from Hacker News

Show HN: Nxtscape – an open-source agentic browser

by felarof on 6/20/25, 4:35 PM with 206 comments

• by kevinsync on 6/20/25, 9:17 PM

  IMO comments so far seem to be not seeing the forest for the trees -- I can imagine incredible value for myself in a browser that hooks into a local LLM, writes everything it sees to a local timestamped database (oversimplification), parses and summarizes everything you interact with (again, oversimplification -- this would be tunable and scriptable), exposes Puppeteer-like functionality that is both scriptable via code and prompt-to-generate-code, helps you map shit out, remember stuff, find forgotten things that are "on the tip of your [digital] tongue", learn what you're interested in (again, local), help proactively filter ads, spam, phishing, bullshit you don't want to see, etc, can be wound up and let go to tackle internet tasks autonomously for (and WITH) you (oversimplification), on and on and on.

  Bookmarks don't cut it anymore when you've got 25 years of them saved.

  Falling down deep rabbit holes because you landed on an attention-desperate website to check one single thing and immediately got distracted can be reduced by running a bodyguard bot to filter junk out. Those sites create deafening noise that you can squash by telling the bot to just let you know when somebody replies to your comment with something of substance that you might actually want to read.

  If it truly works, I can imagine the digital equivalent of a personal assistant + tour manager + doorman + bodyguard + housekeeper + mechanic + etc, that could all be turned off and on with a switch.

  Given that the browser is our main portal to the chaos that is internet in 2025, this is not a bad idea! Really depends on the execution, but yeah.. I'm very curious to see how this project (and projects like it) go.

• by hannob on 6/20/25, 5:49 PM

  Okay, maybe this is a stupid question, but: what is an agentic browser? You seem to assume that everyone knows what that means.

  Is this a common and well-defined term that people use? I've never heard it.

  It would appear to me from the context that it means something like "web browser with AI stuff tackled on".

• by mullingitover on 6/20/25, 10:01 PM

  On the one hand: an agentic browser sounds like a cool idea. Being able to automate things with an agent on the client side is crazy powerful.

  On the other hand: this has the potential to be an absolute security Chernobyl. A browser is likely to be logged into all your sensitive accounts. An agent in your browser is probably going to be exposed to untrusted inputs from the internet by its very nature.

  You have the potential for prompt injection to turn your life upside down in a matter of seconds. I like the concept but I wouldn't touch this thing with a ten foot pole unless everyone in the supply chain was PCI/SOC2/ISO 27001 certified, the whole supply chain has been vetted, and I have blood oaths about its security from third party analysts.

• by gtsop on 6/20/25, 4:51 PM

  Are we still tossing around the 10x productivity boost? Please make this stop. I see first commit on April 28 so by 10x productivity its like you've been working on this for almost 2.5 years, and there is still a waiting list on the website.

  Appreciate the agplv3 licence, kudos on that.

• by Tsarp on 6/21/25, 1:25 PM

  Why not something like https://github.com/nanobrowser/nanobrowser.

  Its kinda built really well without exposing webdriver etc and can comfortably run js and communicate with LLMs.Has full agentic capabilites.

  Why a new browser instead of a robust extension?

• by ppqqrr on 6/21/25, 5:26 PM

  Chromium, Electron and VSCode are all the rage to fork as an LLM application base. In truth it doesn't really matter which you start from, they all lead to the One Local LLM App. There's already a fairly deterministic path of development, we're just waiting to see who wins the race in execution and marketing.

  In the meantime, the bigger opportunity with relatively little competition at the moment is the Web itself, not which application to browse it with. The Web absolutely sucks, and that's most of the reason we even feel the need for an "elevated browser experience" in the first place (i.e. lifted trucks on the information highway).

  The Web sucks, because it was built naively, then optimized for profitable friction. But all of it stood on the assumption that the cost of production on the web involves highly skilled human labor. LLMs have shattered that assumption, but the effects have not manifested. Which is to say that the entire existing Web is probably going to become a marginal, legacy corner of a much bigger base of LLM-driven hypermedia contents that is yet to come.

