Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

Open Computer Use

Local MCP server that drives desktop apps in the background through accessibility APIs, with one toolset on macOS, Windows and Linux.

Reviewed by ReviewRamble editors · Updated

Setup docs Open source

Share this page:XLinkedInFacebookRedditWhatsAppEmail

Our rating

3.5/5

A tidy, cautious computer-use server whose macOS side is solid and whose Windows and Linux sides are still, by its own account, experimental.

What is Open Computer Use?

Open Computer Use, from the GitHub user opensymph, lets an agent see and operate the apps on your own computer. It reads a window's accessibility tree, the labelled list of buttons and fields that screen readers use, plus a screenshot, then clicks, types, scrolls, drags and sets values by element number or by coordinates. Input goes to the target app rather than through your real pointer and keyboard, so you can keep working while it acts. Its tool names and arguments follow the computer-use plugin built into OpenAI's Codex.

Nine core tools are identical on macOS, Windows and Linux; five window-level tools work on macOS and Windows, and macOS adds a full-display capture. The macOS runtime is written in Swift and the Windows and Linux runtimes in Go, all shipped in one npm package. The project is MIT-licensed, was created on 23 August 2026, has about 25 stars and 97 commits, and released v1.3.0 on 17 October 2026.

How to install Open Computer Use

npm i -g @opensymph/open-computer-use && ocu install-claude-mcp

Is Open Computer Use worth it?

Open Computer Use is built around one promise: the agent works in an app while you keep the mouse. On macOS the default click path tries the app's own accessibility actions first, then sends mouse events to that app's process alone; the real pointer only moves if you set OPEN_COMPUTER_USE_ALLOW_GLOBAL_POINTER_FALLBACKS=1. A visible software cursor shows where the agent is acting, and the signed macOS app keeps its Accessibility and Screen Recording grants across upgrades. One contract on three systems means a single prompt can drive the same app on a Mac, a Windows PC and a Linux desktop.

The guard rails are sensible defaults rather than a sandbox. Password managers are always refused. On Windows, the Windows key is blocked, terminals and security apps are on a deny list, and launching apps, stealing focus and foreground input each need their own environment variable. The project's security policy says plainly that the tool has high privileges by nature, should run only on machines you own or are authorised to use, and leaves risks from how you use it out of scope. Anything the agent can see, it can act on: an email or web page it reads is a possible prompt injection with your desktop behind it.

Data stays local in the transport sense. The server speaks stdio, opens no network port, and on macOS reaches its helper app over a local Unix socket. Window screenshots and accessibility text still go to whichever model your agent uses, and the README says nothing about telemetry. Separate command-line tools can capture or record the whole desktop and send global input, but they sit outside the MCP tool list and the input commands are off unless enabled.

The catch: the project's architecture notes call the Windows and Linux runtimes experimental. Windows has no code signing and no visible cursor, and the project's own timings put a typed entry at about 2.0 s and listing apps at about 3.5 s. Linux needs a signed-in desktop session, the window-level tools return "not supported yet", and screenshots fail or come back black under GNOME on Wayland. With about 25 stars, three contributors and no release since 17 October 2026, it is young and quiet.

Our take: for a Mac user who wants Codex-style background computer use in Claude Code or Gemini CLI, this is a careful and easy install; on Windows or Linux, treat it as a promising preview and give it a spare user account, not your main desktop.

Where it earns its place

  • A Mac-based iOS developer at a small studio

    Click through a settings flow in the Simulator and a design app while still typing in the editor

    The agent reads each window with get_app_state and clicks by element number in the background, with the software cursor showing where it acts. Free apart from model tokens.

  • An operations analyst at a 30-person firm with a desktop-only ERP client on Windows

    Copy a weekly report's figures into a spreadsheet without retyping them

    The agent reads the report window through UI Automation and types into the spreadsheet with type_text. Run it on a spare account: the Windows runtime is experimental and each typed entry takes about 2.0 s in the project's timings.

  • A QA lead testing a GTK app on a Linux desktop

    Run the same smoke test the team already uses on macOS

    Because the nine core tools share names and arguments, one script drives both. Use an X11 session: under GNOME on Wayland screenshots fail and coordinate clicks are best effort.

Before you commit

Setup
npm i -g @opensymph/open-computer-use, then ocu doctor (on macOS 14 or later it prompts once for Accessibility and Screen Recording), then ocu install-claude-mcp, install-codex-mcp, install-gemini-mcp or install-opencode-mcp. Any other client runs open-computer-use mcp over stdio.
Integrations
Nine core tools (list_apps, get_app_state, click, perform_secondary_action, scroll, drag, type_text, press_key, set_value), five window tools on macOS and Windows, capture_display on macOS. Also a Codex plugin, a ZCode plugin, an agent skill and the ocu call command for scripting without a client.
Data and privacy
Runs locally over stdio and opens no network port; on macOS the CLI reaches its helper app through a local Unix socket. Screenshots and accessibility text go to your agent's model provider. The README mentions no telemetry. Security fixes are offered only for the latest npm release.
Cost at scale
Free; cost is model tokens. Accessibility trees are capped by default at 1,200 nodes and 64 levels and screenshots are downscaled, which keeps each read bounded, but long tasks still run to many calls.
Limits
Windows has no code signing or visible cursor and needs a signed-in desktop session; over SSH list_apps can come back empty. Linux window-level tools are not supported yet and Wayland screenshots fail or are black. macOS needs version 14 or later.

What’s good

  • Background input on macOS without moving your pointer
  • Same nine tools on macOS, Windows and Linux
  • Global input, app launch and focus behind opt-in flags
  • Password managers refused; signed macOS app

What’s not

  • Windows and Linux runtimes labelled experimental
  • Linux window tools unsupported; Wayland screenshots fail
  • Agent can read and act in any app it can see
  • Young project, three contributors, about 25 stars

Best for

Mac developers who want an agent to operate native apps in the background from Claude Code, Codex or Gemini CLI.

Open Computer Use alternatives

Open Computer Use: common questions

Is Open Computer Use worth it?
A tidy, cautious computer-use server whose macOS side is solid and whose Windows and Linux sides are still, by its own account, experimental. We rate it 3.5 out of 5.
How much does Open Computer Use cost?
Free and open source (MIT). No paid tier; model calls are paid to your own provider.
What are the downsides of Open Computer Use?
Windows and Linux runtimes labelled experimental; linux window tools unsupported; Wayland screenshots fail; agent can read and act in any app it can see; young project, three contributors, about 25 stars.
Who is Open Computer Use best for?
Mac developers who want an agent to operate native apps in the background from Claude Code, Codex or Gemini CLI.
How much does Open Computer Use cost for a team or in production?
Free; cost is model tokens. Accessibility trees are capped by default at 1,200 nodes and 64 levels and screenshots are downscaled, which keeps each read bounded, but long tasks still run to many calls.
Is Open Computer Use safe for company data?
Runs locally over stdio and opens no network port; on macOS the CLI reaches its helper app through a local Unix socket. Screenshots and accessibility text go to your agent's model provider. The README mentions no telemetry. Security fixes are offered only for the latest npm release.
What are the best Open Computer Use alternatives?
The closest alternatives we have reviewed are agent-desktop and Munim Computer Use.

Answers checked Nov 18, 2026.