
Open Computer Use MCP Server: Setup for Claude Code, Codex and Gemini CLI
Lets Claude Code, Codex and Gemini CLI look at desktop apps and click, type and scroll in them for you.
Open Computer Use is a community, open-source version of OpenAI's Codex Computer Use. It lets an assistant read an app's window through accessibility data and screenshots, then click, type and scroll. It aims to work without taking over your mouse.
It runs on your computer and needs no account. On macOS 14 or later you must grant Accessibility and Screen Recording once. Windows and Linux support exists, and the project calls its Linux runtime experimental.
Good for
- Codex and Claude Code users who want computer use without a closed-source add-on
- Developers automating desktop apps that have no API
- Mac users who want an agent to keep working while they use other apps
Works in
12 should work
Tap an app to see how to set it up and any known problems.
Supports local MCP servers.
Supports local MCP servers.
Supports local MCP servers.
Supports local MCP servers.
Supports local MCP servers.
Supports local MCP servers.
Supports local MCP servers.
Supports local MCP servers.
Supports local MCP servers.
Supports local MCP servers.
Supports local MCP servers.
Supports local MCP servers.
Client does not support this transport.
Client does not support this transport.
Install Open Computer Use MCP
Claude Code
- 1
Copy this prompt.
Add the Open Computer Use MCP server to Claude Code. Run this command: npm i -g open-computer-use open-computer-use install-claude-mcp Then check that it loaded and tell me if I need to sign in or restart.
- 2
Paste it into Claude Code. It shows the change and asks before running it.
Requires Node.js and npm (global install), macOS 14+, or Linux or Windows, On macOS: Accessibility and Screen Recording permission.
Read-only mode: No read-only mode.
Install the package globally first with npm i -g open-computer-use, then run open-computer-use once on macOS to grant Accessibility and Screen Recording.
From the project README. Source.
Tools
Lists the apps you have running or used lately
Returns the apps that are running now, plus any used in the last 14 days with how often.
Try: “Which apps do I have open?”
Read ·
list_appsLooks at an app's window
Starts a session with an app and returns a screenshot plus the window's accessibility tree. The assistant must call it once per turn before it acts on an app. On macOS it can also park the window on a hidden display so it keeps working.
Try: “Look at what TextEdit is showing right now.”
Read ·
get_app_stateFinds controls in an app window (macOS)
Searches the current window for buttons and other controls by text or role, without taking a full snapshot. Each match gets an index the click and value tools can use.
Try: “Find the Save button in TextEdit.”
Read ·
queryClicks an element or a point
Clicks a control by its index or by screenshot coordinates. You can choose how the click is sent. The system-wide pointer path moves your real mouse and is off unless an environment setting allows it.
Try: “Click the Save button.”
Runs code ·
clickRuns an extra action on a control
Triggers a secondary accessibility action on an element, such as the actions a control offers beyond a plain click.
Try: “Open the context menu for the selected file.”
Runs code ·
perform_secondary_actionTypes text with the keyboard
Types literal text into the app. The text goes to whichever field has focus.
Try: “Type a short thank-you note.”
Runs code ·
type_textPresses a key or shortcut
Presses a key or key combination, including modifiers and navigation keys, in the xdotool key style.
Try: “Press Cmd+S to save.”
Runs code ·
press_keySets the value of a field directly
Writes a value into a control that accepts one, like a text field or slider, without typing.
Try: “Set the zoom slider to 150%.”
Runs code ·
set_valueSelects text or places the cursor (macOS)
Selects text inside a text element, or puts the cursor before or after it. Add surrounding text to pick between repeats.
Try: “Select the word total in the second paragraph.”
Runs code ·
select_textScrolls an element
Scrolls an element up, down, left or right by a number of pages. Fractions work.
Try: “Scroll down half a page.”
Runs code ·
scrollDrags from one point to another
Drags using screenshot coordinates. By default events go straight to the app and your pointer stays put. Window moves and Finder drag-and-drop need the system-wide pointer setting.
Try: “Drag the slider handle to the right.”
Runs code ·
drag
Read from the source code
Permissions
high risk · 5 things to watch
High risk
The assistant can operate any desktop app you point it at, as you, including sending, buying and deleting. Use it on a spare account or a machine without sensitive apps open.
What it can do
- Read your data
- Create and change things
- Delete things
- Read-only mode
Watch out for
It can do anything you can do in an app
It clicks, types and sets values in any app you name, so it can send messages, buy things or delete files. Run it where that is acceptable and approve actions you do not trust. github.com
Screenshots of your apps go to your AI provider
The tools return screenshots and window text to the model. Close apps that show private data first. The project says it does not upload them anywhere else. github.com
Password managers are blocked, other apps are not
A built-in list blocks 1Password, Bitwarden, Dashlane, LastPass, NordPass and Proton Pass. The project says it has no session approval yet. github.com
Locked Use can unlock your Mac
This opt-in macOS mode unlocks a locked Mac for the agent, behind an admin-installed service. The project marks it experimental. Leave it off unless you need it. github.com
get_app_state can hide an app window
With window placement set to agent_display on macOS, the tool moves the app's window to an invisible display. It disappears from your desktop until restore or until the process exits. github.com
Security advisories
No published security advisories as of Nov 1, 2026.
Technical details
| Turning on read-only | No read-only mode is documented. |
|---|---|
| Sign-in and keys | No account or key. On macOS, the Open Computer Use app holds the system permissions. |
| Scopes |
|
Questions
Is Open Computer Use official?
No. It is a community project by iFurySt, inspired by OpenAI's Codex Computer Use. It is not made by OpenAI, Anthropic or Google.
Which systems does it run on?
macOS 14 or later, Linux and Windows. The project describes the Linux runtime as experimental. On macOS you grant Accessibility and Screen Recording once.
Do I need an API key?
No. It runs on your computer with no account. You only need an AI app that can run local MCP servers.
Will it take over my mouse?
By default it sends events straight to the target app and leaves your pointer alone. A system-wide pointer mode exists, but it stays off unless you set an environment variable.
Can it read my passwords?
It blocks six password managers by name. It can still see and use any other app you name, so keep sensitive windows closed.
Do the tools differ by system?
Yes. The query and select_text tools are macOS tools. This page lists the 11 tools defined in the v1.1.3 macOS source. The npm README still says the MCP server exposes a 9-tool surface, so a Windows or Linux install may show fewer.
How do I add it to Claude Code?
Run npm i -g open-computer-use and then open-computer-use install-claude-mcp, as the README shows. The install box above has the same steps and a project config alternative.
Information
updated Nov 2, 2026
- Publisher
- Community
- Package
open-computer-use- Registry
- npm
- Transport
- stdio
- Auth
- None
- Version
- 1.1.3
- License
- MIT
- Last commit
- Oct 30, 2026
- Open issues
- 17
Sources
Docs, issue trackers and discussions this page is built from. Issues, quotes and findings above link to their own source.
Show all 3Hide
- github.com · github.com
- github.com · github.com
- github.com · github.com