Show HN: BrowserOS – "Claude Cowork" in the browser

Posted by felarof 1 day ago

Hey HN! We're Nithin and Nikhil, twin brothers building BrowserOS (YC S24). We're an open-source, privacy-first alternative to the AI browsers from big labs.

The big differentiator: on BrowserOS you can use local LLMs or BYOK and run the agent entirely on the client side, so your company/sensitive data stays on your machine!

Today we're launching filesystem access... just like Claude Cowork, our browser agent can read files, write files, run shell commands! But honestly, we didn't plan for this. It turns out the privacy decision we made 9 months ago accidentally positioned us for this moment.

The architectural bet we made 9 months ago: Unlike other AI browsers (ChatGPT Atlas, Perplexity Comet) where the agent loop runs server-side, we decided early on to run our agent entirely on your machine (client side).

But building everything on the client side wasn't smooth. We initially built our agent loop inside a Chrome extension. But we kept hitting walls -- service worker being single thread JS; not having access to NodeJS libraries. So we made the hard decision 2 months ago to throw away everything and start from scratch.

In the new architecture, our agent loop sits in a standalone binary that we ship alongside our Chromium. And we use gemini-cli for the agent loop with some tweaks! We wrote a neat adapter to translate between Gemini format and Vercel AI SDK format. You can look at our entire codebase here: https://git.new/browseros-agent

How we give browser access to filesystem: When Claude Cowork launched, we realized something: because Atlas and Comet run their agent loop server-side, there's no good way for their agent to access your files without uploading them to the server first. But our agent was already local. Adding filesystem access meant just... opening the door (with your permissions ofc). Our agent can now read and write files just like Claude Code.

What you can actually do today:

a) Organize files in my desktop folder https://youtu.be/NOZ7xjto6Uc

b) Open top 5 HN links, extract the details and write summary into a HTML file https://youtu.be/uXvqs_TCmMQ

--- Where we are now If you haven't tried us since the last Show HN (https://news.ycombinator.com/item?id=44523409), give us another shot. The new architecture unlocked a ton of new features, and we've grown to 8.5K GitHub stars and 100K+ downloads:

c) You can now build more reliable workflows using n8n-like graph https://youtu.be/H_bFfWIevSY

d) You can also use BrowserOS as an MCP server in Cursor or Claude Code https://youtu.be/5nevh00lckM

We are very bullish on browser being the right platform for a Claude Cowork like agent. Browser is the most commonly used app by knowledge workers (emails, docs, spreadsheets, research, etc). And even Anthropic recognizes this -- for Claude Cowork, they have janky integration with browser via a chrome extension. But owning the entire stack allows us to build differentiated features that wouldn't be possible otherwise. Ex: Browser ACLs.

Agents can do dumb or destructive things, so we're adding browser-level guardrails (think IAM for agents): "role(agent): can never click buy" or "role(agent): read-only access on my bank's homepage."

Curious to hear your take on this and the overall thesis.

We’ll be in the comments. Thanks for reading!

GitHub: https://github.com/browseros-ai/BrowserOS

Download: https://browseros.com (available for Mac, Windows, Linux!)

Comments

Comment by arjunchint 1 day ago

Hey cool stuff since last update!

I still don't buy the we needed it to be a whole Browser and not a Chrome Extension argument:

- your interface is still literally a chrome extension side panel

- none of the agentic browsers from the bigger players like Atlas and Comet really took off either

I do think the server side integration is required:

- with rtrvr.ai a ton of users are integrating our web agent chrome extension via Remote MCP from chatgpt.com as well as triggering as an API endpoint remotely. Your implementation is limited to only local connections as I understand.

- the biggest unlock for users is running at scale, so just being able to launch a hundred cloud browsers, do a task, and return results while you do other things. So we see hybrid cloud/local execution as the key unlock for this year

Your workflow pipeline is really cool! Any blog post/summary on how you set it up?

Last year was a lot of technical builders exploring the capabilities, and I am excited for this year of making these agentic browsers useful!

Comment by johnsmith1840 1 day ago

Extensions are limited though.

One simple example is an extension can't see cross origin iframes. This means it could never do soemthing like fill out a payment form for you if it's an extension.

Limited computation and action space is another as well as bot detection systems.

For example a javascript method trying to automate something like microsoft word in an iframe will have a tough time because the second you inject code in there they will block you.

