# Check your rule (Control › Rules) for a protected MCP tool before it runs

`POST /api/v1/mcp/decide` · operationId `decideMcpCall`

For developers: your server sends this request with a server key; it is not a button in the dashboard.

The AI agent branch of the tool's saved rule decides (every MCP caller is an agent's client); with no saved rule the defaults apply (AI agent: hold for review). A medium-confidence call that the rule would deny is held for review instead unless the site blocks medium confidence. A valid agent token issued with mcpTools turns a hold into allow; a held call approved in the dashboard or by POST /api/v1/mcp/decisions/{id}/step-up is allowed once within 15 minutes with approvedDecisionId. A hold is returned as decision "review" with a decisionId: the dashboard shows it as Held for review; poll GET /api/v1/mcp/decisions/{decisionId} or settle it with your own step-up. Leave client out and confidence is medium. Not authorization.

## Authentication

Send `Authorization: Bearer YOUR_SERVER_KEY` (siteServerKey: Site server key in Authorization: Bearer <key> (the scheme is case-insensitive and required; a bare key is 401). Shown once on creation or rotation. Publishable keys cannot authenticate). Keep the server key in server-only configuration; never send it to a browser.

## Request body

Content-Type: application/json (required)

- `tool` (string, required)
- `client` (object, optional)
- `client.name` (string, required) · maxLength 64
- `client.version` (string, optional) · maxLength 32
- `accountId` (string, optional) · maxLength 512
- `userId` (string, optional) · maxLength 512
- `agentToken` (string, optional)
- `approvedDecisionId` (string, optional)

```json
{
  "tool": "create_invoice",
  "client": {
    "name": "claude-ai",
    "version": "1.0"
  },
  "accountId": "acct_456",
  "userId": "usr_123"
}
```

## Responses

- **200**: Success · body `object`
- **400**: bad_request: the request is invalid; message says what to fix and field names the request field · body `Error`
- **401**: unauthorized (missing or invalid server key, missing Bearer scheme) · body `Error`
- **403**: forbidden: the call came from a browser page or another Site; call from your server with this Site’s server key · body `Error`
- **404**: not_found: unknown path, or an ID in the path that is not on this Site · body `Error`
- **409**: conflict: approvedDecisionId was already used, has expired, or belongs to another tool or account · body `Error`
- **413**: payload_too_large: the request body exceeds 131072 bytes · body `Error`
- **429**: rate_limited: wait for Retry-After · body `Error` · headers `Retry-After`
- **500**: internal_error: fail the protected action; do not retry automatically · body `Error`
- **503**: service_unavailable: temporarily unavailable; only verify retries · body `Error`

### Success body fields

- `decision` ("allow" | "review" | "deny", required)
- `basis` ("rule" | "default_rule" | "no_rule" | "agent_token" | "approval", required)
- `confidence` ("high" | "medium", optional): From the declared client: high for a known AI app, medium for debugging tools (MCP Inspector) and unknown or missing clients.
- `client` (object | null, optional)
- `heldInsteadOfDeny` (boolean, optional)
- `decisionId` (string, optional)
- `tool` (string, optional)
- `accountId` (string | null, optional)
- `userId` (string | null, optional)

Failures are not verdicts. Error bodies follow the `Error` schema: `code` is one of `bad_request`, `unauthorized`, `session_reinit_required`, `forbidden`, `not_found`, `conflict`, `identity_required`, `payload_too_large`, `rate_limited`, `internal_error`, `service_unavailable`, `request_failed`, `approval_invalid`, `approval_used`, `approval_expired`, `approval_mismatch`, `step_up_not_held`, `step_up_mismatch`, `step_up_settled`, `step_up_expired`, `step_up_superseded`, `agent_tokens_disabled`, `tool_not_approved`, `tool_not_read_only`. [Errors and retries](/docs/errors) explains each code and what to do.

## Example

```sh tab=curl
curl -sS -X POST 'https://app.trytelltale.dev/api/v1/mcp/decide' \
  -H "Authorization: Bearer $TELLTALE_SERVER_KEY" \
  -H 'Content-Type: application/json' \
  --data '{"tool":"create_invoice","client":{"name":"claude-ai","version":"1.0"},"accountId":"acct_456","userId":"usr_123"}'
```

```js tab=fetch
const response = await fetch("https://app.trytelltale.dev/api/v1/mcp/decide", {
  method: "POST",
  headers: {
    Authorization: 'Bearer ' + process.env.TELLTALE_SERVER_KEY,
    'Content-Type': 'application/json'
  },
  body: JSON.stringify({
    "tool": "create_invoice",
    "client": {
      "name": "claude-ai",
      "version": "1.0"
    },
    "accountId": "acct_456",
    "userId": "usr_123"
  }),
  signal: AbortSignal.timeout(5000)
});
const result = await response.json();
```

Machine-readable contract: [openapi.json](https://app.trytelltale.dev/openapi.json). Request builder: the [API explorer](/docs/api-explorer) shows every operation as curl; it sends requests only on a local dashboard.