Agent safety and undo
How external MCP clients execute writes compared to Grid82's built-in assistant, confirmation guards, and rolling back agent changes.
External MCP agents vs the built-in assistant
| External MCP client or CLI | Grid82 built-in assistant | |
|---|---|---|
| Execution model | Executes writes directly as you (prompting depends on your client’s approvals) | Generates a proposal card; nothing mutates until you click Apply |
| Permissions | Your user roles, narrowed by approved OAuth apps and scopes | Your user roles inside that UI plane |
| Attribution in Log | Stamped with your user ID and the MCP/CLI surface | Stamped with your user ID and the assistant surface |
Because every external MCP connection acts strictly as you, an agent can never access a table or field hidden from your roles or escalate its own privileges.
Built-in safety guardrails
- Confirmation on destructive targets. Deleting an organisation, suite, app, table, or field requires passing its exact current name in
?confirm=..., preventing accidental loops over empty or wrong IDs. - Schema impact guard. Dropping a table or field returns a list of affected views, formulas, permissions, and Interface pages until
acknowledge=trueis explicitly sent. - Scoped suite consent. On the OAuth consent page, tick only the specific apps an agent needs for its task. Membership checks run on every request, so removing someone from an app revokes agent reach immediately.
- Atomic changesets. A batch insert of 500 records or a bulk
sql.executestatement lands as one changeset in the app’s log. Revert any agent write cleanly using Log in the Studio, My recent changes in Runtime, orchangesets.undo. - Read-only setup verification. The
/agent-setup.mdprompt explicitly instructs agents to test connections solely via read-onlyapps_listandme_getcalls.