Docs

Read what goes to the model, on the laptop

Turn on the conversation gateway so a pasted key is stopped before it leaves, and instructions hidden in a web page are caught before the agent acts on them.

Why

The hook sees what an AI tool does. It doesn't see what the tool sends to the model: the prompt, the files it attached, the page it just fetched. That's where a pasted key leaves the laptop, and where injected instructions arrive. The gateway reads that traffic on the laptop itself.

Nothing to turn on

The gateway is part of ZYBE. Every endpoint runs it from the moment the sensor installs, and Claude Code, Codex CLI and Gemini CLI on that laptop are pointed at it. There's no switch.

Check it's working

Open Inventory, then Endpoints. A laptop that's running the gateway says conversation covered under its last scan. Paste a test key like sk_live_ followed by twenty letters into Claude Code and send it: with Pasted secrets set to Block, the tool shows "ZYBE stopped this message" and the console gets a detection under Secret pasted into the agent chat.

What the developer sees

Nothing, until something is stopped. A stopped message reads:

ZYBE stopped this message: it contains a stripe key. Take it out and send again.

An injected page is never stopped at the gateway. It's recorded, and if the agent then acts on it, that action is judged by the rules under Agent acted after instructions arrived in tool output.

Which AI tools

Cursor and Copilot talk to their own vendor's servers with no way to set the address, so they stay on the hook. Desktop apps are covered through their MCP servers instead.

Turn it off

Flip the same switch off. Within a minute the laptop stops the gateway and the tools talk to their vendors directly again.