# review-gate-mcp

An MCP server that **asks a human when it isn't sure**.

Most extraction demos guess. This one implements the pattern that makes
automation safe to run on documents that matter: fields extracted with high
confidence flow through; anything uncertain is **held in a review queue** for
a person to resolve - never silently guessed into your systems.

```
document ──▶ extract_with_gate ──▶ confident fields ──▶ your system
                     │
                     └──▶ uncertain fields ──▶ review_queue ──▶ a human ──▶ resolve_review
```

## Tools

| tool | what it does |
|---|---|
| `extract_with_gate` | extract `invoice_number`, `total_amount`, `date`, `email`, `vendor_name` from text; confidence below 0.85 goes to the queue instead of the output |
| `review_queue` | list items waiting for a human decision |
| `resolve_review` | a human supplies the verified value, closing the loop |

## Run it

```
npm install
npm test          # end-to-end over real MCP stdio: extract -> hold -> human resolve
```

Add to Claude Code / Claude Desktop / Cursor (any MCP client):

```json
{
  "mcpServers": {
    "review-gate": { "command": "node", "args": ["/path/to/review-gate-mcp/server.js"] }
  }
}
```

Then ask your agent to extract fields from a pasted document and watch what
it does with the smudged one.

## The point

The extractors here are deliberately transparent heuristics - swap in a
model-backed extractor and the gate stays identical. The pattern is the
product: **a system that escalates its hard cases beats one that guesses
them.** This is the standard RavnLab builds automation to; the evaluation
side of the same idea lives in
[ravnlab-eval-harness](https://github.com/ravnlab/ravnlab-eval-harness).

MIT. Built by [RavnLab](https://ravnlab.com).