Trusted Actions
Trusted Actions protect effectful MCP operations. A read-only tool can run under its configured policy; a consequential call waits for the owner unless an active, narrow grant already allows that connector and tool.
Open Dashboard → Approvals to review pending requests, past decisions, active grants, and execution history.
Approval flow
Agent requests a tool call
The request records the agent, connector, tool, input summary, and mission context.
Hatcher checks policy
The call can proceed only if it is read-only, explicitly approved once, or covered by a matching unexpired grant.
Owner decides
Approve the exact request, deny it, or create a narrow grant when repeated access is appropriate.
Execution is recorded
The approval and resulting connector execution stay visible in the activity history.
Safer grants
Keep grants intentionally small:
- Limit them to one agent
- Limit them to one connector and selected tools
- Prefer short expiration windows
- Revoke access when the workflow is finished
- Re-check any tool that writes, posts, sends, purchases, deploys, or changes external state
An approval authorizes the requested external action. Review connector identity, tool name, arguments, target account, and mission context before approving.
Connector setup
Use the agent’s Connectors tab to attach a trusted HTTPS MCP server, store its authentication encrypted, discover tools, and define the allowlist. See MCP Connectors.