Outborn Agent Platform · Technical reference
Errors and retries
Interpret HTTP status and JSON-RPC errors before retrying; verify the outcome of uncertain writes.
Applied guidance
Best practices for errors and retries
Recommended
- Read HTTP status and the JSON-RPC error envelope together.
- Correct invalid input or denied access; reauthorize expired credentials through the supported flow.
- Retry transient failures only for safe operations with a limit and backoff.
Avoid
- Do not retry a write after an uncertain outcome until you have checked whether it already completed.
01
Read both error layers
The MCP endpoint returns JSON-RPC errors with jsonrpc, id, and error.code/message, alongside an HTTP status. Invalid JSON-RPC requests use -32600 and HTTP 400; unknown methods use -32601 and HTTP 404; an unknown tool uses -32602 and HTTP 400; an unknown resource uses -32002 and HTTP 404. Tool failures use -32000 with the underlying HTTP status.
{ "jsonrpc": "2.0", "id": 3, "error": { "code": -32602, "message": "Unknown tool: missing.tool" }}02
Authentication and access
Authentication failures use JSON-RPC -32001 with the HTTP status. A 401 also includes WWW-Authenticate with the protected-resource metadata URL. A 403 can indicate missing required scopes or an organization/access decision. Reauthorize or correct the grant; do not retry with another user's token.
03
Retry only safe operations
Input errors, denied access, and unknown methods or resources need correction. A transient provider or runtime failure may be retried with a bounded backoff when the operation is safe to repeat. For Space updates or removal, first read the current state after a timeout or uncertain response to avoid duplicate or unintended mutations.
Related documentation