Comment by arjunchint 1 day ago

> One simple example is an extension can't see cross origin iframes

Sounds like a skill issue, our web agent is able to interact with cross origin iframes to for example solve captchas: https://www.youtube.com/watch?v=LD3afouKPYc

We honestly haven't faced any bot detection or blocking issues. Owning the browser layer exposes to you much more detection just look at Comet getting blocked on Amazon etc.

Comment by johnsmith1840 1 day ago

With specific user permission to do so sure but in general it is blocked.

Comment by johnsmith1840 1 day ago

You're still limited in lots of annoying ways though

Comment by quarkcarbon279 1 day ago

what permissions are you talking about? No user permissions/any insecure permissions are needed to navigate cross origin iframes, shadow DOMs and likewise. It comes down to your architecture choices and capabilities - rtrvr can navigate these diff realms without ever taking debugger or such insecure permissions

Comment by felarof 1 day ago

Thanks!

> whole Browser and not a Chrome Extension argument

Both of us are definitely biased to think our own approach is better :)

But without owning the binary, we couldn't shipped today's feature -- Agent with access to your filesystem and being able to run shell commands like Claude Cowork.

> your interface is still literally a chrome extension side panel

Yep, our interface is a chrome extension to make iterating on the UX faster. But it uses a ton of C++ APIs that we expose under `chrome.browseros.*`

> Your workflow pipeline is really cool! Any blog post/summary on how you set it up?

Thanks! We'll look into publishing a blog soon!

Comment by arjunchint 1 day ago

> But without owning the binary, we couldn't shipped today's feature -- Agent with access to your filesystem and being able to run shell commands like Claude Cowork

Chrome Extension can also access local files and can also execute LLM generated code in sandboxes

Comment by felarof 19 hours ago

You can't run shell commands though. Bash is the most powerful tool in claude code!

Comment by tekacs 1 day ago

https://github.com/browseros-ai/BrowserOS/issues/99#issuecom...

I didn't hear back there, but huzzah, it looks like this is in there. I'm glad to see it!

Comment by felarof 1 day ago

Thanks for initial feature request! We do read every single request :)

Yes, we expose BrowserOS as an MCP server -- that you can use from claude code, cursor, opencode, etc -- https://docs.browseros.com/features/use-with-claude-code

MCP server works out of box (unlike Chrome DevTools MCP which requires tricky setup).

Comment by 1 day ago

Comment by jm4 1 day ago

Really cool product. How do you plan to monetize it?

You guys need some marketing help. There’s a lot of potential here, but you don’t do a good job of selling it. Tell me what problems I’m going to be able to solve or what headaches it will eliminate. Can it going into that shitty Canvas app my kids’ school uses, identify outstanding assignments or low grades and send me a daily text summary? Can it automate buying everything on my grocery list and setting up delivery? Or look up flight options, ask me what I want and book it for me? Even better, I’m stuck having to look up international flights for 7 people in three households, get everyone to agree on one and then book them. Please build something that will do that.

Keep at it because this thing is cool!

Comment by felarof 19 hours ago

> You guys need some marketing help. There’s a lot of potential here, but you don’t do a good job of selling it.

Thank you for the feedback. Ack, we need to do a better job of marketing.

> How do you plan to monetize it? Our goal is to eventually to sell license for enterprise browsers.

Comment by jm4 17 hours ago

Have you ever thought about a marketplace for premade workflows? Or a library of integrations that are already tested that a user can mix and match to create complex automations? Or access to more MCP servers?

For example, it would be really neat to trigger jobs that perform some task and then make a call to Twilio or something to send an alert. Or some building blocks that tie into my Square account or Amazon account. I want to be able to describe the results I want, but I don’t want to explain how to interact with a particular service and then test that.

I would love to be able to give a prompt like this: “review my item library in Square, identify items that are missing descriptions or are miscategorized, propose the fixes, and confirm with me before making any changes.” That’s an extremely tedious task that requires a lot of clicking and page loads. I hate it and I would pay for your product if you could save me that time.

Or this: “Every month, alert me to any fluctuations in product cost and which items in my Square catalog are affected. Highlight any items where my COGS exceeds 35%. All the invoices are available in my email.” That would be incredibly powerful. Doing this manually can take days.

