Browser Automation for AI Agents (2026)
I compare five ways to give an AI agent a browser, from local Chrome over MCP to hosted cloud browsers, with real install commands and pricing.
By agentools
- Browser Use: Python agents, local or cloud
- Chrome DevTools MCP: Debugging and testing your own site
- Shuttle: Agents with persistent logins you can watch
- Scrapling: Scraping and form workflows on hard sites
- Kernel: Hosted cloud browsers at scale
For browser automation for AI agents, I use Chrome DevTools MCP when a coding agent needs to debug a local Chrome. Shuttle fits when agents need persistent logins in a browser I can watch. Browser Use suits Python agents that may later move to the cloud, Scrapling suits scraping, and Kernel suits hosted browsers at scale.
APIs only cover part of the work a solo founder hands to an agent. Checking a dashboard, filling a form or testing a signup flow often only happens in a browser. Giving the agent a browser closes that gap, but the tools fall into two families that solve different problems.
The first family drives a real browser on your own machine. Chrome DevTools MCP, Shuttle and the Browser Use library are in this group. You get your own network and your own logins, and you pay no hosting bill. The second family is hosted cloud browsers and scraping. Kernel and Browser Use Cloud run the browser on their side and add stealth, proxies, CAPTCHA handling and replays, and you pay per hour or per plan. Scrapling sits in between: a scraping framework that runs locally.
At a glance
| Tool | Best for | Pricing | Type |
|---|---|---|---|
 Browser Use | Python agents, local or cloud | Open source | Library + cloud API |
 Chrome DevTools MCP | Debugging and testing your own site | Open source | MCP Server |
 Shuttle | Agents with persistent logins you can watch | Free and open source | Browser with built-in MCP server |
 Scrapling | Scraping and form workflows on hard sites | Free, open source | Python library + CLI + MCP server |
 Kernel | Hosted cloud browsers at scale | Free tier | Browser API |
The 5 picks
1. Browser Use
Browser agents and cloud browsers for AI
Browser Use is an MIT-licensed Python library that lets an LLM drive a real browser. Browser Use Cloud adds an Agents API (you give it a task and get a result back) and managed cloud browsers you can reach over SDK, REST or CDP. Cloud browsers add residential proxies, stealth, CAPTCHA solving and a live view URL.
It needs Python 3.11 or newer. For a new project:
uv init --python 3.12
uv add browser-use
Put a model key such as OPENAI_API_KEY in .env. To run the browser in their cloud, add BROWSER_USE_API_KEY and use Browser(use_cloud=True).
In Claude Code, the hosted MCP server takes one command:
claude mcp add -t http -H "Authorization: Bearer YOUR_API_KEY" browser-use https://api.browser-use.com/v3/mcp
Pricing: the library is free. Cloud browsers cost $0.02 per browser-hour plus network traffic. Cloud agents cost the model price plus 20%. You start with $1 of free credit, and the minimum top-up is $5.
Caveat: the 20% agent markup is added to your model bill, so long tasks get expensive. If a local browser is enough, use the library on its own.
Pricing: Open sourceLibrary + cloud API
2. Chrome DevTools MCP
Official Chrome DevTools for coding agents
Chrome DevTools MCP is Google's official MCP server, licensed Apache-2.0. It lets an agent control and inspect a live Chrome: performance traces, network, console, screenshots, Puppeteer-backed automation and Lighthouse audits. Chrome only starts when a tool needs it. There is also a chrome-devtools CLI if you don't want MCP.
You need Node.js LTS and Chrome stable. In Claude Code:
claude mcp add chrome-devtools --scope user npx chrome-devtools-mcp@latest
You can also install it as a plugin from inside Claude Code:
/plugin marketplace add ChromeDevTools/chrome-devtools-mcp
/plugin install chrome-devtools-mcp@chrome-devtools-plugins
A good first prompt to test it: "Check the performance of https://developers.chrome.com".
Pricing: free and open source.
Caveat: it collects optional usage statistics unless you opt out. I find it most useful for inspecting pages I build, less so for running errands across many sites.
Pricing: Open sourceMCP Server
3. Shuttle
Chromium fork for AI agents: built-in MCP server, and each agent gets its own logins and tab group
Shuttle is a BSD-3-Clause deep fork of Chromium built for agents. The MCP server runs inside the browser at http://127.0.0.1:9222/mcp, so there is no separate server to start. Each agent gets a named identity with its own persistent, isolated cookies and logins. Agent sessions show up as colored tab groups in your normal window, so you can watch the agent work or take a tab back. Each session also has its own CDP endpoint for Playwright or agent-browser.
It comes with 24 page tools, including snapshot, click, fill, screenshot and get-text. You get no remote debugging prompts and no automation infobar, and Google services, sync and telemetry are removed. It runs on macOS, Linux and WSL2. Tobias Strebitzer built it.
Download it from shuttlebrowser.com/download, start it, and add the endpoint in Claude Code:
claude mcp add -t http shuttle http://127.0.0.1:9222/mcp
Pricing: free and open source.
