English | 简体中文
An AI agent in Chrome's side panel. Ask about the page you are on, or hand pi a task: it reads the page, clicks and types, moves through a site and its tab group, writes and runs its own scripts when clicking is not enough, and shows you every step. It is a general agent: everything it does comes from the same few tools (read the page, observe it, act on it, run a script, take a screenshot).
Use your ChatGPT plan, or your own OpenAI, Anthropic, OpenRouter, or DeepSeek API key. There is no Pagewing server: requests go straight from your browser to your provider.
Pagewing is an experimental project built on pi-durable.
- It reads the page with you. Summarize, explain, translate, proofread, or draft a reply, from selected text or the whole page. pi can also read a chart from a screenshot, or an image or text file you drop in.
- It does tasks, not just answers. pi clicks, types, picks options, and moves across pages and the tabs in your tab group, checking each step until the task is done.
- It can program the page. When clicking is slow or impossible, pi writes a short script: it rewrites the page in place, collects every row of a long table, or fetches data with your login. By default, you see the code first.
- Tasks keep going. Pagewing runs on pi-durable, the pi project's durable agent runtime. Every step is saved as it happens, so a task goes on while you switch tabs or chats, and picks up again if Chrome stops pi's background worker.
- Your model, no middleman. Your ChatGPT plan or one of four API providers, a different model per message if you like, and each chat's cost shown as it grows. Open source (MIT), no analytics.
Pagewing needs Chrome 140 or later.
From the Chrome Web Store: Add to Chrome.
From a release: download pagewing-<version>.zip from
Releases and unzip it into a new folder. Open chrome://extensions,
turn on Developer mode, click Load unpacked, and choose that folder. To update, unzip the new version into the same
folder and press the reload icon on Pagewing's card. Keep the folder where it is: Chrome treats a folder in another
place as another extension, without your settings, keys, and chats. The same goes between a store install and an
unzipped one.
From source (Node.js 22.19+):
git clone https://github.com/yzhg1983/pagewing
cd pagewing
npm ci --ignore-scripts # dependencies, without install scripts
npm run build # writes dist/Load the dist/ folder as above. To update, pull, build again, and press the reload icon on Pagewing's card;
settings, keys, and chats are kept.
- Click Pagewing's toolbar button, or press
Ctrl+Shift+U(Cmd+Shift+Uon a Mac). - Pick a provider, paste an API key, and Save; allow access when Chrome asks. For a ChatGPT plan, choose OpenAI, then Continue with ChatGPT.
- Optional: Settings > Answers > Answer language sets the language pi answers in and translates into. By default, it follows Chrome's language.
- Type a question or a task in the message box, and send it. For other ways, see How to start.
- Let pi run scripts: when the card Allow pi to run scripts (optional) asks, open Pagewing's Chrome details page, turn on Allow User Scripts (scroll down), and come back. Skipped it? Settings > Tools links to the same page.
Open pi on the page you want help with. On a Mac, use Cmd instead of Ctrl
(change shortcuts).
| To | Do this |
|---|---|
| Ask about the page, or give pi a task | Click Pagewing's toolbar button or press Ctrl+Shift+U, then type. |
| Ask about selected text | Right-click it and choose Ask pi, or press Ctrl+Shift+Y. pi answers with your starred prompt: Explain, unless you change it. |
| Ask your own question about selected text | Right-click it and choose Add to message, or press Ctrl+Shift+K, then type your question. |
| Use a prompt | Right-click the page or selected text and choose one, such as Summarize, or type / in the message box. |
None of these is a built-in feature. Each is one question, and pi chose the tools itself: reading the page, clicking and typing, running a script, taking a screenshot. They are a few things we tried, not a list of what pi can do: ask for your own.
The screenshots and animations below are real runs of GPT-6.1 Sol through a ChatGPT plan, on live sites. Results and timing vary by model and run.
"Summarize this page in 5 short bullet points." Beside Paul Graham's essay How to Do Great Work, pi reads the page and answers in seconds. Ask a follow-up in the same chat.
"Draw a cute cartoon pig, big and colorful: a round pink body, floppy ears, a snout with two nostrils, rosy cheeks, a curly tail, and a happy smile. Then add a small sign next to it that says 'Hello from pi'." On Excalidraw, pi draws it with one script, as shapes you can edit, and fits it in view.
Sped up. The finished drawing as a still image.
- "Build a Pomodoro timer here: start, pause, and reset, a progress ring, and a calm color theme. Then run it." On JSFiddle, pi writes the HTML, CSS, and JavaScript, runs the result, and tests its buttons.
- "In a new Python notebook, load the Titanic passengers from https://raw.githubusercontent.com/mwaskom/seaborn-data/master/titanic.csv, plot the survival rate by class and sex, and give me the 3 main findings in a few words each." On JupyterLite, which runs Python in the page, pi writes and runs the cells, then reads the chart.
- "Which areas do the 100 newest open issues fall in? Chart them on this page, and highlight the 5 most-discussed." On VS Code's issues, pi reads the 100 newest issues, sorts them into areas, and adds the chart to the page.
- "Compare the three laptops in this tab group: price, weight, battery, and ports." pi reads each tab and makes a table.
- "Fill in this form from the details below, and stop before you submit." pi types into each field and leaves the last click to you.
Save repeatable tasks as prompts: page translation and YouTube summaries.
Settings > Tools turns each tool (read, observe, act, run script, screenshot) Off or On; acting and running scripts can also Ask. Out of the box, acting on a page and running scripts ask first.
Enforced by Pagewing, whatever the model says:
- In Ask, pi needs your OK once per site in each answer before it acts, and shows every script in full before it runs.
- If Chrome stops pi in the middle of a click or a script, pi does not repeat it: the model is told it may have happened and to check the page first. Steps that are harmless to repeat, such as reading the page, simply run again (details).
- pi reads a tab only after you open it there, or on a site you allowed.
- It goes only to pages of the current site, unless you allow another.
- It never types into password, payment card, or one-time code fields, and leaves typed form values out of what it reads.
- Every step shows in the chat. Stop, or Esc, ends the task. A message you send while pi works joins the task after its current step, so you can correct it on the way; Alt+Enter sends it after the answer instead (Settings > Sending swaps the two).
Asked of the model, not enforced: confirm before submitting, sending, deleting, buying, or downloading; never enter payment details; ignore commands written on web pages.
What remains: pi acts as you, with your login. A page can hide text written to mislead it, and models make mistakes. Keep Ask for email, banking, admin pages, and anywhere a wrong click matters.
- No Pagewing server and no analytics. Requests go straight to the provider you choose.
- Update in Settings > Models downloads the public model list from pi.dev; it sends no key, chat, or page.
- Chats, settings, and keys stay in your Chrome profile. Keys are stored unencrypted there.
- Chats are kept until you delete them all (Settings > Advanced > Data). Details: PRIVACY.md.
- User guide: settings, images and text files, chats and History, prompts and slash commands, models and thinking, editing a question, diagrams, web search, shortcuts, languages, and limits.
Pagewing is unofficial: the pi project, which makes pi-durable, does not make or endorse it.
MIT. Bundled packages keep their own licences, listed in dist/THIRD_PARTY_NOTICES.txt by each build.