/api-reference/rate-limits.json, and as an OpenAPI Overlay that adds each operation’s limits to the OpenAPI as x-rate-limits.
The limits
On the platform-wide limit only responses with a status of 400 or above count toward the ceiling; successful requests are not counted. Health checks and
/.well-known/* discovery documents are exempt.
The sign-in ceiling can be tightened or relaxed by the operator within fixed bounds, so read the RateLimit-* headers rather than hard-coding 200.
The Integration API and MCP per-tool limits are counted by each server process, so under load the effective ceiling can be somewhat higher than the number shown. Do not rely on that headroom.
MCP per-tool limits
Tools not listed have no limit of their own beyond the MCP limit.
Headers
The platform-wide and MCP limits send these on every response they count:
The sign-in, chat, undo and history limits send the standard
RateLimit-Policy, RateLimit-Limit, RateLimit-Remaining and RateLimit-Reset headers instead (RateLimit-Reset is in seconds).
Every 429 carries Retry-After, in seconds.
When you hit a limit
You get429 Too Many Requests. On the platform-wide limit the body looks like this:
/mcp). When the block ends you get a fresh window.
On /mcp the 429 keeps its HTTP status and carries a JSON-RPC error so MCP clients can read it:
tools/call with a JSON-RPC error (code 429, data.tool, data.retryAfter, data.limit, data.window).
The Integration API answers { "success": false, "error": "Rate limit exceeded for integration API", "errorType": "RATE_LIMITED" }.
Backing off well
- Wait at least
Retry-Afterseconds before the next attempt. - If you retry a write, send the same
Idempotency-Keyso the retry cannot create a duplicate. - Spread bulk work out. The Integration API takes up to 500 transactions in one
POST /transactions, which is gentler than 500 single requests.
