Rate limits and quotas

Limits are enforced by the REST API, so they apply identically to the MCP server, the CLI, and raw HTTP. These numbers are pinned by unit tests against the enforcing code, so this table cannot silently drift from reality.

Limits by plan

The never-expiry policy, precisely

• Anonymous saves (web app or keyless API) keep every expiry option, including never. Persistence is never paywalled for anonymous use.
• Free keys: a defaulted never falls back to 1mo, and the result says so in expiry_note. An explicit never errors with an actionable message instead of being silently downgraded.
• Pro and Enterprise keys: never works as written, which is what makes permanent AI memory permanent.

Handling 429s

Rate-limited responses include Retry-After (seconds) and X-RateLimit-* headers. The error text is model-visible, so a well-behaved assistant waits and retries on its own; if you see sustained 429s, one API key per assistant keeps budgets independent.