Currently intended features?

[holdengand]

I am watching this project, can not use it as I am mostly a windows user and don't even have a mac accessible to even test where this is at yet; so I am on the wait list for windows support.

But I am curious about what features are currently planned, like what is the roadmap or what are high priority use cases that are currently driving development? Deep research? Web scraping? Price watching? Browser vision? integration with eternal agent tools like langchain, langflow, or similar? Other things past just chatting about the page or searching your tabs and stuff? Will it be able t9 fun agent tasks in the background while you continue to use the browser? The readme is a bit light on planned direction.

Honestly I by chance found this project looking to see if there were AI tools which could group and organize tabs with AI, and do so locally and not be dependent on some cloud server for a price. But there are a lot of other AI and web tools I just haven't been bothered to learn to set up cause I was tired of installing loads of tools only for them not to cover my needs or importantly were a pain to get to work together. As a non-programmer its gotten exhausting keeping up with the hundreds of tools now and installing so many virtual environments and pip dependencies and fighting docker installs all the time. Having a browser is a potentially smart and easy way to centralize a lot of this kind of stuff I think. So I am curious what tools and features are planed now and in the near future.

I did see the MCP store added to the readme, which is interesting, opens up a lot of potential. But still curious about what the devs here actually have planned inside the browser itself.

[shadowfax92]

@holdengand just added support for windows! Download from here - https://github.com/browseros-ai/BrowserOS/releases/tag/v0.10.0

[shadowfax92]

But I am curious about what features are currently planned, like what is the roadmap or what are high priority use cases that are currently driving development? Deep research? Web scraping? Price watching? Browser vision? integration with eternal agent tools like langchain, langflow, or similar? Other things past just chatting about the page or searching your tabs and stuff? Will it be able t9 fun agent tasks in the background while you continue to use the browser? The readme is a bit light on planned direction.

@holdengand MCP is next on our roadmap and then improvements to our agent experience. This should enable dee research type of tasks. We are happy to build what users what. What is your biggest usecase? :)

Honestly I by chance found this project looking to see if there were AI tools which could group and organize tabs with AI, and do so locally and not be dependent on some cloud server for a price

This is exactly why we build ollama integration! so people could use local models without paying for usage.

[holdengand]

just added support for windows!

Great to hear. I will experiment with this windows release later today.

What is your biggest usecase? :)

I have a lot of on-off niche use cases with AI personally. So I don't think I am the best person for probing for feature ideas. I'm not a developer or programmer, or anyone in a field that NEEDS this stuff. I just have on and off periods where I am experimenting with current tools to keep up to date and waiting for AI to replace us all. So my use cases are all over the place and erratic honestly. Objectively I am somewhere between a casual hobbyist and a wannabe power user.

That said, something I would really like to see that I think could have some wider appeal is some kind of built in webscraping. Of course there are tools for this, but unless you are paying for the services, hosting and using them locally in my experience has always been a bit of a pain and I end up doing a lot of manual work still so its never very efficient or piratical for me to do just for occasional hobby experimentation. And a webscraper built into a browser kind of makes sense to me, for casual users that are not scraping for entire markets of data or something anyway. And It would make powerful, targeted, webacraping readily available to more causal users. I have been interested in finetuning custom models for a while, but honestly have not been bothered enough to curate the datasets I need which would be most efficiently done with web scraping. And having AI agents which help in the process with more targeted scraping and deep research techniques could be quite powerful for a local running, low budget, user.

I might have more thoughts after playing with the current state of the browser.

[shadowfax92]

Thank you for the detailed reply!

AI agents which help in the process with more targeted scraping and deep research techniques

We are actually looking into add MCP support this should allow extensibility of browser for this usecase.

Also, the agent today already to some extent can do basic scraping, but not great at storing the data. We are working on making our agent lot better at this.

[holdengand]

@shadowfax92 Hello again. I have used BrowserOS for a bit, and I have a few thoughts so far. In no specific order:

• I notice that at the moment, to switch between model providers, you have to go to the settings every time. It would be nice to have a drop-down of the different providers the user has provided credentials for. I was switching back and forth between Gemini and Ollama, and a drop-down like you have in the "LLM Sidekick" would be nice. The sidekick seems like it's just loading the webpage for the selected service and nothing else special, but a way to select a service and a model in the Agent page in a similar way would be nice.

• For Ollama, it would be a nice quality-of-life improvement to auto-detect what models the user has installed; I'm pretty sure this can be done with the API pretty easily.

• It would be nice to have support for LM Studio and Kobold as well.

