Limits & Quotas

Guardrails that keep agent traffic predictable, plus what to know about how current the data an agent reads really is.

LimitValueNotes
MCP requests300 / 5 minPer MCP connection. Exceeding it returns HTTP 429 (rate_limited).
MCP request body256 KBPer request. Larger bodies are rejected with HTTP 413.
AI changes120 / 10 minChanges an AI applies by itself, per workspace, shared by every MCP connection, the built-in assistant and AI workflows. Over the limit a write returns rate_limited and is not applied; earlier changes stay saved. Changes the user approves in AI Changes don't count.
Tool executions150 / 5 minPer account, on the REST tool endpoint. Exceeding it returns HTTP 429.
Assistant chat requests30 / 10 minApplies to the built-in AI assistant.
Tools per assistant reply12Across at most 5 model round-trips, so one question always terminates.
Result size~60 KBLarger results are summarised and explicitly marked as truncated.
Project list100 per pagelist_projects returns up to 200 with limit; page with offset. total and hasMore say how many exist.
Deep model inspection4 modelsPer validate_project call with deep: true. Model files are read with a bounded, time-limited request that never downloads full geometry.

Data freshness

Tools read the last saved state of a project. The editor autosaves a few seconds after each change, so an edit made moments ago may not be visible yet. If an agent reports something that contradicts what you see on screen, save the project and ask again.

Continue reading