linux-computer-use
Enables AI agents to control Linux/X11 desktops by providing tools for taking screenshots, clicking, typing, and managing windows via AT-SPI and xdotool.
README
<p align="center"> <img src="./assets/logo.png" width="40%" alt="linux-computer-use"> </p>
<h1 align="center">linux-computer-use</h1>
<p align="center"> <em>Linux/X11 computer-use tools for AI agents — Pi, Claude Code, OpenCode, and any MCP-aware client. AT-SPI + xdotool, ~1k LOC.</em> </p>
<p align="center"> <a href="LICENSE"><img alt="license" src="https://img.shields.io/github/license/tak-uukti/linux-computer-use?style=flat-square"></a> <img alt="platform" src="https://img.shields.io/badge/platform-linux%20%7C%20x11-blue?style=flat-square"> <img alt="node" src="https://img.shields.io/badge/node-%E2%89%A520.6-339933?style=flat-square&logo=node.js&logoColor=white"> <img alt="python" src="https://img.shields.io/badge/python-%E2%89%A53.10-3776ab?style=flat-square&logo=python&logoColor=white"> <a href="https://github.com/tak-uukti/linux-computer-use/releases"><img alt="release" src="https://img.shields.io/github/v/tag/tak-uukti/linux-computer-use?style=flat-square&label=release"></a> </p>
A Linux port of @injaneity/pi-computer-use. One bridge, three frontends:
- Pi extension for
mariozechner/pi-coding-agent - MCP server for Claude Code (and any other MCP host)
- MCP server for OpenCode
The macOS original uses Apple's Accessibility API + AppleScript + ScreenCaptureKit (~6,800 lines of Swift + TS). This port replaces the entire native layer with AT-SPI 2 + xdotool + scrot, ships a single ~470-line Python bridge, and trims the tool surface from 15 → 8 to keep prompts cheap.
| upstream macOS | this port | |
|---|---|---|
| Total LOC | ~6,866 | ~1,200 (-83%) |
| Tools registered | ~15 | 8 |
| Native helper | 2,065 lines Swift | 471 lines Python |
| Runtime deps | Swift toolchain, codesign | python3-gi, xdotool, wmctrl, scrot |
| Frontends | macOS only | Pi · Claude Code · OpenCode · any MCP client |
System dependencies (all installs)
# Debian/Ubuntu
sudo apt-get install -y python3 python3-gi gir1.2-atspi-2.0 xdotool wmctrl scrot
# Enable AT-SPI on the desktop session (GNOME)
gsettings set org.gnome.desktop.interface toolkit-accessibility true
X11 only — Wayland sessions cannot capture other-app windows or synthesize input via xdotool. Run a GNOME-on-Xorg, KDE-on-X11, or XFCE session.
Install
Option 1 — Pi (mariozechner/pi-coding-agent)
pi install git:github.com/tak-uukti/linux-computer-use@v0.2.0
The postinstall script writes a small bash wrapper to ~/.pi/agent/helpers/linux-computer-use/bridge that execs python3 bridge/bridge.py. No build step, no codesign, no native compile.
In a Pi session, call screenshot first — it picks the focused window, returns AT-SPI refs (@e1, @e2, …) plus a PNG, then you can click({ref:"@e3"}), set_text({ref:"@e2", text:"…"}), etc.
Option 2 — Claude Code (MCP)
Installable as an MCP server straight from GitHub via uvx (no clone, no manual venv):
claude mcp add linux-computer-use -- uvx --from git+https://github.com/tak-uukti/linux-computer-use linux-computer-use-mcp
Or, equivalently, drop this into your Claude Code MCP config file (~/.claude.json under mcpServers):
{
"mcpServers": {
"linux-computer-use": {
"command": "uvx",
"args": [
"--from",
"git+https://github.com/tak-uukti/linux-computer-use",
"linux-computer-use-mcp"
]
}
}
}
Restart Claude Code; the 8 tools (list_windows, screenshot, click, type_text, set_text, keypress, scroll, computer_actions) appear under the linux-computer-use namespace.
Option 3 — OpenCode (MCP)
Add to ~/.config/opencode/opencode.json:
{
"$schema": "https://opencode.ai/config.json",
"mcp": {
"linux-computer-use": {
"type": "local",
"command": [
"uvx",
"--from",
"git+https://github.com/tak-uukti/linux-computer-use",
"linux-computer-use-mcp"
],
"enabled": true
}
}
}
Restart OpenCode and the tools become available to the agent.
Tools
8 total. Schemas are deliberately terse — see extensions/computer-use.ts (Pi) or mcp_server/server.py (MCP).
| name | purpose |
|---|---|
list_windows |
enumerate visible X11 windows; returns @wN, title, pid, geometry, focus state |
screenshot |
focus a window, capture PNG, walk AT-SPI tree → @eN targets with role / name / bounds / capabilities |
click |
click @eN, @wN, or x,y; supports button and clickCount |
type_text |
xdotool-type literal text at the cursor |
set_text |
replace value of an @eN text/entry via AT-SPI EditableText (falls back to focus + Ctrl+A + type) |
keypress |
press keys/chords — ["Return"], ["Ctrl","A"], ["ctrl+l","Return"], etc. |
scroll |
scroll at ref/coords by pixel delta |
computer_actions |
batch up to 20 actions in a single call |
Architecture
┌──────────────────────────────────────────────┐
│ Pi Claude Code OpenCode │
└──────┬─────────────┬──────────────────┬──────┘
│ │ │
│ extension │ MCP stdio │ MCP stdio
▼ ▼ ▼
┌──────────────┐ ┌────────────────────────────┐
│ extensions/ │ │ mcp_server/server.py │
│ computer- │ │ FastMCP wrapper (8 tools) │
│ use.ts │ └─────────────┬──────────────┘
└──────┬───────┘ │
│ │
▼ ▼
┌──────────────────────────────────────────────┐
│ bridge/bridge.py newline-JSON over stdio │
│ AT-SPI walk · xdotool · wmctrl · scrot │
└──────────────────────────────────────────────┘
The AT-SPI walker is depth-capped (12) and element-capped (200) to keep prompts lean. Element bounds use SCREEN coords with a fallback to WINDOW coords + window offset (necessary for GTK4 / Xwayland which report SCREEN as 0,0).
