Agent usage guidance

No API key or account is required today. Send a descriptive User-Agentheader identifying your client or agent — it isn’t required, but it helps if something needs debugging on either side, and there is no penalty for doing so.

Every tool endpoint is a pure function of its request body: same input, same output, every time, with no server-side state changed as a side effect. That means it is always safe to retry a request — there is no risk of a duplicate action or a partial write to clean up.

Branch on error.code, not on error.message— the code is the stable contract; the message is for humans and may change. See the API reference for the full set of codes each endpoint can return.

• Retry on 500 INTERNAL_ERROR or a network failure, with exponential backoff. Idempotency (above) makes this safe.
• Do not retry 400/422 responses (INVALID_REQUEST,INVALID_JSON, INVALID_JWT, etc.) without changing the request — the input was rejected deterministically and will be rejected again.
• 413 PAYLOAD_TOO_LARGE means split the work into smaller requests, not retry as-is.

No account-based rate limiting is enforced yet. Every endpoint does enforce a payload-size limit (see the API reference for current values), and the underlying infrastructure may throttle abusive traffic independent of any application-level policy. Don’t build automation that assumes unlimited throughput — see the Acceptable Use Policy.

Tool input is processed in memory and never persisted, logged, or sent to analytics — see the Security page for exactly what that guarantees. That said, treat any third-party API — this one included — the way you’d treat any network hop your data crosses: don’t send live credentials, private keys, or production secrets through a tool if you can substitute a redacted or synthetic example instead.

The same rules apply to agents as to humans — no automated abuse that degrades the service for others, no attempts to bypass payload-size limits, and no submitting content you don’t have the right to process. See the full Acceptable Use Policy.