Blog

Browser Automation for AI Agents: Chrome vs Playwriter vs Agent-Browser

A practical comparison of the ChatGPT Chrome Extension, Playwriter, and Vercel agent-browser for signed-in browser work, automated scripts, and AI-assisted tasks.

On this page

I reran this comparison on July 27, 2026. The original article included Claude in Chrome, but I can’t test Claude now. This refresh compares the current ChatGPT Chrome Extension, Playwriter, and Vercel’s agent-browser on a real Wikipedia search.

The direct recommendation:

  • Use the ChatGPT Chrome Extension for tasks inside a conversation that need the tabs and sign-ins already in Chrome.
  • Use agent-browser for repeatable scripts that must run on their own in a separate browser.
  • Use Playwriter when an agent needs deeper control over the Chrome session you are already using.

Decision table

DecisionChatGPT Chrome ExtensionPlaywriteragent-browser
Time for the full task1.95s9.96s; 7.61s was waiting for Wikipedia to load3.49s; 2.37s was starting its browser and opening Wikipedia
Did it finish?YesYesYes
Signed-in sitesUses your existing Chrome tabs and sign-insUses your Chrome by default; it can also run in a separate or cloud browserUses a separate browser by default; reusing a Chrome profile or saved sign-in is optional
Scripts that run on their ownPoor fit: built for conversation, not scripts that run by themselvesGood when a script needs detailed control over Chrome; requires an extension and a connection to ChromeBest fit: designed for scripts, separate browsers, and running several steps together
How the agent uses itDirectly inside ChatGPT or CodexBy writing and running code that controls ChromeBy sending browser commands from a script or terminal
What it reads and capturesPage text and controls, plus a screenshotPage text and controls, plus a screenshot that includes the Playwriter toolbarPage text and controls, plus a screenshot; it can also add numbered labels

The tools measured time differently. Chrome and Playwriter timed their own actions, while I timed agent-browser from the command line. Playwriter spent most of its run waiting for Wikipedia to load. agent-browser spent most of its run starting its browser and opening the page.

What I tested

Each tool performed the same task on the English Wikipedia Main Page without changing the site:

  1. Open Wikipedia and confirm its URL and title.
  2. Find the search box.
  3. Search for Browser automation.
  4. Confirm that Wikipedia opens its Headless browser article.
  5. Take a screenshot of the visible page.
  6. Ask the tool to find a button that does not exist, without clicking anything else.

Playwriter used a narrower page layout where Search appeared first as a link, so it needed one extra click to reveal the form. All three tools still reached the same article.

For the missing button, Chrome and Playwriter reported that it was not there without clicking anything. agent-browser waited one second, then returned its usual element not found message.

Versions tested

The test ran on an Apple Silicon Mac with macOS 27.0.

Playwriter and agent-browser ran from the command line. The Chrome Extension ran inside ChatGPT or Codex.

ToolTested versionBrowser in the runHow I ran it
ChatGPT Chrome Extensionbundled plugin build 26.721.41059Chrome 150.0.0.0@Chrome, then chrome.tabs.new(), tab.goto(), tab.playwright.domSnapshot(), locators, and tab.screenshot()
Playwriter0.4.0Chrome 150.0.0.0$HOME/.bun/bin/playwriter, session 8, snapshot(), Playwright controls, and screenshot tools
agent-browser0.33.0Chrome for Testing 148.0.7778.97$HOME/.bun/bin/agent-browser, session portfolio-comparison-20260727-wikipedia, limited to Wikipedia, with browser controls and screenshots

The Chrome test used the bundled extension rather than a terminal command. OpenAI’s Chrome Extension documentation explains site permissions, access to existing Chrome tabs, and how to call @Chrome. Playwriter’s official product documentation covers its extension, command line, detailed Chrome controls, and option to use a separate browser. The agent-browser repository and documentation document its command line, running several steps together, numbered screenshots, separate sessions, and ways to reuse sign-ins.

Results

StepChatGPT Chrome ExtensionPlaywriteragent-browser
Open Wikipedia706ms7.61s2.37s
Find or open search113ms323ms130ms
Enter the search61ms25ms200ms
Search and confirm1.01s1.88s300ms
Take the screenshot67ms121ms490ms
Full task1.95s9.96s3.49s
Missing button check57ms, zero matches6ms, zero matches1.47s, element not found error

All three opened Wikipedia, used its search box, opened the Headless browser article, confirmed the heading, and produced a readable screenshot.

ChatGPT Chrome Extension

The extension showed the search box and final heading in a compact text outline of the page:

- search:
  - searchbox "Search Wikipedia"
  - button "Search"
- main:
  - heading "Headless browser" [level=1]

Its screenshot showed the desktop Wikipedia page without extra labels or tool controls. This takes the least setup when the task begins inside ChatGPT or Codex and needs tabs or sign-ins already in Chrome.

It is designed for browser tasks inside a conversation, so it is a poor fit for automated tests or scripts that must run without a person.

Playwriter

Playwriter completed the task in 9.96s after opening a separate browser session. Wikipedia took 7.61s to load, and the narrower layout required one click to reveal the search form.

The screenshot included Playwriter’s toolbar. Its labeled-screenshot option also worked, and the text outline included the toolbar controls. The labels can help an agent understand where controls are, but the tool’s own interface becomes part of the page it sees.

Playwriter first found its local browser connection already in use. It restarted that connection and then completed the task without another problem.

agent-browser

agent-browser opened a separate browser with no saved sign-in. Starting that browser and opening Wikipedia took 2.37s; the full task took 3.49s.

Its text outline showed the search controls, and its screenshot showed Wikipedia without added tool controls. It can also add numbered labels that match items in the text outline.

When I asked it to find the missing button, it returned:

Element not found: #missing-action. Verify the selector, role, or name is correct and the element exists in the DOM.

The update needed one manual fix on this Mac. The Bun package installer skipped part of the setup, so agent-browser would not open. I made the installed file runnable, and its health check then found no problems.

Using your existing sign-ins

The ChatGPT Chrome Extension and Playwriter use your existing Chrome, so they are the natural choices when a task depends on an already signed-in tab. That is convenient, but it also gives the agent access to the tabs you use. Site permissions and a narrowly defined task matter.

agent-browser opens a separate browser and does not share your sign-ins by default. You can choose to reuse a Chrome profile or saved sign-in when a task needs one. For scripts that run on their own, I would keep the separate browser.

See guarded browser automation in a working system.

What I use now

I used to use Playwriter for every browser task. Now I almost always use the ChatGPT Chrome Extension instead. I call @Chrome from the conversation, which lets the agent work in the Chrome tabs and signed-in session I already have open.

Have a workflow that's getting harder to run?

I'll help you identify where it's breaking down and decide whether the next step is a process change, existing software, automation, or a focused build.