• by arjunchint on 6/22/25, 8:56 AM

  As others have mentioned, these agentic capabilities are fully achievable within a Chrome Extension.

  In fact, we built one, rtrvr.ai that has even better Web Agent performance than Open AI's Operator with human assistant and 7x faster than leading competitor: https://www.rtrvr.ai/blog/web-bench-results

  Your Accessibility Tree requirement is a poor excuse, rather you should build up an agent from a first principles understanding of DOM interactions.

  A browser is a SERIOUS security risk, you need a dedicated team to just pull in the latest security patches that Google pushes to Chromium or your users are sitting ducks to exploits and hacks...

• by wanderingmind on 6/21/25, 4:16 AM

  My usecase is to extract CSV or other data files from websites pertaining to water. For example here is the historical reservoir level data in South Australia[1]. Extracting the data was a big pain (at least for me with minimal front end experience). If downloading this can be automated with agents, I'm definitely giving it a shot.

  [1] https://data.sa.gov.au/data/dataset/reservoir-volumes-2018

• by varenc on 6/21/25, 1:06 AM

  This is quite cool! Very excited for this concept. Kudos to launching on HN!

  some genuine feedback on a frustrating early experience:

  - I ran the suggested "Group all my tabs by topic" in productivity agent mode. It worked great.

  - I then asked it to remove all tab groups and reset things, but was told this:

      This is a browser automation task. Please use **Agent Mode** for web interactions like clicking, filling forms, navigating sites, or extracting web content.
  

  - Tried "agent mode" and was told:

      This is a productivity task. Please use **Chat Mode** for tab management, bookmarks, sessions, history, and content analysis.
  

  - Basically was being sent back and forth. Went back to productivity mode and argued with it for a bit. The closest I could come to it removing all tabs groups was creating a new tab group encompassing all tabs, but couldn't get it to remove groups entirely. I'm guessing it might lack that API?

  Overall, it'd be nice if every browser level action it took had an undo button. Or at least if it was smart enough/able to remove the tab groups it just created.

  Will keep playing with it more.

  edit1: one more weird issue: While running the chat interface on chrome internal pages like chrome://extensions, it would randomly browse me to google.com for some reason.

  edit2: confirmed that productivity mode lacks a tool to ungroup tabs, just a tool to create tab groups.

• by lecro on 6/20/25, 10:21 PM

  “nxtscape” gives the good old "SCSI" vibes. People just nailed “GPT”. Please consider something you can say in one breath.

  Great product though.

• by blindriver on 6/21/25, 3:33 PM

  Sounds like a pretty clear case of trademark infringement. Nxtscape is very confusingly similar to Netscape.

• by dvt on 6/20/25, 8:42 PM

  Suffers from the same problem as all other AI "workflows"--no one wants to fucking chat with a computer (Brave sort-of does this, and it's god-awful). A chat interface should only be used as a fallback if the agent is too dumb to figure out what I want.

  A chat interface works for ChatGPT because most folks use it as a pseudo-search, but productivity tools are (broadly speaking) not generative, therefore shouldn't be using freeform inputs. I have many thoughts on fixing this, and it's a very hard problem, but simply slapping an LLM onto Chrome is just lazy. I don't mean to be overly negative, but it's kind of wild to see YC funding slop like this.

  And that's exactly what this is: slop. There's no technical creativity here, this isn't a new product segment, it barely deserves the "hey bro, this might be a feature, not a product" startup 101 criticism. It's what ChatGPT would spit out if you asked it what a good startup idea would be in 2025. All we need to do, even if we were being as charitable as possible, is ask who's doing the heavy lifting here (hint: it's not in the Github repo).

• by deepdarkforest on 6/20/25, 8:08 PM

  This is definitely a winners take all market. Kudos for giving it a shot, but imo browser projects are just too big for a team of 2/3. Plus, google has already demoed at IO the first hint at this. IMO you just cannot move fast enough to grab enough market share as a first/second mover before google just does it on chrome, and that's assuming you can outcompete with Dia in the first place. Even browser-use can do this also, and they have good distribution already.

