Best AI Agent Skills for Browser Automation (2026)

You ask your agent to log into a dashboard, click through three filters, and screenshot the result. Without a browser skill, it guesses at the DOM and fails. With the right one, it reads the page like a screen reader and clicks the actual button. That gap is what this page is about.

Browser automation is the single most-installed category of agent skills right now, and also the messiest. Some ship as Claude Code skills you add with one command. Others run as MCP servers your agent talks to over a socket. A few do both. The trust levels range from Microsoft-official-and-audited down to a community Chrome extension that drives your real, logged-in browser. Knowing which is which matters before you hand any of them your session cookies.

Here are the eight worth your time in 2026, what each one actually does, and exactly how to install it.

Quick comparison

Three of these come from Microsoft, one from Anthropic, one from Vercel Labs, and the rest from independent maintainers with real review history. None carry a copyleft license that would complicate commercial use, which I cover further down.

playwright

The foundation everything else borrows from. Playwright is Microsoft's cross-browser automation framework, sitting at around 91k stars, and the most battle-tested option here for end-to-end web testing. If your agent needs to write and run real test suites against Chromium, Firefox, and WebKit, this is the base layer.

claude skill add microsoft/playwright

Repo: https://github.com/microsoft/playwright

Trust: Official Microsoft project, Apache-2.0, audited. The safest thing on this list to put in a production pipeline.

playwright-mcp

This is the one most agent builders actually want. Instead of bolting the full framework onto your project, playwright-mcp exposes browser control to your agent as an MCP server. It drives the page through the accessibility tree rather than screenshots, so your agent reads structured elements instead of pixel-guessing. At ~34k stars it's already the default browser MCP for a lot of Claude Code setups.

claude mcp add playwright -- npx @playwright/mcp@latest

Repo: https://github.com/microsoft/playwright-mcp

Trust: Official Microsoft, Apache-2.0, audited. No screenshots needed means lower token cost per action, which adds up fast on long browsing tasks.

browser-use

The most-starred project in this whole category at ~99k stars, and for good reason. browser-use exists to make websites legible to AI agents, it hands the model a clean view of clickable elements, forms, and text so it can navigate sites it's never seen. If you want an agent that books a flight or fills a multi-step form, start here.

claude skill add browser-use/browser-use

Repo: https://github.com/browser-use/browser-use

Trust: Verified, MIT-licensed, community-reviewed. MIT is the most permissive option on this page, so it's an easy yes for closed-source commercial work.

chrome-devtools-mcp

When the task shifts from "click things" to "why is this page slow," reach for this one. chrome-devtools-mcp wires the actual Chrome DevTools Protocol into your agent: performance traces, network waterfalls, console output. At ~44k stars it's the go-to for agents doing front-end debugging rather than plain navigation. Pairs naturally with the work in building a testing-first setup for Claude skills.

claude mcp add chrome-devtools-mcp -- npx -y ChromeDevTools/chrome-devtools-mcp

Repo: https://github.com/ChromeDevTools/chrome-devtools-mcp

Trust: Verified, Apache-2.0, community-reviewed. Maintained close to the DevTools team, which is why it tracks protocol changes well.

webapp-testing

Anthropic's own browser skill, and the cleanest entry point if you mostly care about testing your own app. webapp-testing wraps Playwright into a focused workflow: spin up the app, drive it, assert on what happened. Because it ships inside the official Anthropic skills repo, it follows the conventions Claude Code expects out of the box.

claude skill add anthropics/skills/webapp-testing

Repo: https://github.com/anthropics/skills/tree/main/skills/webapp-testing

Trust: Official Anthropic, MIT, audited. If you're picking one skill to trust blind, this is it.

agent-browser

Vercel Labs built this as a browser automation CLI aimed squarely at agents rather than humans. At ~36k stars, agent-browser gives your model a command-line surface to open pages, act on them, and pull results back, handy when you'd rather script the browser than embed a full MCP server. It slots well into developer workflows; our list of AI agent skills for developers goes deeper on that side.

claude skill add vercel-labs/agent-browser

Repo: https://github.com/vercel-labs/agent-browser

Trust: Verified, Apache-2.0, community-reviewed. Vercel's name plus open review history makes this a comfortable pick.

