Use the AI agent

Transit AI’s AI assistant is a read-only investigator. It can list your open sessions, read scrollback (with secrets stripped out), propose commands (gated by a per-vendor permit list and your approval click), and ask you clarifying questions. That’s the whole menu — see Features → AI agent for more detail.

This page covers how to use it day-to-day.

Open the chat panel

Keyboard: press Cmd/Ctrl + J.
Mouse: click the chat icon on the right-hand side of the main area.

The panel opens on the right. Drag the divider to resize.

Pick a model (optional)

The chat panel header shows the active model. Click it to pick a different one from your tier’s allow-list:

Free tier: claude-sonnet-4-6.
Pro / Max: claude-sonnet-4-6, claude-opus-4-7, gpt-4o, gpt-4o-mini.

Switching mid-conversation starts the next turn on the new model; prior turns stay on whatever model answered them.

Ask a question

Open at least one SSH session to a device you want the agent to investigate. The agent can only see sessions you have open.
Type a question in the input box at the bottom of the chat panel.
Press Cmd/Ctrl + Enter (or click Send) to submit.
The AI streams its response. If it needs to read scrollback, it does so silently. If it wants to run a command, an approval dialog pops up.

Approve or deny a command

When the AI proposes a command, an approval dialog appears with:

The command it wants to run (e.g. show interfaces terse).
Which session it’ll run on (device name + session ID).
For Linux / Unix devices: an amber warning banner, because Transit AI has no per-vendor permit list for generic Linux — your click is the only check.

Three buttons:

Approve — runs the command in the named session. Output is captured (until the device goes silent for ~800ms) and fed back to the agent.
Approve and always allow — runs the command AND remembers a regex pattern for this chat-and-session pair so future commands matching the same pattern skip the dialog. The pattern is held in memory only; it clears when you close the chat or quit Transit AI. The per-vendor permit list still runs on every command regardless.
Deny — refuses. The AI is told you said no and decides what to do next.

The dialog cannot be dismissed by Enter, Escape, or clicking outside. You must explicitly choose one of the three.

Always-allow shortcut — what it does and doesn’t do

It skips the approval dialog for commands matching your saved pattern.

It does not skip the per-vendor permit list. The permit check runs on every command. A command the list denies stays denied — the AI is told it can’t run it, regardless of any shortcut you’ve saved.

It is held in memory only, scoped to the current chat-and-session pair. Closing the chat or restarting Transit AI clears it.

On Linux / Unix devices, the always-allow shortcut is unavailable. Transit AI hides the option and refuses to honor it — every command on a Linux box gets its own approval dialog, full stop.

Cancel a running investigation

Press Stop in the chat panel header, or close the chat — either one stops the AI mid-task:

If you stop between turns, the chat ends cleanly with a “User aborted” message.
If you stop while a command is running, Transit AI drops the in-flight result and the chat ends.

Stopping doesn’t undo work the AI has already done — a command that was approved and ran has, in fact, run.

Auto-stop — when an investigation ends by itself

Each investigation is capped at:

12 back-and-forth rounds
50,000 tokens
120 seconds of wall clock

Hitting any cap ends the investigation with a chat message saying which cap fired. Send a follow-up question to start fresh.

What the agent never sees

Your passwords or private keys — the AI has no programmatic path to your credential store. An automated check fails the build if anyone tries to add one.
Bytes that look like secrets in device output — PEM blocks, encrypted passwords, AWS keys, JWTs and similar are stripped on your machine before any scrollback reaches the AI, and replaced with placeholders like [REDACTED:pem#1].

See Security model for the full architectural detail.