The default plan allows 10 parallel SSE streams and 60 requests/min. Burst control is server-side; handle 429 gracefully.
crawl_pages defaults to 30 s; override with the timeout parameter (max 120 s).
exec_py gets 1 vCPU, 512 MB RAM, and 5 min wall-clock. Heavy jobs should call external services instead.
Every exec_py run is jailed (no outbound network, no disk writes outside /tmp, package whitelist). Your API-key scopes logs only to your org.
The MCP adapters cache /list_tools for 24 h; pass refresh=True or hit /list_tools?force=1 after adding new tools.