Skip to main content
Glama
Sign In
linux-computer-use

linux-computer-use

by Roadmvn
OverviewSchemaSourceRelated ServersDiscussionsAdmin
JavaScript
Local

linux-computer-use

linux-computer-use is an open source computer use agent for Linux. It is an MCP server that lets an AI agent such as Claude Code or OpenAI Codex drive a real browser, and native Linux applications, on your own Linux machine, with a live view, human takeover, and safety guardrails.

Computer use for AI agents landed first on macOS and Windows, or inside a cloud VM. This project brings the same capability to Linux, self-hosted, running on your machine with your own browser and your own cookies.

Install

curl -fsSL https://raw.githubusercontent.com/Roadmvn/linux-computer-use/main/scripts/install.sh | bash

That is the whole setup. The script is scripts/install.sh in this repository, so you can read it before running it.

Related MCP server: BrowserMCP Secure

Connect it to your agent

# Claude Code
claude mcp add --scope user linux-computer-use -- node ~/.linux-computer-use/app/src/index.js
# OpenAI Codex
codex mcp add linux-computer-use -- node ~/.linux-computer-use/app/src/index.js

Restart your client afterwards so it picks the server up.

Quick start

Ask your agent: "Open github.com in a session named research, then tell me what is on the page." Then watch it work:

cd ~/.linux-computer-use/app && npx playwright-cli show

What it does

Capability

What you get

Live view

playwright-cli show streams every session, in a window here or served to another machine.

Human takeover

A lease file freezes the agent mid-run, and only you can release it.

Visible cursor

An overlay draws the agent's pointer in the page, so you see where it acts.

Sessions and profiles

Named isolated sessions, cookies kept between runs when profile is set.

Tabs

List, open, select and close tabs with a single tool.

Two ways to act

Accessibility tree with element refs, the robust default, or raw pixel coordinates.

Guardrails

Fail closed on credentials, account pickers, irreversible actions and unreadable targets.

Native applications

Burp Suite, Wireshark, Ghidra and the rest, in a nested desktop that leaves your pointer alone.

Auto mode

set_mode drops the confirmation before a form is committed. The hard stops still apply.

MCP tools

Tool

What it does

Parameters

open

Open or attach a browser session

url, session, profile, browser, cdp

goto

Navigate to a URL

url

snapshot

Accessibility tree of the page, with element refs

-

screenshot

Image capture of the page

-

click

Click an element by ref

ref, confirm

type

Type text

text

fill

Fill a field by ref

ref, text

press

Press a key

key

mouse

Act at pixel coordinates, vision mode

action (move, click, down, up, wheel), x, y, confirm

tabs

Manage tabs

action (list, new, select, close), index, url

history

Navigate the history

direction (back, forward, reload)

status

Session state and control lease

takeover

set_mode

Switch between normal and auto

mode

desktop_start

Start the agent desktop, or shut it down

width, height, stop

desktop_windows

List the visible windows

-

desktop_focus

Bring a window to the front

id

desktop_screenshot

Capture the desktop, or a single window

window

desktop_click

Move the pointer and click

x, y, button, move_only

desktop_type

Type text into the focused window

text, confirm

desktop_key

Press a key or a combination

key, confirm

desktop_launch

Start an application by name

app, confirm

Requirements

  • Linux, Node.js 20 or newer.

  • A graphical session: DISPLAY or WAYLAND_DISPLAY set for the MCP client. No screen, wrap it in xvfb-run.

  • The system libraries of Chromium. The installer pulls them on apt based systems.

  • Desktop backend only: sudo apt install xserver-xephyr xdotool imagemagick xfwm4. Not needed to drive a browser.

Comparison

Platform

Where it runs

Live view

Human takeover

Guardrails

Open source

linux-computer-use

Linux

Your machine

Yes, playwright-cli show, locally or served

Yes, explicit lease released from your machine

Yes

Yes, MIT

OpenAI Operator / ChatGPT agent

Any, used through ChatGPT

Cloud VM

Yes

Yes

Yes

No

Anthropic Claude for Chrome

Chrome extension

Your browser

Your own browser window

-

Yes

No

Playwright MCP

Cross platform

Your machine

Yes

Yes

-

Yes

A dash means not verified rather than absent. These products move fast, so check their own documentation before relying on a row.

FAQ

Can AI control a browser on Linux?

Yes. Install this MCP server, register it with Claude Code, OpenAI Codex, or any MCP client, and the agent drives Chromium, Chrome, or Firefox on your Linux machine while you watch.

What is the Linux alternative to OpenAI Operator?

linux-computer-use. Operator runs a browser in a cloud VM you reach through ChatGPT. This runs on your own Linux machine, with your browser and your profile, and it is open source under MIT.

