- one MCP connection remains open for at most 30 minutes;
- one tool call is limited to 30 minutes;
- deploys may interrupt long in-flight work after the drain window;
- structured responses over 200 KB are truncated as a safety backstop; and
request_proposalsdurably enumerates the marketplace and returnsrunning; each background seller attempt has a 90-second bound, or 5 minutes for a seller that acknowledges its brief with a task id within those 90 seconds, and result pages are byte- and count-bounded.
/murph and /apostra) allow
30 requests a minute per account in production; past that the API answers
429, so wait a minute and retry. Reads and other Murph endpoints are not
counted.
Tools that run a model count against the same 30 calls a minute per account.
Today that is generate_variants. Past the limit the call returns a
RATE_LIMITED error, so wait a minute and retry. Every other tool is
unaffected.
Use returned cursors, keep pages small, and preserve partial results. A retry
must reuse an idempotency key only when it represents the same logical attempt.
Upstream seller and inventory systems can dominate latency even when the v3
facade itself is healthy.
Full rate-limit reference
Review the canonical HTTP and MCP rate-limit behavior and retry guidance.