Pagehand

AI that puts a hand on the current page.

Pagehand opens in Chrome’s side panel and works on the page you are already looking at — reading it, filling it in, clicking through it. You ask in plain words; it does the part you’d rather not.

Add to ChromeSee what it does

Works on any page · Sign in and go · Your chats stay in your browser

What it does

Four things people ask it for.

Open the panel on any page and type. Every line below is something you could paste in as it stands.

  • Read this page for me

    Long article, dense report, a table you don’t want to retype — it reads what is actually on screen, not a search result about it.

    “Summarize this report and pull out the three numbers that matter.”

  • Do it for me

    Forms, checkboxes, multi-step flows. It works in the tab you already have open, so anything you are signed in to, it can reach.

    “Fill this form with the same details as last time, and stop before submitting.”

  • Compare what’s in my tabs

    Type @ to point it at other tabs you have open, and it reads them together instead of one at a time.

    “@ these three listings — which is cheapest once shipping is included?”

  • Tell me why this page is broken

    For the developers in the room: console errors and network requests, read and explained without you opening DevTools.

    “Clicking Save does nothing — what request fails, and what did it return?”

Install

Get it from the Chrome Web Store.

One click, and Chrome keeps it up to date. No zip to keep, no folder to remember, no Developer mode.

  1. Add to Chrome

    Open the Pagehand listing and click Add to Chrome.

  2. Pin it

    Click the puzzle icon in the toolbar and pin Pagehand, so the side panel is always one click away.

  3. Open the side panel

    Click the Pagehand icon on any page — or press ⌘/Ctrl+Shift+K. Sign in, and start asking.

Updates arrive on their own. Prefer building it yourself? See the GitHub setup.

Rather load it unpacked?

  1. Download

    Get pagehand.zip from the latest GitHub Release.

  2. Unzip

    Extract the zip to a folder you will keep — Chrome reads that folder every time it starts. Don’t delete it after installing.

  3. Open Extensions

    In Chrome’s address bar, go to chrome://extensions and press Enter.

  4. Turn on Developer mode

    Toggle Developer mode in the top-right corner of that page.

  5. Load unpacked

    Click Load unpacked, then choose the folder you unzipped (the one that contains manifest.json).

To update later: download the new zip, replace the folder contents, then click the reload icon on chrome://extensions.

Model access

Sign in, or bring your own key.

Two ways to reach a model. The hands are the same either way — the agent loop and every CDP tool run inside your browser.

  • Hosted · default

    Sign in and go

    One email, no API key, no provider account. Pagehand routes the model call through its own endpoint and carries the bill.

    • Nothing to configure on the first run
    • Chat threads still stay in your browser
    • In private preview while usage limits are built

  • Your key · advanced

    Bring your own key

    Point the panel at DeepSeek, OpenAI, Anthropic, or any OpenAI-compatible endpoint. Requests go straight from your browser to that provider.

    • Key stays in chrome.storage.local
    • Nothing touches Pagehand’s servers
    • Any model your provider offers, at its price

Hosted is in private preview: sign-in works, but requests are served to a short list of accounts until quotas land. Email us for access → Your own key works for everyone today, no account needed.

Privacy

It reads the page. Here is the whole of it.

Pagehand looks at a tab when you ask it something, and only that tab. What it reads goes to the AI model answering you and nowhere else — we never store what a page said, and your chat history stays in your browser. While it is working, Chrome puts a “Pagehand is debugging this browser” bar across the top of the window: that is Chrome telling you the truth about what an extension is doing, we are not allowed to hide it, and it goes away when you close the panel. Read the full privacy policy →

Under the hood

Same protocol Puppeteer speaks.

No Node sidecar, no MCP bridge, no DOM scraping. The side panel holds the agent loop and speaks Chrome DevTools Protocol directly, through chrome.debugger on Manifest V3.

  • take_snapshot

    Read the page as a structured accessibility tree with stable uids.

  • click / fill / type

    Drive forms and UI the way a user would — including select-all before replace.

  • navigate / pages

    Move between pages and manage them while the agent keeps context.

  • console / network

    Inspect errors and requests without opening DevTools yourself.

  • evaluate_script

    Run focused page scripts when reading the tree is not enough.

  • take_screenshot

    Capture what the page looks like and keep it in the chat thread.

The full tool list, and the code behind it, are on GitHub.