How do I give Claude Code browser access?

Run the installer, then claude mcp add --scope user linux-computer-use -- node ~/.linux-computer-use/app/src/index.js, and ask Claude Code to open a page. The 21 tools show up in its tool list.

Can it use my existing cookies and logins?

Yes, in two ways: pass profile with a directory path, or start your own browser with a debugging port and attach with cdp, see docs/usage.md. Without either, the profile is in memory and disappears with the browser.

Will it log into my accounts on its own?

No. It stops on an empty login form, and it asks you when several accounts are offered. It also asks before an action it judges irreversible.

Documentation

  • docs/usage.md - first run, your own browser, native applications, guardrails, environment variables, audit trail, more FAQ.

  • docs/design.md - architecture, and the reasoning behind each choice.

Roadmap

  • AT-SPI integration, the Linux accessibility bus, to give the desktop backend the equivalent of the browser's accessibility snapshot, and with it real guardrails by element name. Planned, not available yet.

Contributing

Issues and pull requests are welcome on GitHub. Architecture notes live in docs/design.md.

License

MIT. See LICENSE.

Available Tools

13 tools
clickA

Click an element by its ref from snapshot. Blocked when the element looks irreversible unless confirm is true.

Inputs: ref (required), confirm.

fillB

Fill a field identified by its ref. Refused on credential fields.

Inputs: ref (required), text (required).

gotoC

Navigate the current tab to a URL.

Inputs: url (required).

historyC

Go back, go forward or reload.

Inputs: direction (required).

mouseA

Coordinate based mouse control for canvas, maps and custom widgets that have no accessible element. Take a screenshot first to read the coordinates.

Inputs: action (required), x, y, confirm.

openA

Open a browser session, or attach to a browser already running with a CDP endpoint. Call this first.

Inputs: cdp, url, browser, profile, +1 more.

pressB

Press a key, for example Enter, Tab or ArrowDown.

Inputs: key (required).

screenshotA

Image of the page. Use it to read visual content, verify a result, or work out coordinates for the mouse tool.

Inputs: no named parameters.

set_modeA

normal asks before sensitive actions. auto runs a complete objective end to end, stopping only for an empty login, an account choice, or an irreversible action.

Inputs: mode (required).

snapshotA

Accessibility tree of the page with a ref for each element. Use these refs with click and fill. Prefer this over screenshot for acting: it is exact and cheap.

Inputs: no named parameters.

statusA

Report the session, the mode and who holds the control lease. Pass takeover: true to hand control to the human before they act in the dashboard. Control comes back only when the human releases it on their machine, never through this tool.

Inputs: takeover.

tabsB

List, open, select or close tabs.

Inputs: action (required), url, index.

typeA

Type text into the focused element. Refused when a password field has focus.

Inputs: text (required).

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Deploy Server
    One endpoint for every MCP toolRayrun runs every MCP server behind one URL and holds the credentials. Each call is decided per client, tool, and argument, then recorded.ray.runAd

    TDQS

    A3.5/5.0

    Scored across 13 tools

    Disambiguation4/5

    Most tools have clearly distinct purposes: goto for navigation, snapshot for accessibility tree, screenshot for visual capture, click/type/fill/press for input. The overlaps between click and mouse (ref vs coordinates) and type vs fill (focused vs by ref) are clarified by descriptions, making misselection unlikely.

    Naming Consistency3/5

    Tool names are all lowercase and mostly single words, but they mix verb forms (goto, click, type, fill, press, open) with noun forms (snapshot, screenshot, mouse, tabs, history, status). The only underscore in set_mode breaks the otherwise uniform style, creating a minor inconsistency.

    Tool Count5/5

    13 tools is well within the ideal range for a browser automation server. Each tool covers a distinct aspect of interaction, from session management and navigation to reading state and performing actions, without feeling bloated or sparse.

    Completeness4/5

    The toolset covers the core lifecycle: open, navigate, inspect (snapshot/screenshot), interact (click/type/fill/press/mouse), manage tabs/history, and control session mode. Minor gaps like an explicit wait or scroll tool are workable because snapshot can serve as a synchronization point and mouse can handle scrolling.

    Resources

    Related MCP Connectors

    Related MCP Servers

    • BrowserMCP Secure

      manishpandey9
      A
      license
      Not graded
      quality
      D
      maintenance
      Security-hardened MCP server that gives AI assistants full control over your real browser session, supporting 36 tools for navigation, data extraction, monitoring, and more.
      MIT

    • xfactor

      fveiraswww
      F
      license
      Not graded
      quality
      D
      maintenance
      MCP server that connects AI agents to browser DevTools via CDP, enabling real-time access to console logs, network requests, and page state.
      -