Browser Agents: 11 Tools That Let an AI Agent Use the Web

A browser agent has two halves: a brain that decides what to do next and a pair of hands that does it in a real browser. Most products sell one half. Frameworks like Browser Use and Stagehand are the brain, with a model of your choice. Control layers like agent-browser and Playwright MCP are the hands, meant for a coding agent that already has a brain. Hosted browsers from Browserbase, Browserless and Cloudflare are the arms, keeping a fleet of Chrome instances alive at scale. This list covers all three kinds, plus the data APIs that skip the browser when a page can be fetched instead, and Anthropic's own browser agent for people rather than programs.

Browser agents are AI agents that operate a web browser the way a person does, reading pages and then clicking, typing and submitting forms to finish a task. The tools below split into frameworks that decide what to do, control layers that execute it for a coding agent, hosted browsers that run it at scale, and data APIs that skip the browser.

The 11 browser agents and browser tools compared

Browser Use

Browser Use is the most-used open-source browser agent framework. It is a Python library (Python 3.11 or newer, installed with uv add browser-use) in which you hand an Agent a task string and an LLM and it opens pages, clicks, types and fills forms until the task is done. The README's quickstart shows both its own ChatBrowserUse model wrapper and plain Anthropic or OpenAI clients, and the project publishes a 100-task benchmark as open source. There is also a skill path for coding agents: paste one prompt into Claude Code, Codex or Cursor and it installs the library, runs browser-use skill install and connects to your browser. The repository has 10,220 commits, an MIT licence, and a server.json for the MCP registry added on 16 July 2026. It suits developers who want to own the agent loop and pick the model; you bring the browser infrastructure or pair it with a hosted provider.

Stagehand

Stagehand is Browserbase's SDK for browser agents, and its pitch is precision. You get three AI primitives, act ("click the login button"), extract (typed data against a schema you define) and observe (discover available actions), next to Playwright-style page methods (goto, click, locator, screenshot) for the steps where you already know the selector. The same script can mix both, so a flaky selector becomes a natural-language step and everything else stays deterministic. Version 4 ships first-class TypeScript, Python and Go SDKs, drives the browser over the Chrome DevTools Protocol with no Playwright or Puppeteer dependency, and reports per-method token usage through metrics(). Browserbase recommends running it on its own hosted browsers, which is what enables server-side caching and its model gateway. Stagehand is the hands; it expects you to bring the brain, whether that is LangChain, CrewAI, Mastra or a loop you wrote.

agent-browser

agent-browser is a browser automation CLI from Vercel Labs written in Rust and designed for AI agents that can run shell commands. Its snapshot command returns a compact accessibility tree where each element carries a ref like @e1, and the agent then runs agent-browser click @e2 or agent-browser screenshot page.png. The project claims the text output uses about 200 to 400 tokens against 3,000 to 5,000 for a full DOM, which is the whole point: a coding agent spends its context on your code, not on HTML. There are more than 50 commands covering navigation, forms, network, storage, tabs, frames and debugging, plus video recording, streaming and a profiler. It installs with npm install -g agent-browser or brew install agent-browser and works with Claude Code, Cursor, GitHub Copilot, OpenAI Codex, Gemini and opencode. Firecrawl's own comparison notes the trade-off: a CLI means more overhead per action than an in-process SDK, and there is no reasoning layer, so your agent framework decides what to do.

Playwright MCP

Playwright MCP is Microsoft's MCP server that exposes a Playwright-driven browser to any MCP client, published as @playwright/mcp. Because it speaks MCP, it plugs into Claude Code, Cursor, Claude Desktop and any other client without an SDK on your side. Recent releases show where it is heading: v0.0.78 added browser_find, which searches the page's accessibility snapshot for text or a regular expression and returns only the matching nodes with a little context, cheaper than capturing the whole snapshot when you only need one element. v0.0.80 added browser_start_recording and browser_stop_recording behind --caps=devtools to capture actions a person performs manually. The README was rolled to Playwright 1.63.0-alpha-2026-08-31 on 31 August 2026 and the repository stands at 579 commits. It is listed on AI Agents Listing as Playwright MCP, with a community package at mcp-playwright. Pick it when you want a browser tool inside an MCP-native agent and do not need hosted infrastructure.

Browserbase

Browserbase sells the browser itself. Its Browser-as-a-Service runs headless sessions your agent connects to, and the newer Search API and Fetch API let an agent find pages and pull them back as HTML, JSON or Markdown without a full session. The homepage claims more than 10,000 companies, 800,000 weekly SDK downloads and 100,000 developers, and the company hosts a Navigate 2026 conference in San Francisco on 10 September 2026. The most useful public number this week came from Anthropic's Fable 5.1 launch on 1 September 2026, where Browserbase reported the model completed 82% of tasks on its hardest browser-agent benchmark, against 74% for Opus 5 and 57% for Fable 5. Browserbase is infrastructure, not an agent: you still drive it with Stagehand, Browser Use or your own code. It is listed on AI Agents Listing as Browserbase, and its MCP server as mcp-server-browserbase.

