chrome-use CLI
An agent drives your real logged-in Chrome: a CLI, MCP and an extension with a separate tab group per session
- 257 ★ on the repository
- Last commit
- Apache-2.0
Broad access
Why this level
- The agent acts in your Chrome with your logins: mail, payments and admin panels are open to it as they are to you
- New side panel conversations start with full access and no permission prompts by default
- The stealth and humanize modes target sites with bot protection; using them on third-party services can break the site rules and the law
Install
- Works in
Manual install
curl -fsSL https://raw.githubusercontent.com/leeguooooo/chrome-use/main/install.sh | shmacOS and Linux. For Windows there is `irm https://raw.githubusercontent.com/leeguooooo/chrome-use/main/install.ps1 | iex`. Then the extension and `chrome-use extension install`.
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- The agent acts in your Chrome with your logins: mail, payments and admin panels are open to it as they are to you
- New side panel conversations start with full access and no permission prompts by default
- The stealth and humanize modes target sites with bot protection; using them on third-party services can break the site rules and the law
This is third-party code. Review the repository files before installing.
What it does
chrome-use is a Rust CLI that hands an agent your real, already logged-in Chrome. The extension talks to the command through native messaging, with no debug port and no network socket, so cookies, sessions and the browser fingerprint stay genuine. The owner watches the work in their own window and takes over at a captcha or a two-factor prompt. Each agent session gets its own colored tab group, so several agents share one Chrome without interfering with each other. Commands: `chrome-use open <url>`, DOM and accessibility-tree snapshots, clicks, typing, screenshots, and recording actions for a replay without a model through `chrome-use teach`. The skill is installed with `chrome-use skills install` or the Claude Code plugin, clients without a shell get the MCP server `chrome-use mcp`, and the extension side panel can launch Codex, Claude Code, OpenCode or Pi.
Who it is for. People who want an agent to work with sites under their own accounts (dashboards, mail, internal services) without a separate browser and a fresh login.
Good fit when
- A service demands a login with two-factor confirmation, and repeating it in a clean test browser is a chore
- Several agents have to work in one Chrome without touching each other tabs
- A page has to become structured data in one command, for example a list of Hacker News stories
- An interface change is easier to show with a click: Inspect & Ask sends the agent the element, the source line and a screenshot
- A browser scenario repeats, and it is handy to record it once by demonstration and replay it without a model
Not a fit when
- A server without a display: Chrome with a window on a virtual display is recommended, because headless mode still gives itself away
- You need a clean isolated environment for tests: chrome-use works in your profile and sees all your logins
- Automating banks, brokerages and government sites without you present: on these chrome-use asks for confirmation of actions by default
Example request
Open the pricing page of this service in my Chrome, compare the three plans and print a tableLimitations
The extension from the Chrome Web Store and a one-time `chrome-use extension install` are required. On agent tabs Chrome shows the bar "started debugging this browser"; `chrome-use extension connect --silent` removes it at the cost of one Chrome restart. Bot-detector results in the README are the author measurements and do not guarantee passing on a given site. The commands check for updates once a day and install them in the background; `CHROME_USE_AUTO_UPGRADE=0` turns that off. The installer runs through `curl ... | sh`.
How to disable. Remove the chrome-use extension in Chrome, then the `chrome-use` binary and the `~/.chrome-use` directory with the panel history. The skill goes away from `~/.agents/skills`, `.claude/skills` and the directories of other agents.
FAQ
How do I install chrome-use for Claude Code?
Best through the plugin: `/plugin marketplace add leeguooooo/plugins`, then `/plugin install chrome-use@leeguooooo-plugins`. It installs the skill for all projects and keeps it updated. The binary and the extension are still needed.
How do I connect chrome-use to Claude Desktop?
Through MCP: add a `chrome-use` server to `claude_desktop_config.json` with the command `chrome-use` and the argument `mcp`. In this mode every result is fenced as untrusted page content.
Does chrome-use protect against hidden instructions on pages?
Partly. `snapshot`, `get text` and `read` drop invisible text by default and report how much was left out, and MCP mode and the side panel wrap the result in an untrusted-data label. Agents that never ask before acting, such as Pi and full access in Claude and Codex, are not covered.
What does humanize mode do?
The `--humanize off|fast|human` flag controls pauses and pointer movement. The default is `fast`, `human` adds a seeded 60 to 200 ms pause before a press, and `off` inserts text without key events, which behavior detectors flag.
How do I work with several Chrome profiles?
With the `--browser` flag or ChooseBrowser rules on macOS: when it is installed, `chrome-use open` follows the rule for the site and says so. `chrome-use doctor` shows the state of every profile.
Discussion
No questions or reviews yet.
Related
CLI toolsA CLI, Chrome extension, SDKs and MCP server that let an agent drive your real Chrome, whatever the agent runtime
A skill connects an agent to your Chrome through an extension and CLI for reading pages, managing tabs and verifying actions.
A skill for Amp, Claude Code and Cursor: the agent drives your live Chrome with its open tabs and signed-in accounts through a single debugging toggle
A CLI and skill for driving Chrome from an agent: page snapshots, clicks, console, network and audits in compact output
An MCP server and Chrome extension give any agent 22 tools for your own logged-in browser, verifying the effect of every action.
Claude Code, Codex and Cursor agents work in your everyday Chrome with real logins, in background tabs, with an audit trail.
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| chrome-use (this card) | no | 257 | ||
| agent-browser | yes | not required | 50,530 | The project grew out of agent-browser and kept its CLI, but connects to your real Chrome through an extension instead of launching its own browser. |
| Chrome DevTools MCP | yes | not required | 55,199 | Chrome DevTools MCP targets page debugging over the DevTools protocol, while chrome-use works through an extension and native messaging, with no debug port. |
README in short
The README compares chrome-use with Playwright, Puppeteer, connecting through a debug port and the Claude in Chrome extension, then covers install, connecting to Chrome, the agent side panel, Inspect & Ask, teach by demonstration and prompt-injection defenses. Separate sections cover anti-detection and permissions.