Comment by felarof 15 hours ago

> “Every month, alert me to any fluctuations in product cost and which items in my Square catalog are affected. Highlight any items where my COGS exceeds 35%. All the invoices are available in my email.” That would be incredibly powerful.

You could try this use case on on agent builder even today. We also have a scheduled tasks for you to schedule it to run monthly

> Have you ever thought about a marketplace for premade workflows?

We want to do this and are moving towards that! But we first need to make the premade (or user published) workflows very reliable.

Comment by 1 day ago

Comment by 4b11b4 1 day ago

IAM for agents sounds interesting but how is it reliably enforced? You also built evals?

Comment by felarof 1 day ago

Thanks!

> how is it reliably enforced?

At the chromium level, you have access to every single DOM element and coordinate space around it. So, when a click happens either user or agent, we have a neat way of enforcing required action (either allow it or nullify the click).

We are still at early version. And mostly targeting enterprise sites (like SAP) which don't change that often.

What use case did you have in mind?

Comment by mossTechnician 1 day ago

> we're adding browser-level guardrails (think IAM for agents)

This sounds interesting, but where would I go to see these guardrails and their implementation? I tried searching in the repository and couldn't find them.

Comment by felarof 1 day ago

We are still in early versions of the feature! Haven't released on our repo yet.

What use case did you have? Happy to show a demo of current version we have (you can hit me up on discord or slack -- links available on our repo)

Comment by Johnny_Bonk 1 day ago

What would be great is if it could work in the browser like Claude in chrome and communicate (with my control) back to objects on my desktop like my ide for example or really anything

Comment by felarof 1 day ago

Ohh, interesting, technically this should already be possible. Because we already package gemini-cli into the sidecar (bun) binary. We just have to create a good UX.

What angle are you looking at this from? Is it for convenience? Or do you not like terminal UI and need a web-friendly UI for these agents?

Comment by thawab 1 day ago

Thanks a lot, i wanted to ask about the headless agent use case: How does it compare to using https://github.com/vercel-labs/agent-browser

Comment by felarof 19 hours ago

BrowserOS should work with agent browser as well in headless mode!

On top of that, if you want headful mode, you can use our MCP server https://docs.browseros.com/features/use-with-claude-code

Would love to understand your use case! You can hit me up at nithin[at]browseros.com

Comment by devld 19 hours ago

Genuine question. What's the use case for agentic browsing?

Comment by felarof 16 hours ago

There are bunch of tedious / routine tasks that AI can automate.

I think the big hurdle is mostly education / shift in mindset. We are so used to doing the task manually that most of us (including me) don't pause to think if I should be doing this or can I give to an agent.

Comment by grigio 1 day ago

It seems cool, will it work in headless mode without X11/Wayland/.. ?

Comment by felarof 20 hours ago

At the moment, our product is designed for headful mode.

But if you want to use our browser in headless and use playwright that would work too! (we are chromium fork)

Comment by rahimnathwani 1 day ago

Which local model works best with this? (Assuming MacOS with 32GB unified RAM)

Comment by felarof 1 day ago

gpt-oss 20B works well. You'll want at least 12k context length for agent mode.

Comment by ivysly 1 day ago

why are you calling this an OS

Comment by felarof 1 day ago

Good question. We think the browser is becoming the new OS. It doesn’t really matter anymore if you’re on Windows, macOS, or Linux—the browser is where most work already happens.

We see a future where it’s the main gateway to everything, and where agents live and work alongside you inside the browser. That’s why we call it BrowserOS. :)

Comment by p1necone 1 day ago

Is this really true? Mobile device users are all mostly forced to use apps rather than the browser for most stuff, and people on desktop PCs/laptops are probably either using them for gaming (all desktop apps), or work where a lot of stuff is desktop apps.

Sure regular consumer stuff like social media is webapps (if they're not mobile only), and if you're interacting with like salesforce or a customer support tracker or an issue tracker or something you're likely using a webapp, but the move to mobile devices for most consumer stuff means that people still using PCs are largely power users.

Comment by felarof 1 day ago

> if you're interacting with like salesforce or a customer support tracker or an issue tracker or something you're likely using a webapp

Precisely. I think most knowledge work (especially at business) still happens browser. That is the workflow we want to target!

Comment by ripped_britches 1 day ago

While that may be true, honest feedback is that it is confusing, possibly even misleading. But I hope whatever you pick works for you

Comment by sbsnjsks 1 day ago

[flagged]