Controlled app actions.
Let Claude request approved actions without giving it a raw account session.
Arc treats Claude, ChatGPT, OpenClaw, Cursor, and custom agents as clients requesting user-approved app actions.
Do not build one-off permission models for every agent surface. Define your app actions once, then expose them through the adapter that fits each client.
Let Claude request approved actions without giving it a raw account session.
Serve ChatGPT through supported tool connection paths while Arc enforces the user's grant.
Give coding and internal agents a narrow action surface for product or operational workflows.
When a client expects MCP, Arc can use it to expose approved actions. The thing Arc adds is the permission decision before a real app action runs.
| Adapter path | Use when | Arc responsibility |
|---|---|---|
| MCP-style tool connection | The client expects MCP-compatible tools. | Expose only approved actions and enforce action policy. |
| Direct action API | You control the agent or internal runtime. | Authenticate the agent request and apply allow / ask / block. |
| Custom connector | Your app needs a specific integration path. | Keep the user grant and audit model consistent. |
If your app already has clear API endpoints, you can wrap the operations users want agents to perform.
Allow reads and drafts, ask before sending, block deletes by default.
Let an agent read availability, ask before creating events, block destructive changes.
Let agents help with support and commerce tasks without account-level power.
Start with the agent client your users are already asking for.