Skip to main content
Glama

Mirasim Computer Use

English | 简体中文 | 日本語 | 한국어

npm License: Apache-2.0 Benchmark: DevBrowse Bench on Hugging Face

Computer use for Claude Code, Codex and any MCP agent: an open-source MCP server that drives browsers, macOS apps and Electron apps through their accessibility trees. It runs no model of its own and needs no API key. The Mirasim desktop app ships the same tools built in (Settings → Computer Use).

Claude Code checks its change in the browser with mirasim-cu, finds the Add task button covered by a new header layer, fixes the CSS and checks again

A real run of 24 s, shown in 15 s with the model's waits cut: Claude Code checks its change in the browser, finds a control covered by a new header layer, fixes the CSS and checks again in one call. The red and green marks are added for the video. MP4

  • One call runs a whole chain of steps, and every step reports back. The agent writes a short script (open, type, pick, click, wait, read), and the result says what each step did and ends with one observation, so the model does not look at the screen after every click. The screen reads back as text, a numbered list of elements from the accessibility tree; pixels only when text is not enough.

  • A click lands where a user's mouse would. A click on a control under another layer fails and names the layer in the way («fixed bar at the top of the viewport»), instead of reaching the control a user could not click.

  • Your own browser and your desktop. It drives the Chrome you are signed into through the Mirasim extension, native macOS apps in the background without taking your mouse and keyboard, and Electron apps over their debugging port.

On DevBrowse Bench v1 it took 26% less time and 45% fewer model requests than ChatGPT Computer Use, with half the input tokens.

Install

Needs Node.js 22.19 or later.

claude mcp add mirasim-cu -- npx -y @mirasim/computer-use mcp   # Claude Code
codex mcp add mirasim-cu -- npx -y @mirasim/computer-use mcp    # Codex

Then ask your agent to check something in the mirasim-cu browser. Codex asks before every MCP call until you approve the two tools in ~/.codex/config.toml; that, your signed-in Chrome, macOS and Electron apps, the agent skill and the command line are in docs/setup.md.

Related MCP server: macinput

How it compares

What each project documents, read on 2026-11-18 (sources below the table).

Mirasim Computer Use

Playwright MCP

browser-use

Anthropic computer use

ChatGPT Computer Use

Open source

Apache-2.0

Apache-2.0

MIT; Browser Use Cloud is a paid service

No: an API tool (MIT reference demo)

No: part of the ChatGPT desktop app

MCP server for Claude Code and Codex

Yes

Yes

Yes (browser-use --cli-mcp)

No: an API tool; built into Claude Code

No: inside the ChatGPT desktop app

Model of its own

None; your agent drives it

None

Its agent has one; the MCP mode does not

Claude

OpenAI's models

What the model reads

Accessibility tree as text; screenshots on request

Accessibility snapshot as text; vision opt-in

What its Python prints, plus screenshots

Screenshots and pixel coordinates

Screenshots; more is not documented

Many steps in one call

Yes: a script, each step reported

One tool per action; browser_run_code_unsafe runs a Playwright script

Yes: browser_exec runs Python

Several actions per turn (a batch), one tool call each

Not documented

A click on a covered control

Fails and names the covering layer

Waits for the element to receive the click, and times out if it never does

Not documented

Not documented

Not documented

Your signed-in browser

Mirasim extension

Playwright extension (Chrome, Edge) or a CDP endpoint

Running Chrome through remote debugging

Claude in Chrome extension

ChatGPT extension

Native desktop apps

macOS, in the background

No

No (a separate, experimental macOS project)

Any app in your VM or container; macOS in Claude Code (preview)

macOS (in the background) and Windows

Electron apps

Yes, over the debugging port

Not in the MCP server

Not documented

Through screen control in Claude Code

Not documented

DevBrowse Bench v1

−26% time, −45% model requests

Not run

Not run

Not run

The baseline

Benchmark

DevBrowse Bench v1 core results: time −26%, model requests −45%, input tokens −50%, failed runs 4 : 8

DevBrowse Bench measures browser use in development work: a coding agent changes a web app, then checks its own work in a browser. v1 has 60 tasks in three parts, the three panels of the figure. 51 are on Team Tasks, a simulated React + Vite task manager written for the benchmark: 44 classic and extended tasks (features, bug fixes, QA passes) and 7 interaction tasks ★, bugs where a control the user can see cannot be clicked. 9 are on real projects: merged bug fixes from the open-source repositories of Excalidraw, Lexical, tldraw and react-admin.

