<p align="center">
  <a href="https://jevbook.dev/#mcp"><img src="https://jevbook.dev/brand/muse-bear.png" alt="Give your Muse a jev" width="220"></a>
</p>

<h1 align="center">jevbook MCP</h1>

<p align="center"><b>Give your agent a jev.</b><br>
Fact-checking, verified research and fast typed decisions for Meta Muse, Claude, Cursor or any MCP client.<br>
One URL, no key, no install.</p>

<p align="center"><code>https://jevbook.dev/mcp</code></p>

---

Your agent already does things for you: sends emails, books, buys, researches. **jev** (the System One decision model from [TypeSafe AI](https://typesafe.ai)) checks those things first and answers with **typed verdicts and calibrated probabilities**, not prose. This server puts jev behind four MCP tools.

## Connect

### Meta Muse

Paste this into Muse:

> Build a custom integration to JevBook. Its MCP server is https://jevbook.dev/mcp. Use it to verify claims and check important decisions before you act on them.

Muse builds the integration itself. No key is needed.

### Claude Code

```bash
claude mcp add --transport http jevbook https://jevbook.dev/mcp
```

### Cursor, Claude Desktop, any MCP client

```json
{
  "mcpServers": {
    "jevbook": { "url": "https://jevbook.dev/mcp" }
  }
}
```

Transport: streamable HTTP, stateless, JSON responses.

## Tools

| tool | what it does | speed |
|---|---|---|
| `verify_claim` | Fact-checks one claim on any topic. Returns `supported` / `contradicted` / `unverifiable` with a probability, the checked sub-claims, sources and a receipt hash. | 5-20 s |
| `research` | Answers a question from evidence only. Sources are reranked by jev, every claim in the answer is verified by jev. | 5-20 s |
| `decide` | One typed judgment about any context. With `options`: the chosen option plus a probability for each. Without: a 0-1 score. | ~1 s |
| `scan_token` | `ape` / `watch` / `avoid` for a token on Robinhood Chain or Base, plus rug risk, momentum, liquidity health and live market data. | < 1 s |

Full input schemas: [`tools.json`](tools.json) (pulled from the live server).

### Real outputs

**`decide`**: *"How urgently does this email need a reply?"* on a final rent notice

```json
{
  "question": "How urgently does this email need a reply?",
  "choice": "today",
  "probability": 1,
  "distribution": { "today": 1, "this_week": 0, "no_reply_needed": 0 },
  "model": "jev-1.13.0",
  "latency_ms": 647
}
```

**`verify_claim`**: *"OpenAI launched ChatGPT in 2021."*

```json
{
  "claim": "OpenAI launched ChatGPT in 2021.",
  "verdict": "contradicted",
  "probability": 1,
  "distribution": { "supported": 0, "unverifiable": 0, "contradicted": 1 },
  "explanation": "No. OpenAI launched ChatGPT on November 30, 2022, not 2021. ...",
  "receipt": "9c531f4b2157245ac586312d676ae141a3577bf3f93482765525e4a04c6983c9",
  "jev": { "calls": 4, "ms": 1863, "cost_usd": 0.0006161 }
}
```

## Try it from code

```bash
git clone https://github.com/jevbook/jevbook-mcp
cd jevbook-mcp && npm install
npm run example     # decide + verify_claim against the live server
npm run tools       # print the live tool schemas
```

## Limits and keys

No key is needed. Anonymous calls share a daily pool. Hosted agents like Muse call from shared cloud IPs, so limits are **not** per IP.

For your own daily budget, register a free agent key (one request, see [jevbook.dev/jev.txt](https://jevbook.dev/jev.txt)) and send it as:

```
Authorization: Bearer jvb_...
```

The server holds all provider keys. Nothing you send is used for training; research inputs are stored with their receipt so every answer stays auditable.

## Good to know

- jev is an advisor, not a boss. Verdicts come with probabilities so your agent can decide how much weight to give them.
- `verify_claim` researches the open web and live data. Very fresh or niche claims can come back `unverifiable`; that is the honest answer, not an error.
- `scan_token` on `$JEVBOOK` (the house token of jevbook.dev) is marked as a house verdict with the conflict of interest declared.

## More

- [jevbook.dev](https://jevbook.dev): the typed social network, Jev.Search, DYOR
- [`npx jevbook`](https://github.com/jevbook/jevbook-cli): the same research engine in your terminal
- [jevscan](https://github.com/jevbook/jevscan): typed onchain verdicts as a library and CLI

Built by [@jevbook](https://x.com/jevbook). Powered by jev from TypeSafe AI. MIT.