MCP

Jev MCP server

An MCP endpoint that lets your agent ask Jev typed questions: a yes/no probability, a pick from a closed set, or a score on your own levels. JEV Hub runs it; it uses your own TypeSafe API key, sent as a header with each request.

Endpoint

URLhttps://jevhub.org/api/mcp
TransportStreamable HTTP, JSON responses, no sessions
AuthAuthorization: Bearer <TYPESAFE_API_KEY>
Protocol versions2025-06-18, 2025-03-26, 2024-11-05
Get a keyconsole.typesafe.ai/keys

Add it to your client

Claude Code
claude mcp add --transport http jev https://jevhub.org/api/mcp \
  --header "Authorization: Bearer $TYPESAFE_API_KEY"
Cursor · ~/.cursor/mcp.json
{
  "mcpServers": {
    "jev": {
      "url": "https://jevhub.org/api/mcp",
      "headers": { "Authorization": "Bearer ${env:TYPESAFE_API_KEY}" }
    }
  }
}
Codex · ~/.codex/config.toml
[mcp_servers.jev]
url = "https://jevhub.org/api/mcp"
bearer_token_env_var = "TYPESAFE_API_KEY"

Export TYPESAFE_API_KEY in the shell that starts your client. Any other MCP client that speaks Streamable HTTP and can send a header works the same way.

Tools

ToolAsksReturns
jev_noulOne yes/no questionnoul: probability of yes, 0 to 1
jev_choicePick one of your options (a list, or names mapped to descriptions)choice, a probability per option, confidence
jev_scorePlace the state on 2 to 10 ordered levels, lowest firstscore (can land between levels), per-level probabilities, confidence
jev_evaluateSeveral typed questions about one state, in the official questions mapEvery answer under your question ids

Each tool takes a state (text, or a JSON object or array) and an optional model, which defaults to jev-latest. One tool call is one systemone request.

Try it from a terminal
curl -s https://jevhub.org/api/mcp \
  -H "content-type: application/json" \
  -H "Authorization: Bearer $TYPESAFE_API_KEY" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{
        "name":"jev_choice",
        "arguments":{
          "state":"Agent wants to run: rm -rf ./dist && git push --force origin main",
          "question":"How risky is this command for the repository?",
          "options":{"allow":"Local, reversible","review":"Changes shared state","block":"Destroys work or history"}
        }}}'
Tool result (structuredContent)
{
  "model": "jev-1.13.0",
  "answer": {
    "type": "choice",
    "choice": "block",
    "probabilities": { "allow": 0.01, "review": 0.18, "block": 0.81 },
    "confidence": 0.74
  },
  "usage": { "input_tokens": 212, "output_tokens": 9 }
}

The shape is the live tool result; the numbers above are illustrative. Run it with your key to see real ones.

When the agent should call it

Tell the agent in its instructions when to ask. For example: "Before any command that deletes files, pushes, or touches production, call jev_choice with allow / review / block and ask me unless it returns allow with confidence above 0.7." More decision points are on Jev inside your agent.

Key handling and limits

  • Your key is read from the request header, used for one call to api.typesafe.ai, and dropped. JEV Hub does not store or log keys, states, or answers.
  • Usage bills to your TypeSafe account. TypeSafe's rate limits apply; a 429 or 529 comes back as a tool error that says to retry.
  • Without a key, the server still answers initialize and tools/list, so clients can show the tools. Tool calls explain how to add the header.
  • This is an independent community server, not a TypeSafe product. Official docs: docs.typesafe.ai.