Builder Guide
The AI assistant
A per-user assistant that drafts, triages, and acts — always inside an approval queue you control.
Each enabled user account can have a dedicated GoClaw virtual assistant that watches work routed to that user and prepares it: researching permitted context, classifying, drafting replies, extracting details, and suggesting next steps.
The same permission layer as the user
GoClaw acts on behalf of its user; it is not a privileged service account. Every read is limited by that user's tenant, DataRole, record, field, and location permissions. AI-specific CanAiReadRecords, secure-field exclusions, tool policy, and quotas can narrow access further, but the assistant can never widen it.
The approval queue
Customer-facing or material actions land in the approval queue unless an explicit reviewed policy permits that exact low-risk action. Users can review evidence and inputs, edit where allowed, and approve or dismiss. Planning and research alone never mutate business state.
Daily limits
Each assistant has a daily action budget, enforced by the platform itself. When it's reached, further actions queue for the next day. If users report the assistant "stopped working" late in the day, check the quota first — it's the most common cause.
Attribution
Every artifact the assistant produces is tagged with what created it and from which conversation, so there's never ambiguity about whether a person or an assistant wrote something.
The assistant's effective access is never larger than the user's and may be smaller because AI-specific read, tool, quota, approval, and stale-context gates also apply.
For the complete control model — every layer that keeps the assistant useful without ever being able to run away with it — see GoClaw in depth.
Capability review: 2026-09-14. For exact current technical availability, use the generated API Map and first-class module inventory.