Verified end-to-end
These captures are from the bridge running against a Xvfb :99 + openbox session, driving real Linux apps. Screenshots taken via scrot after the bridge issued the actions.
gnome-calculator — keypress flow
keypress: 7, +, 8, Return → display shows 15. 26 AT-SPI elements detected, every push button reports canPress: true and accurate bounds.
<p align="center"> <img src="./assets/screenshots/01-calc-keypress.png" width="60%"> </p>
gnome-calculator — AT-SPI @eN ref clicks
computer_actions: [click @e3, click @e7] (which the bridge resolves to push buttons "4" and "5") → display shows 45.
<p align="center"> <img src="./assets/screenshots/02-calc-ax-ref-click.png" width="60%"> </p>
gedit — full type_text round-trip
type_text: "Hello sir, … Linux X11 + AT-SPI + xdotool working end-to-end." → 169 characters typed. 190 AT-SPI elements found in gedit's window.
<p align="center"> <img src="./assets/screenshots/03-gedit-typed.png" width="80%"> </p>
gedit — clear and retype
keypress ctrl+a → keypress Delete → type_text "Taksheel". Status bar reads Ln 1, Col 9.
<p align="center"> <img src="./assets/screenshots/04-gedit-taksheel.png" width="80%"> </p>
App compatibility matrix
| App | screenshot | AT-SPI refs | input |
|---|---|---|---|
| gnome-calculator | ✅ | ✅ 26 elements, full action metadata | ✅ |
| gedit | ✅ | ✅ 190 elements | ✅ |
| GTK / Qt apps with AT-SPI | ✅ | ✅ | ✅ |
| Google Chrome / Chromium | ✅ | ⚠️ AT-SPI tree empty unless launched with --force-renderer-accessibility |
✅ (coords / keypress) |
| Firefox | ✅ | ✅ on a real session (gates on gsettings toolkit-accessibility) |
✅ |
| Electron apps | ✅ | ⚠️ same as Chrome — needs --force-renderer-accessibility |
✅ |
| LibreOffice (real Xorg session) | ✅ | ✅ via SAL_USE_COMMON_ONE_ACCESSIBILITY=1 |
✅ |
| Xvfb / nested X | ✅ | partial (some apps misbehave under Xvfb without a real session bus) | ✅ |
Limitations
- X11 only. Wayland sessions cannot capture other-app windows or synthesize input via xdotool.
- Apps must export AT-SPI for
@eNrefs to populate. Most GTK / Qt apps do; Electron / Chromium need--force-renderer-accessibility. - Mouse cursor physically moves — no stealth pointer on X11.
- Dropped vs upstream:
move_mouse,drag,wait,double_click,arrange_window,navigate_browser,list_apps. Usekeypress,type_text, andcomputer_actionsto compose what you need.
Development
git clone https://github.com/tak-uukti/linux-computer-use
cd linux-computer-use
# Pi side (TypeScript)
npm install
npm run typecheck
# Bridge sanity
python3 -c "import ast; ast.parse(open('bridge/bridge.py').read())"
echo '{"id":"1","cmd":"list_windows"}' | python3 bridge/bridge.py
# MCP side
python3 -m venv .venv && .venv/bin/pip install -e .
.venv/bin/linux-computer-use-mcp # speaks MCP over stdio
The Pi extension API surface is stubbed locally in src/types.ts so typecheck runs without @mariozechner/pi-coding-agent installed.
Layout
.
├── assets/ logo + screenshots
├── bridge/
│ ├── bridge.py 471-line Python helper (AT-SPI + xdotool + scrot)
│ └── requirements.txt
├── extensions/
│ └── computer-use.ts Pi tool registration + JSON schemas
├── mcp_server/
│ ├── __init__.py
│ └── server.py FastMCP wrapper around the bridge (8 tools)
├── scripts/
│ └── setup-helper.mjs Pi postinstall — writes ~/.pi/.../bridge wrapper
├── skills/computer-use/SKILL.md pi skill — Quick Start + Pitfalls
├── src/
│ ├── bridge.ts Pi-side subprocess manager + JSON-line protocol
│ └── types.ts local stubs for the pi-coding-agent extension API
├── package.json npm metadata (Pi extension)
├── pyproject.toml MCP server packaging (uvx-installable)
├── tsconfig.json
├── CHANGELOG.md
├── LICENSE
└── README.md
Credits
@injaneity/pi-computer-use— macOS original, design and protocol shape.@mariozechner/pi-coding-agent— the Pi agent.- Model Context Protocol — Claude Code / OpenCode interop.
- AT-SPI 2, xdotool, wmctrl, scrot — the Linux building blocks doing all the real work.
License
MIT © 2026 Tak1tak · built by Tak1tak
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。