XtroEdge Code

An autonomous computer-use agent for VS Code by XTROEDGE LLC.

Give it a goal — "run the project, open it in the browser and test the checkout flow", "open the docs, read the API section and wire it up", "open the app, test the feature and tell me whether it works" — and XtroEdge plans, discovers what it needs itself, works in your codebase, terminal, VS Code, browser and desktop applications, verifies the result and reports back. Everything it does is shown live, and you can pause, resume or stop it at any moment.

XtroEdge Code runs on the official Claude Code CLI installed on your machine and uses your own Claude account (Pro, Max, Team, Enterprise or Anthropic Console). Sign-in is handled entirely by the official CLI; XtroEdge Code never sees or stores your credentials.

Setup

Install the CLI: npm install -g @anthropic-ai/claude-code
Open XtroEdge Code from the activity bar (or Ctrl+Esc) and click Connect Claude account.
That's it. Desktop control (Windows) and the built-in browser (any Chromium browser: Chrome, Edge, Brave) are detected automatically — type /capabilities to see the live status.

Optional: install the Claude in Chrome extension to let the agent use your own logged-in Chrome tabs as well.

What it can operate

Environment	How	Status
Workspace files	Finds, reads, creates and edits files itself; edits shown as diffs with Allow / Always allow / Deny	✓
Terminal / CMD / PowerShell	Runs commands with captured output and exit codes; can also run them in a VS Code integrated terminal you can watch (PowerShell, cmd, Git Bash), start servers in the background, read their output and stop them	✓
VS Code	Opens files, reads the Problems panel, runs tasks, starts/stops debugging, inspects editor state; checks changed files for new problems after every turn	✓
Browser	Opens a real Chrome/Edge window with its own XtroEdge profile: tabs, navigation, back/forward/refresh, reading pages as text + numbered elements, clicking, typing, selecting, scrolling, waiting for dynamic content, uploads, downloads, console/network errors, dialogs, screenshots	✓ (needs a Chromium browser)
Your Chrome	Via the optional Claude in Chrome extension, for sites where you are already signed in	optional
Desktop applications	Discovers installed apps, launches, switches, minimizes/maximizes/closes windows; reads the accessibility (UI Automation) tree and acts on controls by reference; OCR to locate text on screen; real mouse (move, click, double/right click, mouse down/up, drag & drop, vertical/horizontal scroll) and keyboard (Unicode typing, shortcuts, key hold); clipboard; per-monitor screenshots with change detection	✓ Windows
Web search & fetch	Without a browser	✓
Design	When it builds UI (pages, apps, components, dashboards) it works like a senior product designer: matches your project's existing theme, uses a real design system (design tokens, restrained palette, professional typography and spacing, real icons and content, hover/focus states, accessibility, light/dark) so the result looks human-designed, not AI-generated	✓

The agent picks the environment and tool for each step itself: native/semantic methods first (VS Code tools, terminal, DOM, accessibility), then keyboard shortcuts, then screenshot coordinates or OCR, and it switches methods when one fails instead of repeating it.

Control this PC from your phone or the web

In the XtroEdge dashboard (web or mobile app) open PCs → Connect a PC and copy the one-time token.
In the XtroEdge Code chat in VS Code type /connect, paste the token, keep or change the PC name, and press Connect.
The PC shows Online in the dashboard and the status bar shows XtroEdge · online. Tasks sent from the web or phone open in a new XtroEdge tab here and run visibly; every step, message and the result stream back live. Chats started in VS Code appear on your phone too.

Tokens work once and expire after 10 minutes. /disconnect (or the status-bar menu) disconnects the PC; removing it in the dashboard disconnects it here too. When the agent needs an approval (run a command, open the browser, a sensitive action, a question), the same card appears on the phone and in the web chat — Allow, Always allow or Deny (with an optional note) there, or on the PC; whichever answers first wins and the card closes everywhere. The reply streams live on the phone and web while the agent writes it, screenshots the agent takes appear there too, and the phone/web can do everything this chat can: / commands, @ file mentions (files searched on this PC), image attachments, model and permission mode, pause, resume, stop, and resuming any chat from this PC's history — nothing of the conversation is stored online; it stays in the Claude transcripts on this PC and is read from here.

