browser-use vs agent-browser vs Playwright Skills (2026)
A. Frans
Published October 26, 2026
Table of Contents
Browser automation skills have a counting problem. Four of the most popular ones carry 119,051, 44,535, 14,583 and 3,234 GitHub stars, and the star count tells you almost nothing about which one belongs in your project. They solve different problems with different amounts of control.
This comparison covers browser-use, agent-browser, playwright-cli and playwright-skill. The numbers come from our skills database as of late October 2026, and the behavior descriptions come from each project's published description. I haven't benchmarked them against each other, so I won't claim a speed winner.
Side by side
| Skill | Stars | License | What it is | Install |
|---|---|---|---|---|
| browser-use | 119,051 | MIT | Python library that makes websites readable for AI agents | claude skill add browser-use/browser-use |
| agent-browser | 44,535 | Apache-2.0 | Browser automation CLI for AI agents (vercel-labs) | claude skill add vercel-labs/agent-browser |
| playwright-cli | 14,583 | Apache-2.0 | CLI for common Playwright actions, from Microsoft | claude skill add microsoft/playwright-cli |
| playwright-skill | 3,234 | MIT | Model-invoked Playwright scripting | claude skill add lackeyjb/playwright-skill |
Two philosophies
The four split into two camps.
Agent drives the browser step by step. browser-use and agent-browser hand the model a view of the page and let it act: click this, type that, read the result. The model decides each move, which makes them good at sites you've never seen and bad at being predictable.
Agent writes a script, then runs it. playwright-skill has Claude write and execute custom Playwright scripts for testing, scraping and validation. playwright-cli sits between the camps: it's a command-line tool for common Playwright actions, including recording and generating Playwright code, inspecting selectors and taking screenshots.
That difference matters more than any feature on the list. A script can be reviewed, committed and rerun. A step-by-step agent run is closer to a conversation: useful, but harder to repeat exactly.
browser-use: the broad option
browser-use describes itself as making websites accessible to AI agents. Its tags include playwright, python and llm, so it sits on top of Playwright and gives the model a cleaned-up representation of the page.
Pick it when the task is open-ended: "find the cheapest flight on this site," "fill this form from a spreadsheet row," "check whether this page changed." The tradeoff is that you're trusting a model with live browsing. It's the right tool for exploration and a questionable one for anything you need to run identically every night.
Its popularity is real, and it's also the reason to read the repo's recent issues before adopting. A project that big gets fast-moving changes, and API surfaces shift.
agent-browser: the CLI approach
agent-browser, from vercel-labs, is a command-line interface for browser automation. A CLI shape suits Claude Code well, because Claude already works by running shell commands and reading output. There's no extra server to keep alive.
Choose it when you want agent-driven browsing without writing Python. If your project is JavaScript or TypeScript, or you don't want a Python dependency in the repo, this is the more natural fit.
playwright-cli: the official, auditable one
playwright-cli is the only one of the four with an "audited" security status in our database, and it's maintained by Microsoft. Its record-and-generate workflow is the standout feature: perform an action once, get Playwright code out, then commit that code.
That makes it the best pick for test authoring. You want deterministic tests, not an agent improvising each run. Generate the code, review it, and the agent's job is done.
playwright-skill: Claude as script author
playwright-skill is the smallest of the four and the most hands-off. It's model-invoked, so Claude decides when to reach for it and writes custom scripts on its own for testing, scraping and visual validation.
Pick it when you want to say "check that the checkout page still renders after my change" and have Claude write throwaway verification code. The smaller community means fewer eyes on the code, which brings us to security.
Security notes you shouldn't skip
Every one of these gives a model control of a real browser. Treat that as you would a contractor with your laptop.
- Use a clean browser profile. Never point automation at the profile where you're signed in to email, banking or cloud consoles. A page that carries hidden instructions can steer an agent that has your sessions.
- Read before installing. Our database labels three of the four "community-reviewed" and playwright-cli "audited". Community-reviewed is not a code audit. Skim the repo and the install script.
- Limit what it can reach. Run it in a container or a throwaway VM if the task touches anything sensitive.
- Watch for prompt injection. Text on a web page can read like instructions to the model. Never let an agent act on a page's text as if it came from you.
- Keep secrets out of scripts. Generated Playwright code may bake in whatever the agent typed during recording. Scan it before committing.
Which one should you pick?
| If you need... | Pick |
|---|---|
| Open-ended browsing on unfamiliar sites | browser-use |
| Agent-driven browsing from the shell, no Python | agent-browser |
| Repeatable tests with generated, reviewable code | playwright-cli |
| Quick throwaway verification scripts written by Claude | playwright-skill |
For hosted infrastructure instead of local browsers, Browserbase and Apify are the usual next step, and Zapier covers the no-code cases. If you're weighing MCP-based alternatives, our Playwright MCP vs Puppeteer MCP comparison and the wider browser automation skills list cover them, and headless browser tools for AI agents covers hosting.
A sensible install order
1. Start with playwright-cli. It's the lowest-risk and teaches you what Playwright can do. 2. Add browser-use only when a task really needs the model to improvise. 3. Add agent-browser if you want agent-driven browsing but hate the Python dependency. 4. Try playwright-skill last, and only inside a sandbox.
Run claude skill add for one at a time, check what it wrote to your skills folder, and test on a public site with nothing to lose before pointing it anywhere real.
FAQ
Which of these is the best browser automation skill overall? There isn't one. browser-use has by far the largest community, playwright-cli has the strongest security standing, and the right pick depends on whether you need exploration or repeatability.
Are these skills free? All four are listed as free and open source: MIT for browser-use and playwright-skill, Apache-2.0 for agent-browser and playwright-cli. You may still pay for model usage while they run.
Do they work outside Claude Code? Our database lists Claude Code as the compatible agent for all four. Some may work elsewhere, but check each repo's README.
Is it safe to let an AI agent browse while I'm logged in? Not on your everyday profile. Use a separate browser profile or a container, and don't expose accounts you can't afford to have touched.
Should I use a skill or an MCP server for browser automation? Skills load instructions and tools on demand, while MCP servers run as a separate process. Our Playwright MCP vs Puppeteer MCP article covers the MCP route.
Share this article
⚙Related Tools
📄Related Articles
Get More AI Tool Guides
New comparisons and guides every week. Join thousands of professionals staying ahead of the AI curve.