Add an approval step to Claude, ChatGPT and Cursor
Assistants that can use tools do what they are asked, and sometimes what a web page asked. You can add a checkpoint without changing the assistant: connect its tools through a gateway that asks you first.
How it works
These assistants can connect to outside tools using the Model Context Protocol. Instead of connecting the assistant directly to each tool server, you connect it to one address. That address forwards the calls you allow and holds the ones you want to see.
Setting it up
- Get an Operon OS Agenaut account (it is in beta; ask for access on the home page).
- In your assistant's settings, add a custom connector or MCP server and give it Agenaut's address.
- The assistant opens Agenaut's sign-in page. Sign in and choose a level for this agent.
- In Agenaut, add the tool servers you use and switch on the tools the agent may call.
- Add a passkey, so you can approve actions that cannot be undone.
The exact menu names differ between assistants and change over time. The Agenaut docs list the steps for each.
What changes for you
Reads and ordinary edits work as before. When the assistant tries to share something outside, send to someone new or delete, it is told the request is waiting. You get a notification, read the request, and approve or decline. After you approve, ask the assistant to try again and the same request runs once.
What it does not do
It governs only what goes through it. A tool you connected to the assistant directly is not seen. It does not read your conversations with the assistant. And it does not move money: tools classed as money cannot be switched on.
Try it. Operon OS Agenaut is in beta. Request access, or read the docs to see how an agent is connected.
Related guides
- Human in the loop for AI agents: approve risky actions before they run
- AI agent governance: a practical checklist
- What is an MCP gateway, and when do you need one?
Terms used here are explained in the glossary.