Devpilot
Servers

Terminal AI

Ask questions, get suggested commands, and understand errors right inside the Devpilot Web Terminal.

Terminal AI

The Devpilot Web Terminal ships with an inline AI assistant that watches what you're doing and offers context-aware help: suggested commands for a task, plain-English explanations of output, and diagnoses for error messages. It's built to reduce the number of trips you make to search engines and man pages mid-session.

Terminal AI suggests commands. It never runs them for you. Any command the AI offers lands in your prompt — you review it and press Enter yourself.

Opening the Assistant

Inside any terminal session, click the AI button (or use the keyboard shortcut) to open the AI chat panel alongside your shell. You can type a question, paste an error, or describe what you're trying to do.

What You Can Ask

The assistant is tuned for three common terminal moments:

"How do I find the largest files in /var?"

The assistant understands plain English and returns an actual command you can run. When it's confident, it pre-fills the command in your prompt for you to review. You might get back something like sudo du -ahx /var | sort -rh | head -20.

"What does this systemctl status mean?"

Paste the output (or just ask about the last command you ran) and the assistant explains what the active, inactive, failed, and loaded lines mean, plus which service unit file is being referenced.

"I got Permission denied (publickey) — what now?"

The assistant walks through the most likely causes (missing key, wrong permissions on ~/.ssh/, wrong user) and suggests commands to check each one.

How It Uses Context

To give useful answers, the assistant builds a short context window from your session:

  • Recent commands in the current session (up to the last 10).
  • The server's current working directory and connected user, where available.
  • Any frontend context the terminal UI passes along (for example, the command output currently on screen).

Every AI interaction is logged to Devpilot for audit — you can see later what you asked and what the assistant suggested.

Quick Suggestions

Besides freeform chat, the assistant has three canned suggestion modes you can invoke:

  • commands — Common commands for this kind of server or directory.
  • errors — Checks to run when something broke (logs, permissions, syntax).
  • help — Reminders like man <command>, --help, systemctl status, journalctl.

These are instant, don't call the AI, and are useful when you just need a nudge.

Safety Model

Terminal AI is built with a conservative safety posture.

  • No auto-execution. The assistant can suggest a command, but you always press Enter to run it.
  • Dangerous command awareness. Commands that would destroy the server (rm -rf /, mkfs, writing random bytes to /dev/sda, shutdown, and similar) are recognized. If the assistant ever proposes one, it flags the risk prominently.
  • Rate limiting. AI chat is capped at 10 requests per minute per session. If you hit the limit, wait about a minute and try again — this prevents accidental loops from eating quota.

Privacy

  • Your query text and the assistant's response are logged to Devpilot for audit. You can review past interactions on the session.
  • The AI call itself is made through Devpilot's backend to a third-party model provider. Your SSH password and private keys are never included in the prompt.
  • Command output you explicitly paste into the chat is sent as part of your question. Treat that the same way you'd treat any other text you send to a third-party assistant — avoid pasting secrets.

Don't paste passwords, private keys, or production database rows into the AI chat. If you need to ask about sensitive output, redact it first.

What Terminal AI Is Not

  • Not an auto-pilot. It won't run a sequence of commands on your behalf or fix things without your confirmation.
  • Not a package manager. If you want to install software, use the App Store. Terminal AI will happily explain apt install, but it doesn't replace the App Store's tracking.
  • Not persistent across sessions. Context is scoped to the current session's recent commands. New tab, new context.

Next Steps