Jev in Claude Code: MCP tools, skills and a code brief

Before you start

You need Claude Code and a JevStation account. New accounts get 200 free credits with no card. Create a key at Settings › API Keys and put it in your shell profile:

export JEVSTATION_API_KEY="sk_..."

Want to see what Jev returns first? The playground runs presets without an account.

Step 1: add the MCP server

Claude Code registers remote servers with claude mcp add. JevStation’s server uses the HTTP transport and a Bearer header:

claude mcp add --transport http jevstation https://jevstation.com/api/mcp \
  --header "Authorization: Bearer $JEVSTATION_API_KEY"

Because the header is in double quotes, your shell substitutes the key when you run the command, and Claude Code stores the resulting header in its config. That is why the variable must be set first.

Pick a scope. Claude Code has three:

To make Jev available everywhere:

claude mcp add --transport http --scope user jevstation https://jevstation.com/api/mcp \
  --header "Authorization: Bearer $JEVSTATION_API_KEY"

For a team, do not use the command with project scope — it would write your literal key into .mcp.json. Write the file by hand instead and let Claude Code expand the variable from each developer’s environment when it connects:

{
  "mcpServers": {
    "jevstation": {
      "type": "http",
      "url": "https://jevstation.com/api/mcp",
      "headers": { "Authorization": "Bearer ${JEVSTATION_API_KEY}" }
    }
  }
}

Claude Code asks each person to approve project servers the first time they are used.

Step 2: install the skills

Three Agent Skills teach Claude how to use Jev well:

• jev-api — request and answer shapes, errors, costs, the MCP tools.
• jev-question-design — how to write Choice, Score and Noul questions that give reliable probabilities.
• jev-guardrails — screening untrusted input and checking tool calls before they run.

Install them into ~/.claude/skills:

curl -fsSL https://jevstation.com/skills/install.sh | sh

The script only downloads three SKILL.md files. To install into one repository instead, set SKILLS_DIR:

curl -fsSL https://jevstation.com/skills/install.sh | SKILLS_DIR=.claude/skills sh

Claude Code picks up new skills in a running session and loads each one when its description matches the task. You can also call one directly, for example /jev-question-design.

Step 3: brief Claude for a code integration

MCP tools are for decisions Claude makes during a session. When you want Jev inside your own application — a support-ticket router, an answer grader, a guard in your agent loop — open the API docs and press Copy for AI. It copies a compact brief: endpoint, auth through JEVSTATION_API_KEY, question and answer shapes, error codes, costs and a verification checklist. Paste it into Claude Code with your request:

Here is the JevStation integration brief. Add a function to our ticket service that sends each new ticket to Jev with a team Choice and an urgency Score, and routes on the probabilities. Read the key from the environment and handle 402, 429 and 502.

The brief tells Claude that the request model field is ignored and that it should log data.model with each decision.

Prompts to try

With the server connected, these work as written:

• “List my Jev tools, then list my saved question sets.” — free; confirms the connection.
• “Use jev_decide to classify this review as positive, neutral or negative, and tell me the probabilities: …”
• “Before you run this migration, check it with jev_guard_tool_call. My request was: clean up old test data in staging.”
• “Use jev_route_model on each of these five prompts and tell me which ones need a frontier model.”
• “Draft a Jev question set for grading support replies, following the jev-question-design skill, then test it with jev_decide on these two replies.”
• “Run my saved question set qs_… on the text in samples/ticket.txt.”

Depending on your permission settings, Claude Code may ask before it calls a tool. Every evaluation reports credits.charged so you can watch spending.

Verify the setup

1. claude mcp list shows jevstation, and claude mcp get jevstation shows the URL.
2. Inside Claude Code, /mcp shows the server as connected with five tools. A failure usually means a 401 — see below.
3. Ask Claude to run jev_list_question_sets. It should succeed without spending credits.
4. Run one jev_decide call and check that the result has model, answers and credits.remaining.

If something goes wrong

• Failed to connect / 401. The header is missing or the key is wrong. Check echo $JEVSTATION_API_KEY, then claude mcp remove jevstation and add it again.
• A tool returns an error. Tool errors come back as readable results, for example not enough credits, a rate limit, an invalid question set or an evaluation that failed and was refunded. Claude can usually fix the input and retry. The MCP server page lists them.
• Leaked key. Delete it at Settings › API Keys, create a new one and re-run claude mcp add.

Never paste the key into a prompt or into tool arguments; keep it in the environment.

References

Syntax verified on 2026-09-28 against:

• Claude Code MCP: https://code.claude.com/docs/en/mcp
• Claude Code skills: https://code.claude.com/docs/en/skills