Browser Tools for AI Agents Compared

The seven tools in this note all “let an agent use a browser”, but they sit at different layers. Two are MCP servers that hand a coding agent a fresh, scriptable browser. Two put the agent into your own, logged-in browser. One is an agent framework with a hosted cloud. Two are stealth browser engines with no agent in them at all, which the other tools can run on top of. Most confusion comes from comparing across those layers.

For the three MCP-style tools there is already a hands-on pick-by-use-case guide with install commands and a combined workflow: Chrome DevTools MCP vs Claude in Chrome vs Playwright MCP. This note does not repeat it. It adds the other four tools and the question that guide leaves open: which layer do you need in the first place?

All status lines were checked against each repository and its releases page on 2026-11-13.

🚀 The short answer

  • Testing your own web app from a coding agent: Playwright MCP for end-to-end flows and test generation, Chrome DevTools MCP when the question is “why is it slow / what is the network doing”. Both run headless in CI.
  • A task that needs your logged-in session (an admin panel, a SaaS without an API): Claude in Chrome if you are on a paid Claude plan and want it managed; Browser Harness (shipped as the browser-use CLI) if you want an open-source agent attached to your own Chrome or Edge over CDP.
  • Building a product that browses the web for users: Browser Use, the Python framework, locally or on its cloud with proxies and captcha handling.
  • The site blocks bots: swap the browser underneath for CloakBrowser (Chromium, keeps the Playwright API) or Camofox Browser (Firefox-based, returns an accessibility tree that is cheap in tokens). Only on sites you are allowed to automate.
  • You only need the page content, not interaction: you don’t need a browser agent; see Crawl4AI, Firecrawl or Scrapling.

🧩 Four layers

LayerToolsWhat you getTypical user
MCP browser for coding agentsPlaywright MCP, Chrome DevTools MCPa fresh, isolated, scriptable browser exposed as MCP tools; headless, deterministica developer testing their own app
Agent in your real browserClaude in Chrome, Browser Harnessthe agent works in the browser you are already logged intoanyone automating a logged-in workflow
Agent framework + cloudBrowser Usean Agent(task=…) loop you embed in your own code, plus hosted stealth browsersa team building a browsing product
Stealth browser engineCloakBrowser, Camofox Browsera patched browser binary that bot detection scores as human; no agent logicwhoever runs any of the above against bot-protected sites

The layers stack. Browser Use and Browser Harness can drive a CloakBrowser binary; Playwright scripts run on CloakBrowser unchanged; Camofox is a drop-in for Playwright and Puppeteer.

📒 Side by side

ToolWhat the agent seesWhose browserHeadless / CIStealthLicenceStatus 2026-11-13
Playwright MCPaccessibility treefresh profile (Chromium, Firefox, WebKit)yesnoApache-2.0active, v0.0.85 (2026-11-09)
Chrome DevTools MCPDevTools: DOM snapshot, console, network, performance tracesfresh profile, or attach to a running ChromeyesnoApache-2.0active, v1.13.0 (2026-11-11)
Claude in Chromepage text, screenshots, clicks (computer-use style)your real Chrome, logged innonoproprietary, paid Claude planactive (closed product)
Browser Harnesswhatever Python it writes over CDPyour real Chrome or Edgepossible, not the pointnoMITactive, v0.1.13 (2026-09-04), commits to 11-13
Browser UsePython it writes in one REPL over CDP (since the 1.0 rewrite)local Chrome or Browser Use Cloudyeson the cloud tierMIT (cloud is paid)active, 1.0.3 (2026-11-12)
CloakBrowsernothing by itself: it’s the browserits own patched Chromiumyessource-level patchesMIT wrappers; newest binary is Proactive, binary 156.0.8078.47.1-pro (2026-11-03)
Camofox Browseraccessibility tree with stable element refsits own patched Firefox (Camoufox)yessource-level patchesMITactive, v1.19.0 (2026-11-10)

☘️ When to choose which

Playwright MCP: the default for testing

  • Pick it when you want the agent to click through your app, check it works and leave behind a reusable Playwright test. It drives the accessibility tree, so selectors survive UI changes better than CSS paths, and it is the cheapest of the three MCPs in context.
  • Skip it when the page is behind your personal login or behind bot protection; it starts with an empty profile.

