# Autopilot

> **Warning:** Autopilot is experimental. It's off until you turn it on, and how it works may change between releases.

Autopilot is a Claude Code chat tab on your own installed `claude`. It reads every [Session](/docs/sessions) on your machines and every question waiting in [Notify](/docs/notify) (the app calls it Approvals). Then it puts them in order on its **stack**, the list of items it shows you in turn. Without it, you open each Session and answer each card yourself.

| You ask | The agent |
| --- | --- |
| "Is Autopilot running?" | Runs `runpane autopilot status`. It reports whether Autopilot is on, and running or paused. |
| "What's on Autopilot's stack?" | Runs `runpane autopilot stack`, the items in order and the one on screen. |
| "Show me this plan in Autopilot." | Runs `runpane autopilot stack push --file plan.html`. An HTML file shows as a page, any other file as text. It goes at the end of the stack, or right after the current item with `--next`. |
| "From now on, group things by repo." (to Autopilot) | Offers to add that line to your instructions, as its own approval, and adds it once you approve. |

## Turn it on

1. Install Claude Code 2.1.316 or later and sign in. Autopilot runs as a [chat tab](/docs/chat-tabs) on it.
2. Turn on **Chat tabs** in **Settings → AI & Agents**.
3. Turn on **Settings → Advanced → Experimental → Autopilot**. An **Autopilot** row shows in the sidebar, right above Approvals.
4. Open it and press **Start**. Opening it never starts a turn on its own.

To turn it off, flip the same switch. That stops its agent and keeps its stack, chat and documents. Turn it on again and it picks up the same run.

## How it works

Autopilot reads everything first, writes its plan as a document, then works through it item by item. Each question it brings is a decision card in its chat: what it is, who asked, the options with one recommended. Choosing an option answers it. The right side shows that item's documents, app or Session overview as tabs.

- **Pause** stops it from bringing new items. **Go on** resumes.
- **‹ ›** or the arrow keys step through the stack. New items wait their turn and never replace the one on screen.
- **More → New Autopilot Session** starts a fresh run with an empty stack. The old run's documents stay in its folder.
- After Pane restarts, Autopilot shows **Start** again, with its stack kept.

It also passes your questions and instructions to each Session's orchestrator, and asks orchestrators to wrap up finished work. It never archives, closes or stops a Pane, tab or Session itself.

### What it does with your approvals

No switch controls this. Tell it in its chat, like "leave those to me" or "handle the approvals", or put it in [your instructions](#your-instructions). By default it:

- **Brings** approvals from every connected machine, one related group at a time.
- **Closes** the ones that don't matter or have gone stale.
- **Rewrites** an unclear one as a clearer question. Your answer goes to the original asker.
- **Answers** only when the way forward is clear and the item belongs to the Session or group you're talking through with it right now, or when you've told it to handle approvals.

> **Note:** Autopilot can't answer a permission prompt, answer or rewrite a secret card, or type keys into another tab. Pane refuses it with `AUTOPILOT_CANNOT_ANSWER`, even if you ask.

### How it's recorded

Every action shows on the item, on every device:

- "Answered by Autopilot · because …", with the plan, document or your words it relied on
- "Closed by Autopilot: …", with its reason
- "Closed by Autopilot: rewritten as …" on the old item, and "Rewritten by Autopilot from Q-n" on the new one, where Q-n is the old question's number

### Take it back

- **Override** on an item Autopilot answered sends your answer to the asker as a correction.
- **Reopen** on an item it closed puts it back in front of you. A rewritten item has no Reopen, because the new question replaced it. Answer that one instead.

Questions Autopilot writes itself are violet, marked "written by Autopilot", and grouped under **Autopilot** in Approvals. You can still answer any card in Approvals yourself.

## Your instructions

**Settings → Autopilot** shows only while Autopilot is on. It has a **Your instructions** box. Tell it how to work: "be brief", "always ask before closing anything", "start with the website". It reads them after Pane's own rules, and they can't override the limits above.

Save while it runs and it gets them at once; the box says **Applied**. While it's stopped, the box says **Applies on next Start**. **View built-in instructions** shows Pane's rules, and **Reset** clears yours. They're kept in `<PANE_DIR>/autopilot/instructions.md`, which you can edit directly.

## Sentinel

**Sentinel**, a switch in Autopilot's header, keeps your Sessions moving while you're away. Turning it on also starts Autopilot. Pane wakes Autopilot every 5 minutes while it's idle. Each check is a turn on your own Claude Code sign-in, so it uses your Claude usage. It answers what a plan or decision already settles, nudges a Session that's gone quiet, and sends the rest to you as its own approval, which reaches your [phone](/docs/phone). It follows the same approval rules, and never archives or creates Sessions.

When you come back, a **While you were away** card lists what it did. After Pane restarts, Sentinel is off and Pane asks whether to turn it back on.

## Debug

- **The switch is greyed out.** Chat tabs are off. Turn them on in **Settings → AI & Agents**.
- **`AUTOPILOT_OFF`.** Every `runpane autopilot` command except `status` needs Autopilot on.
- **An item reads "&lt;machine&gt; offline".** Autopilot can't reach that machine, and an answer to it is refused with nothing queued.
- **"Can't see: &lt;machine&gt;".** That's one of your Tailscale machines that isn't saved on this one. Add it in [your computers](/docs/your-computers).
- **Autopilot stopped by itself.** Its agent exited. Press **Start** again.

The [`runpane` CLI](/docs/runpane-cli)'s `autopilot` commands are also tools in the [MCP server](/docs/mcp)'s `autopilot` toolset, such as `autopilot_status` and `autopilot_stack_push`.