  Good luck, but in your place i would at least start with something that a certain ICP needs more. Many, many manhours have been wasted by ambitious technical founders on taking down Chrome. (many also starting from a chrome fork itself). But none of them succeeded. We only have limited energy

• by gwhr on 6/24/25, 2:44 PM

  Would there be rate limiting applied for these / IP banned if used for web automation tasks? Is this treated as a bot by when visiting other websites?

• by zahlman on 6/20/25, 10:04 PM

  > And honestly, we feel like we're constantly fighting the browser we use every day. It's not one big thing, but a series of small, constant frustrations. I'll have 70+ tabs open from three different projects and completely lose my train of thought. And simple stuff like reordering tide pods from amazon or filling out forms shouldn't need our full attention anymore. AI can handle all of this, and that's exactly what we're building.

  So your thesis is that an AI agent should decide what I pay attention to, rather than me?

  What could possibly go wrong?

• by symbolford on 6/22/25, 7:59 AM

  btw, if you want to speed up the build times you should look to use RE API or Goma. it can make a build finish in about 15 to 30 minutes and caches changes to speed things up further once a full build is completed.

  https://chromium.googlesource.com/chromium/src/+/main/docs/l...

  It seems like gemini 2.5 is pre-trained on chromium given it has a lot of context into the current state of the codebase. That may help a bit.

  Also, I've not tried this yet, but maybe using an MCP server over a clangd language server is a solution to the context problem?

  https://chromium.googlesource.com/chromium/src/+/master/docs...

• by xena on 6/20/25, 5:08 PM

  Do you respect robots.txt?

• by joshstrange on 6/21/25, 3:55 PM

  I got a chance to play with this a bit and while I see the promise it's still a far way off.

  I asked it to find the best price for a new Seagate Exos 24TB drive on Amazon (GPT-4o):

  > The best price found for a new Seagate EXOS 24TB enterprise hard drive on Amazon is $274.99 (model ST24000NM000C). This is a renewed/refurbished unit with 4.1/5 stars from 39 reviews. Note: The search results show mostly renewed units, and new units may be available at different prices on individual product pages.

  Ok, fair, you couldn't find a new one on Amazon. I would have prefered a "Couldn't find one" instead of leading with the price and only in the second sentence mentioning it's a refurb. Note, this is mostly an LLM failing, not Nxtscape, that part seemed to work well overall.

  I then followed up with "I only want a new drive, please look at other popular electronic retailers". It replied with (formatted slightly):

  > Best Buy - WD Blue 2TB Internal Hard Drive ($59.99, on sale from $70)

  > Newegg - Desktop Internal Hard Drives (500GB up to 8TB available) - Price range: $180-$299 - 4.5 store rating - Features: 30-day returns, 5-9 day delivery

  > Amazon - Seagate EXOS Enterprise Hard Drives (up to 24TB)

  > Additional options include SSDs from Newegg ($60-$201) if you're interested in faster storage solutions.

  As almost always with LLMs, I see where it went off the beaten path. I didn't specify "a new one with the same specs I originally asked for", I shouldn't have to. This is probably mostly on the LLM, I don't know if Nxtscape could improve that with prompting (I don't know exactly what they are sending to judge either way). Also it got lazy with the Amazon response (no price).

  One way that Nxtscape might be able to improve is to parse out what the user is asking for, creating a data structure to define a "result" (in my case: url, name, price, description?), use that to prompt the LLM to conform to that shape, then take the results and pass them all through a one-off LLM instance to summarize the data. I think that would help with the inconsistencies in the results. Then again, that's very "Data extraction"/"data lookup"-focused and I haven't even played with using it for input: Fill out this form, loop this process for input (mail merge), etc.

  Really cool idea and I'll try throwing some other problems at it as I think of them, but mostly for fun/research, this doesn't seem like force-multiplier for my normal workflows (yet).

• by admiralrohan on 6/21/25, 6:49 PM

  1. How will you earn money? Sustainig the project? 2. Which LLM is used? Only "local LLM" is mentioned. Hardware requirements?

