Show HN: PageAgent, A GUI agent that lives inside your web app | Hacker News Reader

Show HN: PageAgent, A GUI agent that lives inside your web app

alibaba.github.io

147 pointssimon_luv_pho76 comments

Title: Show HN: PageAgent, A GUI agent that lives inside your web app

Hi HN,

I'm building PageAgent, an open-source (MIT) library that embeds an AI agent directly into your frontend.

I built this because I believe there's a massive design space for deploying general agents natively inside the web apps we already use, rather than treating the web merely as a dumb target for isolated bots.

Currently, most AI agents operate from external clients or server-side programs, effectively leaving web development out of the AI ecosystem. I'm experimenting with an "inside-out" paradigm instead. By dropping the library into a page, you get a client-side agent that interacts natively with the live DOM tree and inherits the user's active session out of the box, which works perfectly for SPAs.

To handle cross-page tasks, I built an optional browser extension that acts as a "bridge". This allows the web-page agent to control the entire browser with explicit user authorization. Instead of a desktop app controlling your browser, your web app is empowered to act as a general agent that can navigate the broader web.

I'd love to start a conversation about the viability of this architecture, and what you all think about the future of in-app general agents. Happy to answer any questions!

Open articleView on HN

This is highly experimental right now, but here are some quick links for anyone wanting to dig deeper:

- GitHub: https://github.com/alibaba/page-agent

- Live Demo (No sign-up): https://alibaba.github.io/page-agent/ (you can drag the bookmarklet from here to try it on other sites)

- Browser Extension: https://chromewebstore.google.com/detail/page-agent-ext/akld...

I'd be really interested in feedback on the security model of client-side agents giving extension-bridge access, and taking questions on the implementation!

I tried setting the LLM to "http://0.0.0.0:8080" and the extension crashed and now continues to crash at startup.

Is http://0.0.0.0:8080 a OpenAI compatible API?

Even it’s not, it’s not supposed to crash on startup. Can you post some screenshots and details on GitHub issues? I’m looking into this.

Hah, wow. I hope this is an LLM response, otherwise, what a huge blindspot for a "developer"...

I don't get it. It's just docs. I don't see anything. Even the video in your GitHub readme doesn't work in my browser.

It sounds like a network issue or browser compatibility issue. Can you please add an issue on GitHub so I can look into this.

I mean, not even the readme video?

Your readme video is https://github.com/user-attachments/assets/11aed429-b69c-49d...

That gives me 404

Confirmed. Have to fix that asap. About other issues. Can you see the homepage? What’s the browser version you use?

https://imgur.com/a/eEQtWQp

I see the homepage but no chat or anything else that could be an agent.

Does the "run" button work?

No, it doesn't. I get this error in console

Uncaught (in promise) Error: WebGL2 is required but not available. setupGL https://alibaba.github.io/page-agent/assets/SimulatorMask-B8... K https://alibaba.github.io/page-agent/assets/SimulatorMask-B8... <anonymous> https://alibaba.github.io/page-agent/assets/SimulatorMask-B8... nt https://alibaba.github.io/page-agent/assets/SimulatorMask-B8... maskReady https://alibaba.github.io/page-agent/assets/PageAgent-oX13Jj...

Because I have WebGL disabled.

I see. The visual effect requires the browser to support webgl2.

The core functionality should not crash because the visual effect crashed. Not a good practice. I will fix that asap.

Thanks for noticing. Btw the video should work now.

Yes it works, nice effects.

on chrome, the extension should be limited to a defined tab group

Currently, the extension only has access to the active tab when the task starts (configurable) and any new tabs it opens — all automatically placed into a dedicated tab group. It won’t touch other existing tabs.

Are you looking for something like scoping the agent to a predefined tab group?

Scoping to a dedicated tab group is a smart approach for sandboxing. The activeTab permission model already limits access nicely, but combining it with tab groups adds a visible boundary that users can actually see and trust. Would be great if Chrome's extensions API gave more granular tab group controls natively.

