Errors and limits

The MCP server uses the same API keys, sign-in rules and limits as the REST API, so the same ideas apply.

How the server answers

The server speaks Streamable HTTP in stateless JSON-RPC. Each POST carries one JSON-RPC message, or a batch of up to 20, and gets a JSON reply. Notifications get 202 with no body. There is no session and no event stream, so GET and DELETE answer 405.

Errors

Problems before the request reaches the MCP protocol use the HTTP status and the same JSON error body as the REST API, with a request_id.

Status	Code	What to do
401	missing_api_key or invalid_api_key	Create a new key and update the client.
413	payload_too_large	Send a smaller request. The limit is 256 KB.
429	rate_limited	Wait for Retry-After seconds.
500	internal_error	Try again in a moment.

Protocol mistakes are JSON-RPC errors: -32700 for a body that is not JSON, -32600 for an invalid request, -32601 for an unknown method, -32602 for invalid parameters or an unknown tool, and -32002 for a resource that is not found.

When a tool itself fails, because an argument is wrong, the record does not exist, the key lacks the scope, the module is switched off or a limit was hit, the assistant gets a normal result marked isError with a plain message it can read to you. Ask it to try again, or to explain what it needs.

Limits

• 120 requests a minute per key. REST and MCP requests made with the same key share this.
• ask is limited to 10 a minute per key, on top of the 120.

Limits are counted per server instance, so they are best effort. Long searches across many conversations are best done in several short steps. See rate limits for the headers and the full explanation.

Logs

Tirver records every request, including rejected ones, with the key prefix, the tool name and the status, so problems can be investigated. It does not record arguments, results or the key itself.