Using the AI panel¶

The AI panel sits on the right of the workspace and works on the session in the active tab. The steps below follow one investigation: ask, read the proposal, decide, and let the AI carry on with the result. They assume you have configured at least one AI Provider.

Ctrl+\ shows or hides the panel for every tab; the panel's own ✕ hides it for the current session only.

1. Open a session with AI on¶

AI works on SSH and Local Console tabs. The session needs:

Enable AI on for its connection (or turned on in the tab, from AI Assistant → AI for This Host: toggle)
an active free trial or subscription. During the trial, AI is on for 2 of your connections at a time, which you choose; see Free trial and limits

When the license keeps AI off in a tab, the composer shows the reason instead of the input box — for example AI off · free trial: AI on 2 connections at a time. followed by the connections that have it — with the button that helps: Use AI here…, Turn AI on here, Subscribe, License settings or Fix clock.

On a Local Console tab the model is also told your operating system and shell, so it proposes commands in the right syntax — PowerShell or cmd on Windows, zsh or bash on macOS and Linux. Those commands run on your own workstation.

An SSH tab where nginx has just failed, with the AI panel open beside it showing the Analyze Error, Explain Command, Review Config and Optimize cards web-01 has AI on: the connection banner says ai: enabled and the side list shows AI. The empty panel offers the four quick actions, with the provider dropdown at the top and the Auto-run off chip beside the composer. In the trial, web-02 holds the other AI place, and bastion-01 and staging-app, which also have AI turned on, show AI off · trial limit.

2. Choose the provider for this session¶

The provider dropdown at the top of the panel sets which configured provider answers this session's questions. It starts on the provider that is active in Settings → AI Providers.

The choice is per session rather than global: a fast local model on the tab where you are reading logs, a larger cloud model on the tab where you are stuck. Switching does not restart the conversation.

3. Ask a question, or use Analyze Error¶

Type into the composer at the bottom of the panel (Ask a question…) and press Enter. The question is answered against the recent scrollback of the session you are in, so you rarely need to paste anything — "why did that fail?" after a failed command is usually the whole question.

Or start from a quick action. They appear as cards on an empty panel, and in the + menu beside the composer once a conversation has started:

Quick action	What it does
Analyze Error	Sends the recent scrollback for diagnosis straight away — use it when something just failed. Ctrl+Shift+A does the same
Explain Command	Starts the question with "Explain this command:" for you to finish
Review Config	Starts the question with "Review this config for issues:"
Optimize	Starts the question with "Suggest performance optimizations for:"

The + menu also holds Upload from computer and Browse remote files for attaching a file to the question. Attachments are redacted before they are attached, and OpsPilot tells you how many secrets it removed; a large file is cut to its first part, and OpsPilot says so.

Session content is redacted before it is sent. The question you type is sent as you typed it, so do not paste a secret into it.

4. Read the note¶

The reply streams in as it is generated, and the panel header shows thinking… or investigating… while it does. A reply is a short note — what the model sees and why it wants the next step — plus, usually, one proposed command rather than a multi-step plan. OpsPilot asks the model for a single next action, runs it if you approve, and then asks again with the real result.

5. Read the proposal card¶

Each proposed command arrives as a card with:

the model's reason for running it, and a risk badge — Read-only, Low risk or High risk
the exact command text you are about to run, verbatim, in a shell block

The tier is not the model's word alone. The session's Command Safety profile applies your dangerous-pattern list on top of the model's own self-assessment, and a pattern match makes a command High risk. If the profile is set to Only my list and the model flagged a command your list does not match, the card carries an amber note: The AI flagged this command as destructive. This connection's Command Safety profile only treats your own patterns as dangerous, so that warning was not applied.

What the card asks of you depends on the tier and on the profile's rules:

will run automatically — the profile lets this command run without a click, so it runs at once (see the auto-run chip below).
approve & run and dismiss — a click decides.
a reason box and confirm & run — a High risk command in a profile that asks for a typed reason. The box names why the command was flagged: Flagged dangerous by the AI., or the pattern it matched.