Browserless

Browserless has run headless Chrome as a service for eight years, and its pitch to agent builders is that the model was never the bottleneck: the browser underneath gets blocked, loses the login and crashes three steps into the task. It answers with stealth fingerprints, built-in CAPTCHA solving, residential proxies and authenticated profiles that persist a login across runs. Existing Puppeteer or Playwright scripts move over by swapping launch() for connect() against wss://chrome.browserless.io. There is an MCP server that gives Claude, Cursor or any MCP client search, scrape and crawl tools, a hosted browser agent of its own, and a self-hosted option with the same API. The company cites 173 million Docker pulls, 99.9% measured uptime and more than 2,000 paying customers, and the free plan includes 1,000 runs a month with no card. It is listed on AI Agents Listing as Browserless.

Cloudflare Browser Run

Browser Run is the product Cloudflare used to call Browser Rendering, renamed and rebuilt for agents. It runs Chrome sessions on Cloudflare's network and, alongside Puppeteer and Playwright, now exposes the Chrome DevTools Protocol directly, so any CDP script or agent framework connects with a one-line change to the WebSocket URL and no Worker is required. The agent-specific additions are Live View (watch what the agent sees in real time), Human in the Loop (the agent hands a login page or an edge case to a person and takes control back afterwards), session recordings for debugging, MCP client support through the chrome-devtools-mcp package for Claude Desktop, Cursor, Codex and OpenCode, and support for WebMCP, the Chrome API from Chromium 146 onward that lets a site declare tools like search_flights for agents to call instead of clicking through a UI. Limits went to 120 concurrent browsers from 30, and Quick Actions run at 10 requests a second. It is the natural choice if your agent already lives on Workers.

Firecrawl

Firecrawl approaches the problem from the data side. Its API searches the web, scrapes any URL into Markdown or JSON, and now interacts with a page after scraping it: the homepage example scrapes Amazon, then sends interact(scrape_id, prompt="Search for 'mechanical keyboard'") and "Click the first result". One command, npx -y firecrawl-cli@latest init --all --browser, installs its skills and MCP server into Cursor, Claude Code and Windsurf, and there is an agent-onboarding SKILL.md an agent can fetch to get its own API key. The company reports a P95 latency of 3.4 seconds, coverage of 96% of the web including JavaScript-heavy pages, and 176,500 GitHub stars on the open-source repository. Its own browser-agent comparison lists the free tier at 1,000 credits a month including five hours of browser use, paid plans from $16 a month, and the Browser Sandbox at 2 credits per browser minute. Firecrawl is listed on AI Agents Listing as Firecrawl; it is the right tool when the job is reading the web at scale rather than acting on it.

Bright Data Web Unlocker

Web Unlocker is Bright Data's one-call answer to blocked pages. You send a URL and a zone name to api.brightdata.com/request and get back clean HTML, JSON, Markdown or a screenshot; proxy rotation across more than 400 million residential IPs, browser fingerprinting, CAPTCHA solving and JavaScript rendering happen behind the call. Bright Data claims a 98% success rate and bills only for successful requests, with a free tier of 5,000 requests a month, a pay-as-you-go rate per 1,000 requests, and a Scale plan that includes 383,000 requests. It is not an agent and does not drive a session; it is what you put in front of an agent whose scraping step keeps failing. Bright Data is listed on AI Agents Listing as Bright Data.

Apify

Apify is a marketplace of 67,620 Actors, prebuilt scrapers and automations for specific sites, from a TikTok scraper with 250,000 users to Google Maps and LinkedIn tools, that an agent can call as tools. The Apify MCP server at mcp.apify.com lets an MCP client pick and run Actors, and a new Apify AI beta finds the right Actor from a description. Underneath sits Crawlee, Apify's open-source crawling library with 25,652 GitHub stars, and Actors run on Apify's infrastructure with proxies, scaling and monitoring handled. Developers who publish Actors get paid: Apify says it paid out $1.5 million last month. It reports 99.95% uptime and SOC 2, GDPR and CCPA compliance. Reach for it when someone has already written the scraper you need. It is listed on AI Agents Listing as Apify.

