CUAWright plugin
Plugin for Claude Code and Codex: the agent writes and runs Playwright scripts, browses sites and distills reusable skills.
- 6,096 ★ on the repository
- Last commit
- MIT
Broad access
Why this level
- The agent writes and runs code in a browser and virtual machine without confirming each step, and web pages may carry hostile instructions.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add plugins/cuawrightDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- The agent writes and runs code in a browser and virtual machine without confirming each step, and web pages may carry hostile instructions.
This is third-party code. Review the repository files before installing.
What it does
CUAWright from Microsoft Research gives a coding model a terminal to drive a browser and a desktop. For the browser, the agent writes Playwright scripts, runs them, reads screenshots and logs and revises the code. Instead of one model call per action, a reusable script remains. The repository ships plugin manifests for Claude Code and Codex, a `cuawright-web` skill for browsers and a separate `cuawright desktop` mode for tasks inside an Ubuntu virtual machine (the OSWorld benchmark). A skill factory distills successful solves into parametrized code that reruns without a model.
Who it is for. Developers automating websites who want the agent to leave behind a verifiable script rather than a one-off session.
Good fit when
- Data from a site is needed and the script must run weekly.
- A date-driven form repeats and must be wrapped in a function.
- An agent is weak at browser tasks and needs a terminal loop with Playwright.
- A researcher reproduces Online-Mind2Web or OSWorld results.
Not a fit when
- A page only needs opening and reading and a plain browser MCP is enough.
- You have no right to automate the target site under its rules.
- Unattended work inside a real account is needed.
Example request
Open the site, find three matching plans and save a script that repeats the search with other dates.Limitations
The plugin needs `pip install -e .` and `playwright install chromium`. The desktop mode runs in an Ubuntu virtual machine and needs separate setup and extra dependencies. Models connect through OpenAI, Anthropic or OpenRouter, so keys and token costs apply. The agent executes its own code in a browser, so do not hand it working accounts and watch the requests. After the rename from Webwright, old commands still work.
How to disable. Run `/plugin uninstall cuawright@cuawright` in Claude Code and remove the marketplace source. Delete the virtual machine and the Python package separately.
FAQ
How do I install it in Claude Code?
Run `/plugin marketplace add microsoft/Webwright`, then `/plugin install cuawright@cuawright` and open a new session, since plugins load at session start.
How does CUAWright differ from Webwright?
It is the project new name since October 2026. The browser mode stays (`cuawright web`) and a desktop mode was added.
Which models are supported?
OpenAI, Anthropic and OpenRouter through small separate modules. Pick a provider and set your own key.
What is the skill factory?
A mode that turns a solved task script into verified parametrized code. Rerunning does not call a model or spend tokens.
What license does the project use?
MIT, according to GitHub. The sources are small separate modules under `src/cuawright/`.
Discussion
No questions or reviews yet.
Related
PluginsBetterWright keeps a persistent Playwright browser for agents with network rules, a credential vault and human handoff of steps.
A Claude Code plugin with an MCP server and seven skills: the agent drives a real browser by writing Python code
An MCP and CLI that let an agent control your real Chrome, with its own logins and extensions, through Playwright
Turns any website into a CLI command and lets an agent drive the user's already logged-in browser via Playwright
Agent browser on direct CDP without Playwright: login profiles, a CLI, an MCP server and recording workflows into a skill
An MCP server that lets an agent drive a real browser: open pages, click and fill forms through Playwright
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| CUAWright (this card) | no | 6,096 | ||
| Playwright MCP | yes | not required | 39,570 | an MCP server with one model call per action; CUAWright leaves a script for repeated runs |
| Browser Use | no | 120,809 | a Python browser agent library; CUAWright works through the coding model terminal and writes skills |
README in short
The README explains the terminal-for-web-agents idea, lists benchmark results, shows a quick start for browser and desktop and gives separate instructions for Claude Code and Codex.