agent-browser: a Claude Code skill by Vercel Labs
Official
For developers who want Claude to drive Chrome from the terminal to click, fill forms and take screenshots. Skip it if you cannot install a global npm tool.
High-risk warning At least one scanner rated it high risk, so we keep it out of the top ten. Read its code before you install it.
Before you install: Its working instructions come from the installed CLI at run time, not from the file you install, so you cannot read them in advance.
The skill's text changed on Nov 12, 2026, after we last read it. Our description may be behind.
- Score
- 95of 100
- Rank
- #11of 239
- Installs
- 1.4Mon skills.sh
- Security
- Warning 5 scanners
- Updated
- 2026
- Context
- ~234tokens always
Install
npx skills add https://github.com/vercel-labs/agent-browser --skill agent-browser Run it in your terminal. View the source
Using the Claude app rather than Claude Code? You can add a skill there too. How to add it in the app
What does agent-browser do?
agent-browser is a Claude skill that points Claude to the agent-browser command line tool, which controls Chrome or Chromium over the DevTools protocol. The skill file is only a short stub. Before any browser command, Claude runs agent-browser skills get core to load the usage guide that matches the installed version. Pages are read as accessibility-tree snapshots, and elements are picked by short refs such as @e1. Extra guides cover Electron apps like VS Code and Slack, exploratory QA, recording a HAR file to build an API client, and cloud browsers on Vercel Sandbox or AWS Bedrock AgentCore. A dashboard on port 4848 shows live sessions.
When should you use agent-browser?
Use it when
- Filling a web form or clicking through a site
- Taking screenshots or pulling text from a page
- Running an exploratory QA pass on your web app
Skip it when
- You cannot install global npm packages
- You want every instruction visible before you install
What it needs and what it can touch
You need
- Node.js and the agent-browser CLI
What it can do
- Runs agent-browser shell commands
- Opens websites and acts in a browser for you
- Can control desktop apps such as Slack or VS Code
From reading its instructions, not from running it.
Why it scores 95
Strongest on upkeep (100 of 100), weakest on trust (92).
Is agent-browser safe to install?
At least one scanner rated it high risk, so we keep it out of the top ten. Read its code before you install it. Scanners often miss what other scanners catch, and a skill can change after a scan. Read its source before you install it.
| Scanner | Result | Date |
|---|---|---|
| Gen Agent Trust Hub | Passed | Nov 3, 2026 |
| Socket | Passed | Nov 4, 2026 |
| Snyk | Medium risk | Nov 3, 2026 |
| Runlayer | High risk | Mar 15, 2026 |
| ZeroLeaks | Passed | Apr 16, 2026 |
Details
- Installs
- 1,411,072
- Repository stars
- 49,494
- Last repository change
- Tokens always loaded
- about 234
- Tokens when used
- about 602
- License
- Apache-2.0
- Data confidence
- High
Did it do the job for you?
Side-by-side comparisons
Alternatives to agent-browser
All skills for this job-
Get browser-usebrowser-use
For developers who want Claude to drive their own Chrome, or a hosted cloud browser, through the DevTools protocol.
-
Get opencli-browseropencli-browser
For developers who want Claude to work inside their logged-in Chrome through OpenCLI and capture site APIs.
-
Get agent-deviceagent-device
For mobile and TV app developers who want Claude to tap through apps on simulators and devices.
-
Get webapp-testingwebapp-testing
For developers who want Claude to click through a web app running on their own machine and see what breaks.
agent-browser Get agent-browser