Approve, answer and reply from Perch

Turn it on deliberately

Answering is off by default. Enable it in Settings after connecting an agent. You can keep monitoring without using any answering features.

Claude Code supports allow or deny for tools, numbered questions, plan review and replies. Codex supports allow, deny and replies. Cursor supports a reply after a turn. Other interactions belong in the agent’s own app.

Review a tool request

1. Check the session and project so you know which work the request belongs to.
2. Read the command or proposed change and the agent’s reason.
3. Choose allow or deny, or return to the agent’s own app if you need more context.
4. Watch the Sending state. Use Esc or Undo within two seconds to cancel a mistaken response.

Questions, plans and the next instruction

For Claude Code, numbered choices let you answer a question with the matching key, while a free-text response covers a more specific answer. Plan review lets you approve or send feedback while Claude keeps planning.

When a supported turn ends, a reply becomes the next instruction. A reply is different from approving the changes the agent just made. You still need to inspect the work in your project.

Return to the original prompt

If the session’s app is in front when a request arrives, Perch steps aside. Switching back to the app and staying also returns the request. You can set how long Perch waits before handing it back, from 2 to 30 minutes.

If Perch quits, the agent gets the request back. You do not have to finish every interaction in the shelf, and the original agent interface stays part of the workflow.

What happens to request content

The pending request reaches Perch over a private local socket, stays in memory and is discarded once answered. Prompts, commands, diffs and transcripts are not written as session history. See the privacy policy for the full data description.

Common questions