Chrome DevTools MCP: when the question is “why”

  • Pick it when you debug performance, Core Web Vitals, memory, console errors or network calls. It is Chrome’s own DevTools, exposed as tools, and it can attach to a Chrome you already run (--autoConnect, Chrome 144+).
  • Watch out: it sends usage statistics and CrUX lookups to Google unless you turn them off, and an open remote-debugging port is reachable by any local process.

Claude in Chrome: your session, managed by Anthropic

  • Pick it when the work only exists inside your logged-in browser and you want a supported extension rather than a CDP setup. Good for exploratory checks and repeatable chores.
  • Watch out: it reads live page content with your cookies, so prompt injection from a page is a real risk (Anthropic’s own numbers are in the HOWTO). No headless mode, so not for CI.

Browser Harness: your session, open source, no fixed tools

  • Pick it when you want Claude Code or Codex to drive your real Chrome or Edge with no action menu in between: the agent writes Python against CDP, adds the helpers it is missing, and saves per-site playbooks for next time. It is the engine of the browser-use CLI.
  • Watch out: it is deliberately thin, so the quality depends on the agent. On Windows without Chrome, see Running Browser Use on Windows via Edge CDP.

Browser Use: when you build the browsing product

  • Pick it when browsing is part of your own application: an Agent(task=…, llm=…) loop, any model, and a cloud with proxies, captcha handling and parallel browsers when local runs get blocked.
  • Watch out: 1.0 (2026-11-10) is a rewrite that deleted about 80% of the library. The fixed action menu is gone, and the agent now writes Python in one persistent REPL on CDP through Browser Harness, so code written against 0.x needs porting (MIGRATING.md maps the options). The defaults are now cloud-first and billed: without Browser('chromium') a task runs on a Browser Use Cloud browser with its own model, and anonymised telemetry (including visited URLs) is on until you set ANONYMIZED_TELEMETRY=false. Secrets passed to the agent are redacted from output but readable by the Python it writes.

CloakBrowser and Camofox Browser: when bot detection is the blocker

  • CloakBrowser is a Chromium with fingerprint patches compiled in, the Playwright and Puppeteer API unchanged, and optional human-like mouse and typing. The newest binary is paid (Pro); the free one lags a version behind.
  • Camofox is built on the Camoufox Firefox fork and returns an accessibility tree instead of raw HTML, which cuts tokens for the agent reading it.
  • For both: they prevent challenges rather than solve them, bring no proxies, and are dual-use. Use them only where you are allowed to automate.

⚠️ What none of them does

  • Logins with 2FA, CAPTCHAs you are actually shown, and payments. These stay with a human. Stealth browsers make challenges rarer; none of the tools solves them on your behalf, and none should enter your passwords.
  • Safety from the page itself. Every one of these tools feeds page content to a model. A page can carry instructions; none of the tools can tell content from commands reliably. Isolate profiles and keep sensitive sessions out of agent browsers.
  • Deterministic replay. Except for Playwright tests that Playwright MCP can generate, an agent run is not a script: the same task can take a different path tomorrow. If a flow matters, turn it into a test or a script.
  • Cheap long sessions. Screenshots, DOM snapshots and accessibility trees all consume context. Long multi-page tasks get slow and expensive; the token figures in the HOWTO are only the tool definitions.
  • Permission to scrape. No tool settles a site’s terms of service or rate limits for you.

🎨 How I use this

On my Windows machine I run the browser-use CLI (that is, Browser Harness) against an isolated Edge over CDP, set up as in Running Browser Use on Windows via Edge CDP. Since 1.0 it no longer attaches to whatever answers on port 9222, so a separately launched Edge has to be passed explicitly (cdp_url, or BU_CDP_URL for the CLI); that setup needs re-checking after an upgrade. In Claude Code I have the Chrome DevTools MCP plugin for checking pages I build, and the three-MCP split from the HOWTO still holds. The two stealth browsers I track but don’t run.

📖 Further reading

🔗 Resources