> ## Documentation Index
> Fetch the complete documentation index at: https://apie.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# MCP enforcement

> Walkthrough — Enforce mode and approval routing through the MCP proxy.

You use MCP tools through Cursor or Claude Desktop. You want unknown tools to require human approval before executing — without changing your agent code. This recipe configures the MCP proxy in Enforce mode with approval routing.

**Source:** [mcp-proxy-enforcement-loop.ts](https://github.com/apie-sh/javascript-sdk/blob/main/examples/mcp-proxy-enforcement-loop.ts)

## Scenario

A release gate orchestrator uses an internal CI/CD MCP server. Unknown tools should require approval. Known risky tools should be blocked in production.

## Step 1 — Enable approval guardrail in dashboard

In the Apie dashboard, enable `require_approval_unknown_mcp_tools` (or your workspace equivalent) for the agent.

## Step 2 — Configure Enforce mode

```json apie.mcp.json theme={null}
{
  "agentKey": "release-gate-orchestrator",
  "serverName": "internal-cicd",
  "mode": "enforce",
  "approvalTimeoutMs": 600000,
  "environment": "production",
  "upstream": {
    "command": "node",
    "args": ["examples/mcp-upstream-fixture.mjs"]
  }
}
```

| Field                       | Why                                 |
| --------------------------- | ----------------------------------- |
| `mode: "enforce"`           | Enforce blocks and approvals        |
| `approvalTimeoutMs: 600000` | 10-minute approval window           |
| `environment: "production"` | Tags events for production policies |

## Step 3 — Point MCP host at proxy

```json theme={null}
{
  "mcpServers": {
    "internal-cicd": {
      "command": "npx",
      "args": ["@apie/cli", "mcp", "proxy", "--config", "apie.mcp.json"]
    }
  }
}
```

## Step 4 — Trigger an unknown tool

Call an MCP tool your agent hasn't used before. In Enforce mode:

1. Proxy evaluates the guardrail
2. Decision is `require_approval`
3. Approval request appears in the dashboard
4. Agent waits until you approve or reject

### What you'll see

* **Approved:** Tool call proceeds, telemetry events in dashboard
* **Rejected:** MCP host receives JSON-RPC error `-32002`
* **Blocked:** MCP host receives JSON-RPC error `-32001`

## Validate setup

<CodeGroup>
  ```bash TypeScript theme={null}
  npx apie doctor --mcp --mcp-config apie.mcp.json
  ```

  ```bash Python theme={null}
  apie doctor --mcp --mcp-config apie.mcp.json
  ```
</CodeGroup>

## Error codes

| Code     | Meaning                         |
| -------- | ------------------------------- |
| `-32001` | Guardrail blocked the tool call |
| `-32002` | Approval denied or timed out    |

See [Errors reference](/reference/errors).

## Next steps

<CardGroup cols={2}>
  <Card title="MCP proxy" icon="plug" href="/mcp/proxy">
    Full proxy setup guide.
  </Card>

  <Card title="Human approval" icon="user-check" href="/guardrails/human-approval">
    Approval flow details.
  </Card>
</CardGroup>