The "inside-out" framing resonates with me. I have been building embeddable scripts that get dropped into third-party sites via a script tag, and the architectural decisions you are making here mirror a lot of the same trade-offs I have encountered.

The biggest challenge with any in-page tool is the tension between needing deep DOM access and maintaining isolation. For the agent UI itself, you almost certainly want iframe isolation -- CSS conflicts with the host page are a constant headache otherwise. But for the actual DOM interaction (reading page state, simulating events), you need to be in the host page context. This dual architecture (iframe for your UI, direct access for page interaction) adds complexity but is worth it for reliability across diverse sites.

One thing I would flag as a real production concern: Content Security Policy. A significant number of enterprise and SaaS sites set strict CSP headers that will block inline scripts, eval, and sometimes even dynamically created script elements. If your target audience includes embedding this in production apps, you will hit CSP issues quickly. The bookmarklet approach cleverly sidesteps this for demos, but for a proper integration the host app needs to explicitly whitelist your script origin.

The HTML dehydration approach you described in the comments (parsing live HTML, stripping to semantic essentials, indexing interactive elements) is smart. In my experience, the fidelity of that serialization step is where most of the edge cases live. Shadow DOM, canvas elements, dynamically loaded content, iframes-within-iframes -- each one needs special handling and you end up building a progressively more complex serializer over time. Keeping that layer thin and well-tested is probably the highest-leverage investment for long-term maintainability.

Really appreciate the in-depth feedback.

Iframe and CSP are big problems. For the in-page version, I chose to leave out Shadow DOM, canvas, and iframes. Although I know one of the developers forked a version to control same-origin iframes. I don't think it's practical to try to hack around browser security (and website security) — that's why I built the browser extension. I'm hoping the bridge that lets a page call the extension can cover most use cases.

My original HTML dehydration script was ported from `browser-use`. You're absolutely right that it's getting heavier over time, and it's the key factor influencing the overall task success rate. I'm looking to refactor that part and add an extension system for developers to patch their own sites. Hope it turns out well.

Thank you for the feedback. I'll be extra cautious to keep the dehydration code maintainable.

> Data processed via servers in Mainland China

Appreciate the transparency, but maybe you could add some European (preferably) alternatives ?

Please use your own LLM api instead!

The free testing LLM is Qwen hosted by Aliyun. Qwen and DeepSeek are the only ones I can afford to offer for free. It's just there to lower the try-out barrier; please DO NOT rely on it.

The library itself does NOT include any backend service. Your data only goes to the LLM api you configured.

I tested it on local Ollama models it works fine.

Or why not stay fully local with WebLLM... https://webllm.mlc.ai

That looks great! I also thought about calling the Gemini nano model embedded into Chrome (only extensions can do that). But after some testing on smaller models I found that anything smaller than 9b can’t really handle the complex tool call schema I use.

Qwen3.5 4b is quite good but still gives messy json quite often. But it’s very promising!

Maybe after one more model iteration or some fine-toning we can go fully embedded?

I'm looking into a European testing endpoint. The legal and compliance requirements are quite hassle, and persuading my company to pay for that infrastructure is gonna be a tough sell.

Very interesting. Is this related to CoPaw and AgentScope? I think the AG-UI integration for dynamic UI would be useful here, are you using that?

I'm building a web UI workspace right now where I have been planning to integrate the agent as an app or component instead of having it be the entire UI. I may fork PageAgent for that, lets see.

Currently the only dependency is zod for schema parsing.

I'm intentionally building on a lightweight, in-page JavaScript foundation to carve out some differentiation from the Python-heavy agent ecosystem.

The "protocol" layer of AG-UI does look interesting. I'll look into it to see if I can reuse something, although it seems to be evolving more toward an integration framework rather than an open protocol.

Really glad this resonates with your use case. Lightweight embedding is exactly my priority scenario. Would love to hear how the work goes!

