# Jevx MCP

Run **[Jev AI](https://jevx.org)** typed decisions from any MCP client — Claude, Cursor, Cline, Codex. Give it a state and the questions you branch on, get back a choice, a score or a yes/no, with a calibrated probability on every option.

Jev AI is a zero-shot decision model from TypeSafe: text or JSON in, typed values out — no dataset, fine-tune or retrain when the label set changes. Jevx is an independent playground and API for Jev AI and is not affiliated with TypeSafe AI.

## Tools

| Tool | What it does |
|---|---|
| `decide` | Send a state and up to 16 typed questions, get back choices, scores and noul values with probabilities |
| `validate_questions` | Check a question's shape locally, free and without an API key, before spending credits on a malformed request |
| `open_in_jevx` | The URL to try or continue a decision in the browser playground |

## Install

```bash
npx jevx-mcp
```

### Claude Desktop / Claude Code

```json
{
  "mcpServers": {
    "jevx": {
      "command": "npx",
      "args": ["-y", "jevx-mcp"],
      "env": { "JEVX_API_KEY": "sk-..." }
    }
  }
}
```

Create a key at [jevx.org/settings/apikeys](https://jevx.org/settings/apikeys?utm_source=mcp). There is no anonymous decisions endpoint — a key is required, same as running the playground signed out.

### Cursor / Cline

Point the client at `npx -y jevx-mcp` (stdio transport) and set `JEVX_API_KEY` in its environment.

## Question shapes

| type | criteria | returns |
|---|---|---|
| `choice` | object of option to meaning, two or more options | `choice`, `probabilities`, `confidence` |
| `score` | ordered array, low to high | `score`, `probabilities`, `confidence` |
| `noul` | exactly the keys `true` and `false` | value between 0 and 1, where exactly `0.5` means the model declined |

The API answers HTTP 200 with a non-zero `code` on a refusal (bad key, malformed request); `decide` and `validate_questions` both surface that as a normal tool error, not a silent success.

## License

MIT