• by b0a04gl on 6/20/25, 5:04 PM

  so agents can control tabs, forms, clicks—like a real user would.so what about undo. if an agent clicks the wrong thing, how do you roll that back without reloading the world?

• by awongh on 6/20/25, 6:21 PM

  I think LLMs could have a reasonable chance at solving tab-related workflows (keeping track of tabs or the idea/concept of tabs) - that is tracking and sorting lots of small related research ideas.

  Sort of like a backwards perplexity search. (LLM context is from open tabs rather than the tool that brings you to those tabs)

  I built a tab manager extension a long time ago that people used but ran into the same problem- the concept of tab management runs deeper than just the tabs themselves.

• by lucasyvas on 6/21/25, 3:44 PM

  This is the wrong product. The right product is an AI-powered operating system that presents webviews.

  Web browsers as applications are made completely useless by the AI wave and I fully expect only the webview portion to survive long run.

  Microsoft, Apple, and Google are best positioned to capitalize on this. Meta is further behind but has decent glasses.

  The pendulum is on its way back to native.

• by ajb on 6/20/25, 10:04 PM

  I think there is big value in something that is on the side of the user, in particular in processing user-hostile material. Here are some concrete use cases:

  * Buying a sofa. You want to filter for sofas of a specific size, with certain features, marketing sites want to feed you a bunch of marketing slop for each sofa before giving you the details . This generalises to many domains.

  * You have a few friends who are still stuck on Facebook, you want to be notified if they post anything and avoid other rubbish

  * The local neighborhood is stuck organising in a Facebook group or even worse, nextdoor. You want to see any new posts except for those couple of guys who are always posting the same thing.

  * A government consultation website has been put up, but as a hurdle the consultation document has been combinatorially expanded to 763 pages by bureaucratic authoring techniques. You want to undo the combinatorial expansion do you can identify things you actually care about.

• by midenginedcoupe on 6/22/25, 11:22 AM

  Is it me or does the browser feel like the wrong level for this sort of helper? MS have the right idea in doing it at the OS level, it's just their implementation sucks becuase of the non-trivial security implications.

• by ellisd on 6/20/25, 8:08 PM

  Why obfuscate the LLM system prompt in your Github repo when it's going to be completely visible in the network inspector?

• by mahoro on 6/20/25, 6:01 PM

  This is great, I'd like to test! Is there any recommendations on which ollama models works best with this kind of tasks?

• by anilgulecha on 6/20/25, 4:51 PM

  Very interesting approach. Why a browser, and not a fantastic chrome extension? Grouping tabs, summarizing, even taking open ended actions, seem very doable with permissions extensions have..

  edit: Just read about the accessibility thing, but that's thin. Is there any usecase in the future that a browser can, but an extension can't?

• by lxe on 6/20/25, 4:53 PM

  Before I dive into the source code... how do you pass the page content, and the locations of interactive components to the LLM? And how do you dispatch events to interact with the page? I just want to verify if it's ARIA tree like the others, or it's something else.

• by kerv on 6/20/25, 8:12 PM

  This is cool thanks for sharing

  What is the tech around the thing that segments out DOM elements automatically and shows the visual representation. I think something like this would be great for automated UI testing agents?

• by 8organicbits on 6/20/25, 11:11 PM

  > simple stuff like reordering tide pods from amazon [..] shouldn't need our full attention anymore

  There's a straw man here. If you want to reorder an item on Amazon: click on 'order history', scroll, and click buy. This is a well-optimized path already and it doesn't require your full attention. I suspect the agent approach takes more effort as you need to type and then monitor what the AI is doing.

• by b0ner_t0ner on 6/21/25, 3:10 AM

  How is your product compared to Browser Base https://www.browserbase.com/ ?

• by OsrsNeedsf2P on 6/20/25, 5:31 PM

  Is this only for MacOS? If it's a Chromium fork, what's the reason for no Linux/Windows?

  Also what's the business model?

• by Babkock on 6/20/25, 5:35 PM

  Yeah that's just what we need, more AI shit, more slop slapped on top of Chromium.

