429 until the window resets.
Limits by Plan
Each window lasts 60 seconds.
Every request that passes authentication and the plan check counts, including requests refused because the key can’t reach that project or lacks the endpoint’s scope.
Agent-run limit
These endpoints use the agent-run limit instead of the request limit, and don’t count toward the request limit:
MCP tools that start the same work use the same agent-run limit. MCP code mode also has its own execution allowance; see Code mode.
Who shares a limit
- API keys. Each key has its own limits; two keys never share a window. A key keeps the limits of the plan the account had when the key was created, and changing plan doesn’t update existing keys. After an upgrade, create a new key to get the higher limits, then revoke the old one. MCP requests made with a key count toward that key’s limits.
- CLI and MCP sign-ins. Requests made through MCP sign-in or after
devtune loginare limited per member and account, at the account’s current plan. All of one member’s CLI and MCP sign-ins on an account share the same limits, including MCP code mode.
Rate Limit Headers
Responses include headers that show where you are in the current window. On agent-run endpoints they report the agent-run limit. The429 returned for a sixth new-project draft within an hour on POST /projects/{projectId}/setup/new-project-draft reports the draft window. A 401, or a 403 for plan, project or scope, has no rate-limit headers.
Example Response Headers
When You Exceed the Limit
If you exceed your rate limit, the API returns a429 Too Many Requests response:
Retry-After with the number of seconds to wait before retrying. See Error handling for retry hints.
Best Practices
- Check the headers before making rapid requests. Use
X-RateLimit-Remainingto pace your calls. - Use exponential backoff when you receive a 429. Wait the seconds in
Retry-Afterbefore retrying. - Cache responses where possible. Most DevTune data updates once per day (after scheduled tracking runs), so caching for several minutes is usually appropriate.
- Batch your reads rather than polling frequently. Pull data once and process it locally.
Related Documentation
- Authentication - API key setup
- Error Handling - All error responses