Each task runs as a pair, both sides at the same time in Codex in the ChatGPT desktop app, with the same model (gpt-6.1-sol at medium reasoning effort), the same prompt and the same starting code. The only difference is the browser tool the agent is told to use: ChatGPT Computer Use (the in-app browser of Codex), or Mirasim Computer Use 0.1.0-rc.5 as an MCP server. 200 runs per side, on 2026-11-06 and 2026-11-07; time, model requests and tokens come from Codex's own session record, and every run was graded by rule-based judges.

ChatGPT Computer Use

Mirasim Computer Use

Runs passed

192 / 200

196 / 200

Over the 192 pairs where both sides passed (change = Mirasim total / ChatGPT total − 1, 95% interval from resampling tasks):

Metric

Change

95% interval

Time

−26%

−32% to −20%

Model requests

−45%

−48% to −42%

Input tokens

−50%

−54% to −46%

Output tokens

−8%

−13% to −3%

Browser calls

−37%

−41% to −32%

Most of the time was saved in round trips: the agent made 45% fewer model requests, while the output tokens it wrote fell by only 8%. observe() returns elements as data, so the agent puts its checks in the script and gets back a verdict instead of reading the page after every click. And a click lands where a user's mouse would: a click on a covered control fails and names the layer that took it. The four extra failures on the other side are all one interaction task ★, a QA pass where a transparent layer under a sticky header covers a toolbar for mouse clicks: the report written with ChatGPT Computer Use missed it in 4 of 5 runs, the one written with Mirasim Computer Use in none.

These numbers describe the two tools as they were on those two days. Around the end of the runs, ChatGPT Computer Use changed how it uses the browser and began to need fewer model requests per run. When 39 of the 60 tasks were run again the same way on 2026-11-10 and 2026-11-11, with a later release of Mirasim Computer Use, the time saving on them was about −16%, against −25% for the same tasks here. That rerun is not part of v1.

Tasks, judges, every run and the full report are on Hugging Face, in the devbrowse-bench dataset (v1 report).

Tools

Two tools, assembled per driver at session start:

tool

role

computer_repl

JavaScript over the GUI in a PERSISTENT session: reads (observe(), find(name), text(), page(), screenshot(), diagnostics()), actions (click, hover, setText, keys, scroll, wait, focus + the driver's verbs — openUrl, newTab, selectOption, paste, launchApp, menu, …), and surface HANDLES (computer.surfaces() / computer.surface(id)) that bind verbs to one tab / window / app. Top-level declarations persist across calls; the first result carries the API reference; every run ends with one observation and a changed/frozen verdict

computer_repl_reset

discard the JavaScript bindings — surfaces untouched

Element observations are incremental by default: when little changed since the previous one, only the diff is sent (refs are sticky, so unchanged lines drop out), with a full list re-sent periodically.

License

Apache License 2.0 © MIRASIM, INC. Parts of the browser extension bridge are adapted from Playwright (Apache-2.0, © Microsoft Corporation); see NOTICE. The Windows helper's statically linked Rust crates and their licenses are listed in THIRD_PARTY_NOTICES.

"Mirasim" and the Mirasim logo are trademarks of MIRASIM, INC. and are not licensed under the Apache License. Chrome, Claude, Codex, ChatGPT, macOS and Windows are trademarks of their respective owners; this project is not affiliated with or endorsed by them.

Tool Schema Changelog

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

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      D
      maintenance
      An open-source MCP server for macOS and Windows that provides native desktop control via Accessibility APIs, OCR, and Chrome CDP. It enables AI agents to interact with applications, manage browser sessions, and automate workflows with high-speed native UI actions.
      49 npm
      15
      AGPL 3.0
    • A
      license
      Not graded
      quality
      D
      maintenance
      An MCP server for macOS that enables AI agents to control the desktop GUI through keyboard input, mouse actions, and screen captures. It provides stable low-level primitives for UI automation and agent-driven desktop workflows.
      MIT
    • F
      license
      Not graded
      quality
      D
      maintenance
      macOS MCP server that enables AI agents to directly control the host OS, including mouse, keyboard, windows, files, and accessibility automation for computer-use workflows.
      1
      -
    • A
      license
      Not graded
      quality
      C
      maintenance
      MCP server that gives AI agents hands and eyes on macOS, enabling them to see and operate native and Electron applications via structured accessibility queries, screenshots, OCR, and application-specific skills.
      4
      MIT