429 instead of a result.
Limits
How requests are counted:
- Per company. Requests count against the company of the API key, so every key in the company, and every tool that uses one (the CLI, the MCP server, Terraform), shares the same limits. Requests without a valid key are counted per IP address.
- Per service. In the service mutation and deployment tiers, each service has its own count: the service is the ID in the path, such as
{id}inPOST /applications/{id}/deployments. - Per operation. In the first five tiers, each operation has its own count: creating 20 applications in an hour doesn’t stop you from creating a database. All operations in the read tier share one count.
- Every request. A request counts when it arrives, before it’s checked, so requests that fail with
4xxcount too.
Response headers
Each response carries the state of the limit that applied to it:Handle 429 responses
A request over the limit returns429:
retry-after, then send the request again. If it fails again, wait longer each time and add jitter, so that parallel workers don’t retry in step:
retry.ts
- Watch
x-ratelimit-remainingand slow down before it reaches0. - Poll long-running operations every 10 seconds or more, not in a tight loop. See Long-running operations.
- Use webhooks to learn when a deployment ends, instead of polling.