Open ButtonsCLI
Name the important terminal tabs so the agent can target them deliberately.
The agent says it ran the tests. Then it says it fixed the failure. Then it says the server is ready. But where did those commands run? What is still running? Which output did it read? The answer should not require an act of faith.
Give the agent a terminal you can both see.
Allow your AI agent to create new terminal tabs, send commands, read tabs, etc.
ButtonsCLI exposes a local control surface to the agent while the actual terminal tabs stay open in front of you. The agent can create and target named tabs, send commands, read recent output, wait for a result, and send control keys.
You see the same workspace. Keep the development server in one pane, tests in another, and live logs in a third. Watch the sequence unfold, type into a terminal yourself, or stop a process when the plan changes.
The handoff begins with a status-bar button called Agent Inst. It copies the installed helper instructions and current local connection details. Paste that into the agent you are already using.
The app must be open. The control endpoint stays on loopback. No screen scraping is needed, and the direct CLI helper remains something you can inspect and run yourself.
“I want the agent to do the work. I also want to see the work.”
Three moves.
No new agent framework. No hidden remote machine. Just a direct handoff into the terminal workspace you already opened.
Click Agent Inst.
Copy the current local helper instructions from the status bar.
Paste into the agent
Let Codex, Claude Code, Reasonix, Gemini CLI, or another local agent operate while you watch.