Rate limits
The per-minute cap and per-plan daily allowances for the AI agent, MCP tools, and API.
The two limits
Recited applies two limits to programmatic and agent usage:
A per-minute cap, same for everyone: 30 requests per minute. This exists so a runaway script or a misbehaving client stalls fast instead of burning a day's allowance in one loop.
A daily allowance, by plan, counted per project and reset on the UTC day:
| Plan | Requests per day |
|---|---|
| Launch | 300 |
| Scale | 1,000 |
| Advanced | 3,000 |
What counts against the allowance
The daily allowance is shared by everything agent-shaped on a project:
- chat turns with the in-app AI agent
- tool calls made through the MCP server by connected clients
Article API pulls with a project API key are not counted today. Treat that as room to build, not room to hammer: cache your pulls regardless.
One MCP tool call is one request. A single question in an MCP client can trigger several tool calls, so budget for that: an agent that checks five reports to answer one question spent five requests.
Daily scans, dashboards, and everything you do by hand in the app do not count. The allowance governs machines, not people.
What hitting a limit looks like
Requests over the limit are refused with a clear message rather than silently dropped. The in-app agent will tell you the allowance is spent; MCP clients receive an error result the model can read and relay. Per-minute refusals clear within the minute; daily refusals clear at the UTC day boundary.
Staying inside the limits
- Cache article pulls. An hourly ISR pull is the intended pattern. Pulls are not counted against the allowance today, but a pull on every page view is still the wrong shape; always cache or revalidate.
- Space out batch scripts. Anything iterating over prompts or articles should throttle itself under 30 per minute.
- Watch multi-client setups. Two teammates with MCP-connected editors share one project allowance. If you are routinely refused, that is the signal you have outgrown your tier's allowance; see plans and limits.
Design intent
The numbers are sized so that real use, an integrated website, a couple of connected AI tools, daily agent conversations, never notices them. They exist to keep runaway automation from degrading the product for your own team.