The AI panel with the question, the AI's one-line note and a Read-only proposal for nginx -t with approve & run and dismiss The question, the AI's short note, and one proposal. nginx -t only tests the configuration, so it is Read-only. The profile here asks every time, so it waits for approve & run.

6. Approve and run, or dismiss¶

approve & run — OpsPilot types the command into the real session, and you watch it run in the terminal.
dismiss — nothing runs. The card collapses and the conversation continues.

The first step folded to one line, the configuration test's output in the terminal, and a Low risk proposal to correct the file with approve & run and dismiss After the first approval the finished step folds to one line and its output is in the terminal. The model reads the result and proposes the fix. Editing the file changes something, so the card is Low risk and waits for a click.

For a High risk command with the typed reason on, type why you are running it. confirm & run stays disabled until you have typed at least ten characters, then needs a click or Enter. The reason stays with the command, shown as Approved: followed by your words, in the session transcript. That transcript is in memory only, so capture it yourself if you need the record kept — see Approvals & auto-run.

A High risk proposal to clear the nginx cache and reload, with the matched pattern named, a typed reason and the confirm & run button A High risk proposal. The command contains rm -rf, one of the Default profile's dangerous patterns, and the card says so. confirm & run works once a reason of at least ten characters has been typed.

A dangerous command always needs this click. No setting runs one by itself.

7. Follow the next turn¶

Once a command runs, OpsPilot captures what it printed, strips terminal escape sequences, redacts it, and feeds it back as the next turn. The model then explains what the output shows and proposes the next step, so you do not have to re-ask after every command.

A command that is still running when OpsPilot stops waiting shows still running — ask a follow-up once it finishes.
After 12 follow-up turns since your last question, the panel pauses with Paused after 12 automatic steps — ask a follow-up to keep going.

8. Read the summary¶

When the model has an answer, the investigation ends with a summary card: the full write-up, with a Copy button for pasting it into a ticket or a runbook. The panel header shows done. If the model needs a decision only you can make, it says so and the header shows needs input; ask a follow-up to continue.

A finished investigation: three steps folded to one line each, followed by the summary of what was wrong and what was fixed The whole loop: test, fix, test again, each step folded to one line, then the summary. The header shows done.

The New AI session button (+) in the panel header starts afresh for the tab: it clears the conversation and leaves the connection untouched.

9. Stop¶

While the model is answering, or while OpsPilot waits for a command's output, the send button becomes a Stop button. Stopping cancels the answer (the turn reads Stopped.) or stops waiting for the output. Typing wins over stopping: as soon as there is something to send, the button goes back to Send, and sending starts a new turn.

The auto-run chip¶

The chip beside the composer shows the rule for this tab's Command Safety profile: Auto-run off (every command waits for your click), Auto-run read-only (read-only commands run by themselves) or Auto-run all (every command runs by itself except dangerous ones). Clicking it opens Settings → Security, where the rules are set per profile. See Approvals & auto-run.

The redaction badge¶

If a 🔒 badge with a count appears on a turn or a command, redaction fired before anything was sent. The number is how many items were replaced; hover it to see which categories they came from.

The badge is an indicator, not a control. Turning it off in settings hides the indicator only; redaction still happens on every turn either way.

When AI goes off during an answer¶

If the license stops allowing AI in this session while the model is answering — the trial or the subscription ends, for example — the answer already under way finishes. Nothing it proposes runs: its card shows, for example, Expired: AI is off. New questions are refused with the reason, and no session is closed.

If you turn AI off in the session yourself, or move the trial's AI to another connection, the answer stops there and the turn says why. Either way, a proposal still waiting for your decision expires.

When a proposal came from an assistant¶

If a proposal came from a connected assistant rather than the panel, its turn carries a via … badge naming the assistant, and the composer shows Use … chat with a Switch to AI Provider button, because OpsPilot cannot type into the assistant's own chat. Those proposals use the same cards, the same tiers, the same profile rules and the same buttons — see AI Assistants (MCP).