playwright-cli

Smaller at ~11k stars, but it solves a specific pain: writing the automation code in the first place. playwright-cli records what you do in a browser and generates the Playwright code to reproduce it, and it inspects live pages for selectors. Think of it as the scaffolding tool your agent uses before the real automation runs.

claude skill add microsoft/playwright-cli

Microsoft-maintained, Apache-2.0, audited. Trust: Official. Most useful paired with the main playwright skill rather than on its own.

mcp-chrome

The outlier, and the one to think hardest about. mcp-chrome runs as a Chrome extension that drives your real, already-logged-in browser. That's its whole appeal: no fresh headless session, no re-authenticating, your agent acts as you across sites where you're signed in. At ~12k stars it has real traction. It also means handing an agent your live sessions, so install it knowing exactly that.

claude mcp add mcp-chrome -- npx -y hangwin/mcp-chrome

Repo: https://github.com/hangwin/mcp-chrome

Trust: Verified, MIT, community-reviewed. The license and review are clean; the risk is operational, not legal, scope what you let it touch.

Skill install vs MCP server: which form to pick

The split on this page isn't random. A skill is code and instructions Claude Code loads directly into the agent — claude skill add pulls it in, and it lives alongside your other skills. An MCP server is a separate process your agent connects to; claude mcp add registers it, and it runs as its own service over the Model Context Protocol.

Rules of thumb that hold up in practice:

• Reach for the skill form when the work is self-contained and you want it versioned with your project, webapp-testing, playwright, agent-browser.
• Reach for the MCP form when you want a long-running browser the agent talks to across many turns, or when the tool already ships an MCP build, playwright-mcp and chrome-devtools-mcp are built MCP-first.
• A few tools, like Playwright, exist in both shapes. Match the shape to how your agent runs, not to which has more stars.

MCP servers cost a little more setup but give you a persistent browser session, which matters for multi-step tasks. Skills are lighter to add and easier to commit to a repo. For a wider view of the persistent-session tradeoff, the agentic browsers roundup and the headless browser tools comparison both go deeper than I can here.

When AGPL licensing actually matters

None of the eight skills above carry a copyleft license, which is deliberate on my part, for commercial work, that's the safe pool. But you'll meet AGPL-licensed browser tools elsewhere, and the distinction is worth holding.

Apache-2.0 and MIT (every license on this page) let you build closed-source commercial products on top with no obligation to publish your code. AGPL is different: if you run AGPL software as a network service, you may have to release your own source to your users. For an internal automation script, nobody cares. For a SaaS product where the browser tool is part of what you sell over the network, AGPL can force your hand. Check the license before a tool reaches production, not after. When the license is MIT or Apache-2.0, like all eight here, you can stop worrying and ship.

If you're staffing a wider toolkit, our guide to the best AI tools for developers covers the non-browser pieces that pair with these.

My picks

If you want one MCP server for general agent browsing, install playwright-mcp — official, audited, accessibility-driven, cheap on tokens. If you're testing your own app, webapp-testing is the cleanest start. If you need an agent to roam unfamiliar sites and fill forms, browser-use earns its 99k stars. Add chrome-devtools-mcp only when you hit performance questions, and treat mcp-chrome as a power tool you scope carefully.

FAQ

What's the difference between a Claude skill and an MCP server for browser automation? A skill is loaded into the agent directly with claude skill add and lives with your project. An MCP server runs as a separate process the agent connects to via claude mcp add. Skills are lighter to commit to a repo; MCP servers give you a persistent browser session across many turns.

Which browser automation skill is safest to install? The four Apache-2.0 and MIT projects from Microsoft and Anthropic, playwright, playwright-mcp, webapp-testing, and playwright-cli, are official and audited. webapp-testing from Anthropic is the one to trust with the least scrutiny.

Do I need Playwright if I install playwright-mcp? No. playwright-mcp bundles what it needs to drive the browser through the accessibility tree. You'd add the full playwright skill only when your agent has to author and run complete test suites, not just navigate pages.

Is it safe to let mcp-chrome use my logged-in browser? The license and code review are clean, so the legal and supply-chain risk is low. The real risk is operational: it acts inside your live sessions. Limit which sites and tabs it can touch, and don't point it at anything you wouldn't let a script click on your behalf.

Can I use these in a commercial product? Yes. Every skill on this page is MIT or Apache-2.0, so you can build closed-source commercial software on top with no obligation to publish your code. Watch for AGPL only on browser tools from outside this list.