Not exactly the same but I'd also point to Paul Kinlan's FolioLM as a very interesting project in this space. A very nice browser extension,

> Collect and query content from tabs, bookmarks, and history - your AI research companion. FolioLM helps you collect sources from tabs, bookmarks, and history, then query and transform that content using AI.

https://github.com/PaulKinlan/NotebookLM-Chrome https://chromewebstore.google.com/detail/foliolm/eeejhgacmlh...

Thanks for sharing! We need more projects like this in the JS ecosystem.

I've been trying to arrive to something like this with my own sidepanel extension called Klue but its more of a user notes + web page context approach. Nice to see another take on this! https://chromewebstore.google.com/detail/cackjmmgcmnkjnffabk...

Very cool!

I'm particularly impressed by the bookmark "trick" to install it on a page. Despite having spent 15 years developing for the browser, I had somehow missed that feature of the bookmarks bar. But awesome UX for people to try out the tool. Congrats!

Thanks!

Bookmarklets are such an underrated feature. It's super convenient to inject and test scripts on any page. Seemed like the perfect low-friction entry point for people to try it out.

Spent some time on that UX because the concept is a bit hard to explain. Glad it worked!

Oh whoa, we are working in parallel on a similar angle!

We just launched Rover (https://rover.rtrvr.ai/) as the first Embeddable Web Agent.

Similar principles, just embed a script tag and you get an agent that can type/click/select to onboard/demo/checkout users.

I tried on your website and it was reeaaaally slow. Quick question:

- you are injecting numbering on to the UI. Are you taking screenshots? But I don't see any screenshots in the request being sent, what is the point of the numbering?

I don't think building on browser-use is the way to go, it was the worst performing harness of all we tested [https://www.rtrvr.ai/blog/web-bench-results]. We built out our own logic to build custom Action Trees that don't require any ARIA or accessibility setup from websites.

Would love to meet and trade notes, if possible (rtrvr.ai/request-demo)!

Is this open source?

We are about to open source next week!

If an AI agent runs inside the page and can see the DOM and the user’s session, how do you keep it safe without limiting what it can actually do?

Advantages and disadvantages of sandboxing agents with OS DAC/MAC, VM, container, user-space, WASM runtime, browser extension permissions, and IDK IFrames and Origins?

How are AI agents built into browsers sandboxed by comparison?

Recent work in sandboxing agents; https://news.ycombinator.com/item?id=47223974

This is the problem every agent has to face.

PageAgent’s differentiator is that site developers can embed it directly into their own pages. In that scenario, with proper system instructions plus a built-in whitelist/blacklist API for interactive elements, the risk is pretty manageable.

For the general-agent case, operating on pages you don’t control, the risk is definitely higher. I’m currently working on the human-in-the-loop feature so the user can intervene before sensitive actions.

Would love to hear other approaches if anyone has ideas.

This looks really useful! I'm having a hard time understanding how it might be used by each specific user, using their own LLM instance, though. Is that because it does not support that type of bring-your-own-llm scheme, or am I just not putting two and two together with some kind of chain of user authentication, then token exchange?

This library does not include a LLM services. The one on the homepage is only for demonstration and testing. The npm package and extension requires your own LLM api config. Doc here https://alibaba.github.io/page-agent/docs/features/models

I’ve been thinking about something like this. If it’s just a one line script import, how the heck are you trusting natural language to translate to commands for an arbitrary ui?

The only thing I can think of is you had the AI rewrite and embed selectors on the entire build file and work with that?

Everything happens at runtime, on the HTML level.

It uses a similiar process as `browser-use` but all in the web page. A script parses the live HTML, strips it down to its semantic essentials (HTML dehydration), and indexes every interactive element. That snapshot goes to the LLM, which returns actions referencing elements by index. The agent then simulates mouse/keyboard events on those elements via JS.

This works best on pages with proper semantic HTML and accessibility markup. You can test it right now on any page using the bookmarklet on the homepage (unless that page CSP blocks script injection of course).

Is this Affiliated with the Chinese company Alibaba? Any chance data goes there too?