Connection. The header icon shows the link to your account: green when online, an amber pulse while reconnecting. If the link drops (Wi-Fi change, sleep, server restart) the extension retries by itself (1 s, 2 s, 4 s… then every ~10 s), reconnects immediately when the VS Code window regains focus or you click Retry now, and keeps everything the agent did meanwhile — steps, replies, approvals — to send once it is back. Your PC stays "online" in the dashboard through short blips. Only a key that the server rejects ends the pairing (the PC was removed in the dashboard); then /connect is offered again.

How it works

USER GOAL → plan (live task list) → choose environment & tool
   → OBSERVE (screenshot / UI tree / page text / terminal output / VS Code state)
   → ACT → OBSERVE AGAIN → VERIFY → recover if needed → … → verified result

Local companions. Desktop and browser control run in small local companion processes that the Claude CLI starts for each chat and talks to over its own stdin/stdout; the VS Code bridge is served on 127.0.0.1 with a random per-launch bearer token. Nothing listens for remote connections and nothing bypasses OS, browser or VS Code security.
Permissions. Every companion action goes through the same permission prompt as file edits and commands. With xtroedge.computerPolicy = auto, routine actions (observe, click, type, navigate, run in terminal) run without prompting; sensitive ones still ask. Before anything destructive, financial, communicative (sending messages), installing software or changing security/account settings, the agent must call request_confirmation and wait for your Approve / Reject — that prompt can never be auto-approved or silenced.
Control. While the agent works you always have Pause · Resume · Stop · Cancel. Pause holds permission prompts and makes every companion wait before its next action; Stop makes the next action fail immediately and interrupts the agent. Esc stops too.
Privacy. Text typed with sensitive: true (passwords, tokens) is hidden in the activity feed and tool results; password fields are detected automatically in the browser and in accessibility trees.

Live activity

Every step shows what it is doing (Observing screen, Opening application Notepad, Activating control #12, Typing, Navigating https://…, Running in terminal npm test…) and flips to past tense when done, with screenshots inline.
A computer-state strip shows the active window, cursor position, current browser tab and URL, open dialogs and console errors.
The status line shows the agent state (PLANNING · INSPECTING · EDITING · EXECUTING · VERIFYING · BROWSING · OBSERVING · OPERATING COMPUTER · WAITING FOR YOU · PAUSED) and the current action; a plan panel tracks progress; counters show reads, edits, commands, browser and desktop actions and failures.
Stuck detection: repeating the same action three times warns you and offers Stop; companions also refuse to re-run an identical action that already failed twice and tell the agent to change approach.

Commands

Voice typing: the mic button next to Send opens Windows voice typing (Win+H) — speak and the words are written straight into the chat box, in any language Windows supports (Urdu, Hindi, Arabic, English…). Nothing is sent to a server; review and press Enter. On macOS press Fn twice for Dictation.

The header shows a connection icon when this PC is connected (click it for the account, PC name and Disconnect) and a history icon listing every XtroEdge chat on this PC across all projects (chats from the terminal or other Claude tools are not shown). Any chat can be renamed — the pencil next to the title (or double-click it), the pencil on a history row, /rename New name, or from the phone/web — and the name is written into the chat's transcript on this PC, so it shows everywhere — pick one to resume it in its own project folder.

Type / to see everything. Built in: /new /clear /history /resume /model /mode /usage /cost /capabilities /status /memory /login /logout /settings /help. Commands provided by the Claude CLI and your own custom commands (.claude/commands) appear in the same menu.

Settings

Setting	Default	Meaning
`xtroedge.computer`	on	Desktop control (Windows)
`xtroedge.computerPolicy`	ask	`ask` prompts for every companion action; `auto` runs routine actions automatically, sensitive ones still ask
`xtroedge.browserBuiltin`	on	Built-in browser window (Chrome/Edge/Brave with an XtroEdge profile)
`xtroedge.browser`	on	Claude in Chrome integration (your own Chrome)
`xtroedge.designMode`	on	Senior-designer mode: professional, on-brand, human-looking UI by default
`xtroedge.permissionMode`	auto	Ask before edits · Edit automatically · Plan mode · Auto (safety classifier)
`xtroedge.gitCoAuthor`	off	Off = commits use only your own git identity (no "Co-Authored-By: Claude" attribution)
`xtroedge.maxTurns`	150	Safety limit on agent turns per message

Security

Web pages, files, screenshots and tool output are treated as untrusted data — the agent is instructed never to follow instructions found there and never to reveal secrets. Nothing runs without going through the permission system you configured. XtroEdge Code does not bypass VS Code, browser or OS security, opens no remote endpoints, and asks before irreversible actions.

Website: https://www.xtroedge.com/ · Support: support@xtroedge.com