Terminal & CLI

Watch the agent run commands in a branded VS Code terminal, hand terminal output to the chat, and run the full LocalPilot agent from any shell with the `localpilot` CLI.

Ask AI about this page

ChatGPT Claude Perplexity Gemini View as Markdown

The agent in your terminal

When the agent runs a command (run_command), it executes in a visible, branded LocalPilot Agent terminal through VS Code’s shell integration. You watch the command run live — with the shell’s own prompt, colours and exit-code markers — while the agent receives the captured output and exit code.

The LocalPilot Agent terminal running npm test twice — failing before the fix, passing after — while the chat shows the captured exit codes and “1 file changed”

One agent terminal per chat session, reused across commands.
Output is captured through shell integration, so it is exactly what you see; ANSI colours and shell-integration markers are stripped before it reaches the model.
Stopping the chat sends Ctrl+C to the running command; long commands time out (timeout_seconds, max 10 minutes).
If the shell has no integration (or localpilot.terminal.runInIntegratedTerminal is off), the command runs in a hidden process instead — same result, just not visible.
background: true commands (dev servers, watchers) open their own terminal and are not awaited.

Setting	Default	Purpose
`localpilot.terminal.runInIntegratedTerminal`	`true`	Run agent commands in the visible LocalPilot Agent terminal
`localpilot.terminal.reveal`	`true`	Show the terminal panel (without stealing focus) when a command runs

Terminal output as context

LocalPilot records the commands you run in any VS Code terminal (via shell integration), so the last command and its output are one click away:

Right-click a terminal → LocalPilot: Explain Terminal Output, Fix Failed Command or Add Terminal Output to Chat.
Type @terminal in the chat to attach the last command + output of the active terminal.
Fix Failed Command picks the most recent command that exited non-zero, opens an Agent chat and asks it to find the cause, fix it and re-run the command.
⌘⌥K (Ctrl+Alt+K) in a terminal still generates a shell command from a description.

The `localpilot` CLI

localpilot is the complete LocalPilot agent for the shell — the same agent loop, tools, skills, memory, MCP servers and change tracking as the extension, running under Node against your local model servers. Nothing is sent anywhere else.

The localpilot CLI running inside the branded LocalPilot terminal: it edits math.js, runs npm test, prints a summary and lists the changed file

Getting it

Inside VS Code — LocalPilot: Open LocalPilot Terminal (also in the chat toolbar’s ⋯ menu and the terminal + dropdown as the LocalPilot profile). The CLI is on PATH in that terminal; nothing to install.
Everywhere — LocalPilot: Install ‘localpilot’ CLI in PATH writes ~/.local/bin/localpilot (Windows: %USERPROFILE%\bin\localpilot.cmd). The shim finds the newest installed LocalPilot after updates and runs with node, or with VS Code’s own runtime when Node isn’t installed.

Usage

localpilot "fix the failing tests in src/api"     # one task in agent mode, then exit
localpilot                                        # interactive session (REPL)
localpilot --ask "how does the retry logic work?"
localpilot --plan "migrate the auth module to JWT"
localpilot -f src/db.ts "add connection pooling"   # attach files (path#L10-L20 for a range)
localpilot --yes "update all deps and run tests"   # auto-approve commands/edits (CI, scripts)
localpilot -p "summarise this repo" > NOTES.md     # print only the final answer
git diff | localpilot --ask "review this diff"     # piped stdin becomes the prompt
localpilot -m ollama/qwen3:8b "…"                   # pick a model for this run
localpilot models                                 # models from every reachable server
localpilot doctor                                 # servers, settings, skills, MCP, ripgrep
localpilot --ide "…"  ·  --settings  ·  --setup   # hand off to VS Code

In a session: /mode agent|ask|plan, /model <name>, /models, /changes, /undo, /keep, /file <path>, /clear, /doctor, /exit. Ctrl+C stops the current run; press it again to quit.

Approvals

Every command, file edit, delete and network call goes through the same approval policies as the extension (localpilot.approvals.*). When a policy says ask, the CLI shows the command or diff and prompts [y]es · [n]o · [a]lways this session. --yes approves everything; without a TTY (CI) anything that needs approval is denied unless --yes is given.

Edits are tracked per run: a one-shot run prints the changed files at the end; in a session /undo reverts them and /keep accepts.

Settings

The CLI reads, in order: VS Code’s user settings.json, the workspace’s .vscode/settings.json, ~/.localpilot/config.json, then flags. So a model server configured in LocalPilot Settings is already known to the CLI. Changes made from the CLI (/model) are saved to ~/.localpilot/config.json, never to VS Code’s settings. Logs go to ~/.localpilot/cli.log.

~/.localpilot/config.json uses the same keys without the localpilot. prefix:

{
  "providers": [{ "id": "ollama", "type": "ollama", "baseUrl": "http://localhost:11434" }],
  "defaultModel": "ollama/qwen2.5-coder:7b",
  "approvals": { "terminal": "ask", "write": "always" }
}

Deep links

The CLI’s --ide, --settings and --setup flags use LocalPilot’s URI handler, which you can call from anything: vscode://localpilot.localpilot-agent/run?mode=agent&prompt=…, /new, /settings?section=providers, /setup, /terminal?run=localpilot%20--help.