• I have Searxng set up locally, so support for that might be nice in the search engines; also, it could be beneficial to support it if you deal with more web scraping and deep research, as I think it's a bit more friendly for free searching without a bunch of APIs aand so on and believe a number of existing web scrapers and deep researching tools use or prefer it. I'm pretty sure DuckDuckGo is alright for this too, but I have had issues in the past where it complains about too many requests and so on because I often have a VPN active. If users don't have searxng set up for themselves there are plenty of people who host and make it public that I think could be used. I'm no expert on setting this stuff up though.

• I suspect this is on the to-do list, but a way to set up and save custom common prompts, like "Organize my tabs" or whatever.

• I haven't spent time thinking of a bunch of use cases, but support for vision models might be good, even if it's a separate agent that's called when needed. Currently, even if using a model that has vision, it can't detect anything in my testing. It could be beneficial for navigating some aspects of sites, just being able to pull information from images or graphics that aren't transcribed in text anywhere, or being able to provide live support for the user if they are working on something together so the user doesn't always have to type out what's going on; or even providing an image for context or help searching for something. Those are a few ideas where vision could be beneficial off the top of my head.

• It would be nice to have local LLMs accessible in the sidekick too, so the user can just chat with their LLM. I know this feature is in every single other application. But honestly, it seems rather easy to add and could easily just make this a good all-in-one tool for casual users. Also, I don't know what the long-term goals are for the 'Sidekick' are, but being able to access the local model makes sense if the rest of the app is aimed at supporting local anyway. Alternatively, make the "Agent" able to be toggled between 'conversational' mode and 'agent' mode, and maybe have an 'auto' mode where it tries to detect when it and the user are just chatting versus when it's actually being requested to start using agent capabilities to do work rather than chat about content on the page or something else. This mixed with STT and TTS could make it easy for the user and AI to engage in tasks without even needing to type and would allow this to be able to become a hands-free browser.

• This is more long term, I think, but a 'background mode' where the agent can do more complex tasks in the background without the user messing with it. In my testing, you can do this now by running two windows and having the agent work in one and you work in the other. And I am usually running more than one window, so this would be nothing new to me. But I think having some way for the agent to work behind the scenes sometimes would be handy. I don't know, it might not be practical to have an essentially headless browser running for the AI for this project. But just something I have thought about a few times testing.

---

And that's my thoughts so far with the current windows release.
In my experience so far the "Agent' often struggles with tasks or complains it doesn't have access to tools capable of navigation or searching and so on. I often give it slightly more complex tasks I'm guessing, but not stuff that seems unreasonable in that the steps to do it should be pretty easy to figure out. Stuff like "google dogs and open an image in a new tab" failed, even just having it navigate to google failed with Gemini 2.5 pro model. So, the agent needs some work as you already are looking to improve. I am curious what model/s are being used by the developers to test this so far?

There may have been other thoughts I had, but thats what I got for now.

[matrox75]

@holdengand
I only just read these comments and saw that you had already asked about LMStudio before. I would like to support your very good thoughts on features. When the MCP support is implemented, a whole new world opens up, e.g. to integrate your own langflow workflows into the browser. Otherwise, after reading your self-description I had one dejavu after another... that's me, that's me too ... caught between ambitious hobbyist and trail and error power user of langflow, n8n, anythingllm, vscodium etc. Fighting with docker, permanently disabling telemetry calls home in config.json, banning outgoing connections in little snitch. I'm tired of it. That's why browseros sounds so promising, if the developers keep their intentions and stick to them. Privacy and on premise have top priority.

[holdengand]

@matrox75

langflow, n8n, anythingllm, vscodium

Yup, tested all of these and like a dozen others.
Currently I honestly use google's AI studio so often just cause they have made it so easy to use for general AI stuff, as much as i don't lik just handing them all my data ALL the time. The past few days I discovered Dyad, and thats so so close to making it easy to make apps on demand, particularly for a non-coder (as long as you don't mind apps build in react, I have warmed up to it with Dyad). So if thats something you are interested in, between google gemini's free api and the free deepseek option in open router that it supports, for the cost of electricity you can at least get foundational apps up and running pretty easy and run them locally with a few conversational chats (it does support local llms to, but I haven't really tested them simply because I know they are not likely to hold up to my expectations on my limited powered pc). I have successfully made a few simple test apps that can use my local Ollama install quite easily.

This and that are currently the two open source projects that most have my attention atm.

Anyway, your LM studio request was more targeted and mines a bit easier to miss or ignore. I just wanted to try and open a dialogue about what this project was looking to be exactly.

[shadowfax92]

Thanks @holdengand sharing your detailed thoughts and feedback!

Closing this issue as it's more suited for discussions pane. Let take it forward there.

PS: Really glad to have you 🙂