Approve or deny agent actions without sitting at the terminal

[0xMag1ci4n]

I want something like a supervised mode — where each change is first shown and vetted by me and only applied if approved. Right now I have to either be at the terminal or just trust the agent blindly. Is this something OpenACP handles?

[jasondbranding]

OpenACP streams tool calls, file edits, and agent thinking in real time to your chat platform. When the agent proposes an action, you see it immediately and can approve or deny from your phone or desktop chat. The permission timeout is 10 minutes by default — if you don't respond, the action is auto-rejected (not auto-approved). This is the human-in-the-loop pattern OpenACP is built around. Works with Telegram, Discord, and Slack — so you're never blocked from reviewing actions just because you're not at your desk.

[jasondbranding]

You can also configure dangerouslyAutoApprove per session if you trust a specific task fully — but the default is always human approval. Good balance between safety and not being spammed with confirmations for read-only operations.

[Cznorth]

This supervised mode requirement resonates — we built WinkTerm around human-in-the-loop at the shell:

AI and user share the same PTY. The agent pre-fills commands into your input line; nothing runs until you press Enter (or backspace to edit). In-terminal chat via # at the prompt.

Self-hosted, SSH/SFTP, Docker, MIT: https://github.com/Cznorth/winkterm

It is not async approval-from-phone (you are still at the terminal), but it avoids blind trust for command execution. Curious whether OpenACP aims for remote approval queues vs in-session vetting.