Full transparency: I work at Alibaba and published this under Alibaba's open-source org. I sometines maintain it during work hours, so yes, Alibaba technically pays me for it. That said, this is my project — it's MIT-licensed, includes no backend service, and is open for anyone to audit.

The free testing LLM endpoint is hosted on Alibaba Cloud because I happen to have some company quota to spend, but it's not part of the library. Bring your own LLM and there is zero data transmission to Alibaba or anywhere else you haven't configured yourself.

I highly recommend using it with a local Ollama setup.

Thank you for sharing this!

Am I right in thinking you’re asking me to put an API in frontend code?

No and please don’t do that.

If you only use it as a personal assistant. You can connect to your llm service directly.

If you plan to integrate it into your web app. It’s better to have a proxy api for the llm and auth the request with cookie or something.

Any plans to support WebMCP? https://developer.chrome.com/blog/webmcp-epp

WebMCP doesn’t seem to be available for use inside webpages or extensions.

After looking into it more, I think PageAgent is actually a very good fit for WebMCP...

Cool, but likely to become obsolete with the rise of agents that ship with the browser.

I would much rather have a stack I control. Additionally this can be used as a library.

Curious - how does it perform with captchas and other "are you human" stuff on the web?

I added in the system prompt that it should skip CAPTCHAs and hand control back to the user. Currently working on a proper human-in-the-loop feature. That's actually one of the key advantages of running the agent inside your own browser.

Makes sense.

For curiosity's sake, have you had it try to attempt captchas?

If so, what were the results?

I haven’t. I don’t think it will work well.

I use a text-based approach. Captchas like “crossroad” usually need a screenshot, a visual model and coordinate-based mouse events.

i tested the chrome extension, it worked great - i asked it to change the light/dark mode of a website, it navigated to settings, clicked a few tabs, scrolled and found it to toggle the setting.

thanks for sharing!

Glad it worked well! The Chrome extension is my focus right now. It handles simple tasks pretty reliably and fast, but still has a long way to go for more complex workflows. Lots to improve.

Looks cool! Are you open to adding AWS Bedrock or LiteLLM support?

Thanks!

It supports any OpenAI-compatible API out of the box, so AWS Bedrock, LiteLLM, Ollama, etc. should all work. The free testing LLM is just there for a quick demo. Please bring your own LLM for long-time usage.

Does it support long-click / click-and-drag?

Not yet. Currently focused on the more common interaction patterns. PRs welcome though!

Gotcha. Still very cool! Congrats on the release.

Thanks!

How is this secure? Seems like this PageAgent could be the user pretty easily and cause all kinds of problems.

Could you elaborate on what kind of security problems you’re referring to? Like hallucination?

The PageAgent has access to the security tokens of the currently logged in user. They can do anything the user can on the site, including become them. What is to prevent the PageAgent from being exploited and send these security tokens elsewhere? It would be trivial for some other package to look for your PageAgent and override key functions, and then it is all over.

PageAgent operates at the HTML/DOM level with the same privileges as any other JavaScript running on the page and nothing more. The security token concern you're describing applies equally to every third-party script, npm package, or browser extension that runs in-page. It's not unique to PageAgent.

The browser extension can be more risky because it's more privileged. I've designed a simple authorization mechanism so that only pages explicitly approved by the user can call the extension.

That said, I'd welcome more eyes on this. If anyone wants to review the security model, the code is fully open source.

Confusing name because of the existence of pageant, the putty agent.

Came here to say missed opportunity to call it "PAgent". Rolls off the tongue better than Page Agent.

I'm 2 years too late for that one...

Darn. Pageant would've been a nice name though. Maybe `page-agent.js` is more relevant in web dev community.

I think page agent is good. I've never heard of putty's pageant. And I think it's better to distinguish it from general meaning of pageant (for beauty).

Thanks!

I think every successful Show HN post ends up with a "thought this was about X" or "didn't look up the name first?" comment. Consider it a win! I don't think anyone will mistake a tool for putty with your tool, but you might share a google search page with it.

Firefox support?

In my plan. Should be easy since I use wxt as the extension framework.

Reply on news.ycombinator.com