Caveat: it came out in October 2026 and has about 40 GitHub stars. Expect rough edges and a small community.
Pricing: Free and open sourceBrowser with built-in MCP server
4. Scrapling
Python scraping framework with a stealth browser and an MCP server for agents
Scrapling is a BSD-3-Clause Python scraping framework by Karim Shoair, with about 88k GitHub stars. Version 0.5, released October 17, rebuilt the MCP server around full browser workflows. An agent can fill forms, click, scroll, type and wait in one ordered sequence. It can also read or search the current page without reloading, run JavaScript, take screenshots, inspect recorded network requests, and save and restore login state.
It solves hCaptcha, reCAPTCHA and Cloudflare Turnstile. Pages are stripped of scripts and hidden prompt-injection text before they reach the model. The adaptive parser finds elements again when a layout changes. You also get a spider framework, Response.markdown() and a site-to-Markdown crawler.
pip install "scrapling[ai]"
scrapling install
claude mcp add ScraplingServer "/path/to/.venv/bin/scrapling-mcp"
scrapling install adds the browser fetchers. Change the path to point at your own virtual environment.
Pricing: free and open source.
Caveat: v0.5 has many breaking changes, so pin your version if you have older scripts. Respect each site's terms.
Pricing: Free, open sourcePython library + CLI + MCP server
5. Kernel
Fast cloud browsers for web agents
Kernel is a browser sandbox that hosts cloud browsers for web agents. The browsing happens on Kernel's side, and your agent connects over CDP from Playwright or Puppeteer. You get live view, session replays, managed stealth mode, managed auth and browser pools, and you are not charged while a browser sits idle on standby. A hosted remote MCP server (OAuth 2.1 or API key) can launch Chromium browsers and run Playwright code.
Install the CLI with either command:
brew install kernel/tap/kernel
npm install -g @onkernel/cli
Then set up MCP for your client:
kernel mcp install --target <target>
Pricing: the Developer plan is free plus usage, with $5 per month in credits. Hobbyist costs $30 per month and Start-Up costs $200 per month.
Caveat: the service is proprietary, though the browser images are open source.
Pricing: Free tierBrowser API
More to consider
Comparison table
| Tool | Type | Runs where | Pricing | Free tier | MCP support | Open source |
|---|---|---|---|---|---|---|
| Browser Use | Python library plus cloud | Local or cloud | $0.02 per browser-hour plus traffic; agents at model cost + 20% | $1 credit, library free | Hosted MCP | Library: MIT |
| Chrome DevTools MCP | MCP server for Chrome | Local Chrome | Free | Yes | Yes | Apache-2.0 |
| Shuttle | Chromium fork for agents | Local (macOS, Linux, WSL2) | Free | Yes | Built into browser | BSD-3-Clause |
| Scrapling | Scraping framework | Local Python | Free | Yes | Yes | BSD-3-Clause |
| Kernel | Hosted browser sandbox | Cloud | Developer free plus usage; $30 or $200 per month plans | $5/month credits | Hosted remote MCP | No (browser images are) |
How I would choose
- Debugging my own site or app: Chrome DevTools MCP. Performance traces, console, network and Lighthouse cover most of what I check by hand.
- An agent that stays logged in to my accounts while I watch: Shuttle. Each agent gets isolated cookies, and its tabs sit in my normal window.
- A Python agent that finishes tasks end to end: Browser Use. I start with the free library and switch to the cloud when I need proxies, stealth or CAPTCHA solving.
- Pulling data from sites that change layout or show CAPTCHAs: Scrapling. The adaptive parser and the prompt-injection cleanup are why I pick it.
- Running many browsers away from my laptop: Kernel, for browser pools, session replays and managed auth.
- A light skill instead of an MCP server: agent-browser. It is Vercel's CLI, and it drives Chrome through accessibility snapshots to open pages, fill forms, click, take screenshots and do exploratory QA. Install it with
npx skills add vercel-labs/agent-browser. It also needs a global npm install and a Chrome download. It can connect to a Shuttle session through that session's CDP endpoint.
FAQ
What is the cheapest way to start with browser automation for AI agents?
Chrome DevTools MCP, Shuttle and Scrapling are free and open source. The Browser Use library is free too. With all four, you only pay for your model.
When do I need a cloud browser instead of a local one?
When the browser has to run off your machine, or when you need residential proxies, stealth, CAPTCHA solving or browser pools. For those I look at Browser Use Cloud or Kernel.
Do these work with Claude Code?
Yes. Browser Use, Chrome DevTools MCP, Shuttle and Scrapling each have a claude mcp add command above. Kernel sets up its MCP server with kernel mcp install --target <target>.
Can an agent use my logged-in accounts?
Shuttle gives each agent a named identity with its own isolated cookies and logins. Scrapling can save and restore login state, and Kernel offers managed auth.
How do I reduce prompt injection from web pages?
Scrapling strips scripts and hidden prompt-injection text before a page reaches the model. With any tool, I keep agents on sites I trust and watch the session.