Is this command safe to run without asking the user? Destructive paths outside the project, credential files or force-pushes are not safe."
    }
  }
}

# → answers.allow.noul ≥ 0.9: run it
# → 0.6–0.9: show the command and ask
# → < 0.6: block

Billing and limits

• Same balance as the playground and /v1/systemone — 100,000 free input tokens at sign-in, packs from $9.90 that never expire
• Input tokens only; output is free; failed upstream calls refund automatically
• 1–8 questions per jev_decide call, ~8,000 character state, 120 calls/minute per key
• MCP initialize and tools/list are unauthenticated; only tools/call uses your key and balance