• by _fw on 6/20/25, 5:14 PM

  I’m using Dia a lot for work at the moment and frankly it’s a gamechanger. granted I’m not a developer but being able to interact with an LLM that has access to the page I’m on is extremely useful:

  Instead of manually hunting across half a dozen different elements, then copy/paste and retype to put something into a format I want…

  I can just get Dia do it. In fact, I can create a shortcut to get it to do it the same way every single time. It’s the first time I’ve used something that actually feels like an extension of the web, instead of a new way to simply act on it at the surface level.

  I think the obvious extension of that is agentic browsers. I can’t wait for this to get built to a standard where I can use it every day… But how well is it going to run on my 16GB M1 Pro?

• by zahirbmirza on 6/20/25, 5:04 PM

  Does it support MP4 playback?

• by finolex on 6/20/25, 6:01 PM

  This is cool! Congrats on launch! How do you store user data? Do you write to device? Curious if there's a basic.tech x nxtscape collab possible here where you can store each user's info to their dedicated PDS

• by pnw_throwaway on 6/21/25, 12:41 AM

  Feels like yet another solution in search of a problem.

  All the same, looks like y’all are having fun working on it, and maybe some unforeseen usecase will bubble up.

• by revskill on 6/20/25, 6:09 PM

  When windows ?

• by ugh123 on 6/20/25, 7:46 PM

  That tab grouping capability is leaps and bounds better than the junk Google put into Chrome for "AI suggested" grouping.

• by afeigenbaum on 6/20/25, 6:49 PM

  What is on your roadmap?

• by thisislife2 on 6/20/25, 5:39 PM

  I've upvoted to encourage your initiative, but I personally will not support any "AI" software unless it 100% runs locally and supports old platforms and hardware. Otherwise it is nothing but another conduit to get access to, and suck all my personal data for surveillance capitalism.

• by bosky101 on 6/21/25, 5:53 PM

  As a consumer I'm asking myself why not just make an extension. I can see myself trying and picking 2 or more browser addons to to turbo charge + chat with any page.

  From a founder hat, I can see why the code base is a moat, hard problem. I hope the effort is worth the cost.

  Good luck.

• by rodolphoarruda on 6/20/25, 6:44 PM

  This is the missing piece from Karpathy's keynote: the browser.

• by Lammy on 6/20/25, 5:35 PM

  AOL still have active trademarks for “Netscape” which might trouble you here:

  - https://tsdr.uspto.gov/#caseNumber=76017078&caseSearchType=U...

  > PROVIDING MULTIPLE-USER ACCESS TO A GLOBAL COMPUTER INFORMATION NETWORK FOR THE TRANSFER AND DISSEMINATION OF A WIDE RANGE OF INFORMATION; ELECTRONIC TRANSMISSION OF DATA, IMAGES, AND DOCUMENTS VIA COMPUTER NETWORKS; [ELECTRONIC MAIL SERVICES; PROVIDING ON-LINE CHAT ROOMS FOR TRANSMISSION OF MESSAGES AMONG COMPUTER USERS CONCERNING A WIDE VARIETY OF FIELDS]

  - https://tsdr.uspto.gov/#caseNumber=76017079&caseSearchType=U...

  > PROVIDING INFORMATION IN THE FIELD OF COMPUTERS VIA A GLOBAL COMPUTER NETWORK; PROVIDING A WIDE RANGE OF GENERAL INTEREST INFORMATION VIA COMPUTER NETWORKS

  - https://tsdr.uspto.gov/#caseNumber=74574057&caseSearchType=U...

  > computer software for use in the transfer of information and the conduct of commercial transactions across local, national and world-wide information networks

• by wongarsu on 6/20/25, 5:05 PM

  Name derived from Netscape (Firefox's great-grandfather), icon is a red fox, but based on Chrome? Was this originally designed as a Firefox fork or what happened there

• by jklinger410 on 6/20/25, 4:55 PM

  What is with Mac users forking Chromium and then only making releases for Mac?

• by taylorius on 6/20/25, 5:12 PM

  I'll get voted down, but I hate that cute AI fox, and hope I never see it again.