# VerifiedX Claude + MCP compliance agent+human starter

Protect a composed compliance workflow delivered through MCP without rebuilding the agent loop.

This starter mirrors a real compliance or risk workflow:

- Claude inspects the account record
- Claude checks enhanced due diligence state
- Claude attempts to clear an active sanctions / EDD alert
- VerifiedX blocks the alert-clear because the evidence does not justify it
- a human analyst approval is added to upstream workflow context
- Claude retries the same high-impact action and the clear is allowed on the second pass

## Run it

```bash
npm install
npx @verifiedx-core/sdk doctor
```

Set env:

```bash
export ANTHROPIC_API_KEY=...
export VERIFIEDX_API_KEY=...
export VERIFIEDX_BASE_URL=https://api.verifiedx.me
```

Run:

```bash
npm start
```

What to check:

- `artifacts/scenario_report.json`
- `artifacts/verifiedx_diagnostics.jsonl`

You should see two phases in `artifacts/scenario_report.json`:

- `blocked_attempt` shows the unsafe alert clear getting blocked
- `approved_retry` shows the same workflow succeeding once a human analyst approval is present upstream