Claude in Chrome

Claude in Chrome is the one entry built for a person rather than a program. Anthropic's extension became generally available on every paid Claude plan on 26 August 2026, and Claude can now take actions in your browser autonomously instead of asking approval for each one; a safety classifier checks every action before it runs, and you can switch back to manual approval. It reads the page you are on, types, clicks, navigates and fills forms using your existing logins, which is how it reaches internal dashboards and vendor portals that have no connector. The prompt injection numbers are the reason it took a year to leave pilot: on Anthropic's current red-team evaluation, attacks that reached Opus 4.5 succeeded 17.6% of the time before safeguards, while with probes plus the automatic-approval classifier no attack succeeded against Sonnet 5 or Opus 5 and 0.3% succeeded against Fable 5. It runs only in Chrome, not in other Chromium browsers or on mobile, and Enterprise admins can restrict it to approved domains. There is a Claude Code integration too: build in the terminal, verify in the browser, with Claude reading console errors and DOM state.

How to choose

Start from what you already have. If you have a coding agent such as Claude Code or Cursor and want it to test a web app, agent-browser or Playwright MCP adds a browser without adding a framework. If you are building an autonomous agent in Python and want to own the loop, Browser Use is the framework and Browserbase or Browserless is the browser under it. If you write TypeScript and want deterministic steps where possible, Stagehand on Browserbase. If your stack is Cloudflare Workers, Browser Run. If the task is reading rather than acting, Firecrawl, Web Unlocker or an Apify Actor will be cheaper and faster than a browser session. And if the task is yours, in your own accounts, Claude in Chrome is the only tool here that needs no code.

Two costs are easy to miss. Token cost scales with how much page content reaches the model, which is why agent-browser's refs and Playwright MCP's browser_find exist. Failure cost scales with blocking, which is why the hosted providers spend their homepages on CAPTCHAs, fingerprints and persistent logins.

What's new (as of 2 September 2026)

• Claude in Chrome went generally available on 26 August 2026 on all paid plans, with autonomous actions checked by a safety classifier. The same day Anthropic gave Claude Cowork a built-in browser.
• Fable 5.1, released 1 September 2026, scored 82% on Browserbase's hardest browser-agent benchmark against 74% for Opus 5, per Browserbase's statement in Anthropic's launch post.
• Playwright MCP rolled to Playwright 1.63.0-alpha-2026-08-31 on 31 August 2026; v0.0.78 added browser_find and v0.0.80 added recording tools.
• Cloudflare renamed Browser Rendering to Browser Run, added a CDP endpoint, Live View, Human in the Loop, WebMCP support and 120 concurrent browsers.
• Browser Use added an MCP registry server.json on 16 July 2026 and a one-prompt skill install for Claude Code, Codex and Cursor.
• Firecrawl marked its Interact endpoint as new and now ships an agent-onboarding SKILL.md.
• Browserbase holds its Navigate 2026 conference in San Francisco on 10 September 2026.

Key takeaways

• Browser agents split into frameworks (Browser Use, Stagehand), control layers for coding agents (agent-browser, Playwright MCP) and hosted browsers (Browserbase, Browserless, Cloudflare Browser Run); most builds need one of each.
• Firecrawl, Bright Data Web Unlocker and Apify return page data without a browser session and are the cheaper path when the job is reading.
• Claude in Chrome became generally available on 26 August 2026 and is the only entry here that needs no code.
• Free tiers exist at Browserless (1,000 runs a month), Firecrawl (1,000 credits a month) and Bright Data (5,000 requests a month).
• Prompt injection is the open risk: Anthropic reports 0.3% attack success against Fable 5 with its full safeguards, and Google reports gains on Gray Swan for Gemini 3.8.

FAQ

What is a browser agent?

A browser agent is an AI agent that completes tasks by operating a web browser: opening pages, reading their content, clicking, typing, filling forms and downloading files. It combines a model that decides each step with a browser-control layer that executes it, and often a hosted browser that runs the session.

Do I need a hosted browser to run a browser agent?

Not to start. Browser Use, Stagehand, agent-browser and Playwright MCP all run against a local Chrome. Hosted browsers from Browserbase, Browserless and Cloudflare matter once you need concurrency, persistent logins, proxies and CAPTCHA handling, which is where Browserless says most demos die in production.

Which browser agent tools are free?

Browser Use (MIT), Stagehand, agent-browser and Playwright MCP are open source. Browserless includes 1,000 runs a month free, Firecrawl 1,000 credits a month, and Bright Data Web Unlocker 5,000 requests a month. Claude in Chrome is included in paid Claude plans.

Are browser agents safe from prompt injection?

Not fully. A page can contain hidden instructions that redirect an agent, and Anthropic's own testing found 17.6% of red-team attacks succeeded against Opus 4.5 before safeguards. Its current stack of trained models, probes and an action classifier brings that to 0% for Sonnet 5 and Opus 5 and 0.3% for Fable 5. Treat every page an agent reads as untrusted input.

Every tool above with a listing sits in the Browser automation category on AI Agents Listing.