An open-source browser with an assistant built in. It works in the tabs you already have open — reading, clicking, filling things in — and runs on whichever model you prefer, including one that lives on your own machine. It asks before it acts.
It reads the page,
connects the dots,
and keeps things moving.
A sample workflow · this illustration doesn’t run a real model.
Most of your work already happens in a browser. Nabsun lets an assistant help with it — in the same tabs, behind the same logins, while you watch it happen.
A small model ships inside the app and runs on your own CPU, so there’s nothing to sign up for and nothing to pay for. It keeps working with no internet connection at all, and you can point it at something larger whenever you like.
Reading a page happens on its own. Anything that changes something — a click, a keystroke, a navigation, a line of JavaScript — stops and asks you first. That holds wherever the request came from: the sidebar, your editor, or a plugin you added.
Nabsun is an MCP server too, so Claude Code or Codex in your editor can work in the browser you’re already signed into. Same tools, same approval prompts, whichever side the request comes from.
Keep a plain Markdown file — soul.md — with your name, your time zone, how you like answers. The assistant reads it so you stop reintroducing yourself. By default it is shared only with models running on your own machine; cloud providers are given none of it.
Ask about what you’re reading, pull details together from a handful of tabs — or hand over the whole errand and watch it click through the pages, fill in the fields, and come back with the result.
Try the sample workflows ↗Open the sidebar and ask about whatever you’re reading. The page comes along as context, so there’s nothing to paste in.
It finds the right elements and gets on with it — clicking through pages, typing into fields, picking from dropdowns, ticking boxes, moving between tabs — then re-reads what came back and carries on. The same loop you’d run yourself, without the tab juggling.
Anything that changes something waits for your yes — or switch on Autopilot for the session. When a call is genuinely yours, like which result you meant or what belongs in a field, it asks in the sidebar rather than guessing, and it won’t type passwords or card numbers at all.
Start with the model that comes in the box, or connect the one you already use. Switching is a couple of clicks, so you can pick whatever suits the task in front of you.
Qwen3 · 1.7B parameters
Runs on your CPU. No account, no key.
Add an API key and pick any model your account can reach — the field is free text, so newer ids work too.
YOUR PROVIDERAlready running models locally? Point Nabsun at your server and carry on using them.
YOUR HARDWAREAdd either CLI from the Extensions panel and Nabsun runs it for you, using the login it already has. No key is kept here.
YOUR WORKFLOWThe built-in model is happiest with everyday questions and getting around the web. For long, multi-step research, it’s worth switching to something larger.
Chromium and Electron underneath, MIT licensed on top.
Clone it, read it, and change whatever you’d rather have work differently.
git clone https://github.com/naveenalavilli/nabsun.git
cd nabsun
npm install
npm run setup
npm run fetch:model
npm run build
npm startThe model is about 1.2 GB to fetch, and packaged builds already include it, so an installed copy downloads nothing — which is why the installer is around 1.3 GB. Builds are not code-signed yet, so Windows SmartScreen will warn: choose More info → Run anyway, after checking your download against the SHA256SUMS.txt published with it.
It’s early and it’s moving quickly, so expect a few rough edges. Do have a read of the security notes before you sign it into accounts you care about.
What it is, which models it works with, and what you’ll need to get going.
It’s an open-source browser built with Chromium and Electron, with an assistant that can read pages, compare things across tabs, and use the browser’s own tools to get a task done. The rest of the time it’s an ordinary browser — bookmarks, saved passwords, downloads, session restore, all where you’d expect them.
Yes — the built-in Qwen3 model runs on your CPU, with no key and nothing leaving the machine. An installed copy already has it; if you build from source, you fetch the engine and weights once. Browsing the actual web, of course, still needs a network.
It works both ways. Nabsun runs an MCP server, so an agent in your editor — Claude Code, Codex, or anything else that speaks MCP — can drive your real, signed-in tabs. It also hosts MCP servers of its own, whose tools join the same list the sidebar uses. Either way the tools and the approval prompts are the same, and the bridge listens only on loopback, with a token that changes every launch.
Yes. Reading a page is automatic; clicking, typing, navigating, and running code all stop and ask. Autopilot waves through clicks, typing, and navigation for the session if you’d rather work uninterrupted. It deliberately stops short of running code on a page — that sits behind its own setting, off by default. The assistant also can’t touch password or one-time-code fields at all, and a denial is final — it’s told not to look for a way round it.
The built-in model, your own Ollama server, an Anthropic or OpenAI key, or the Claude Code and Codex CLIs you’ve already signed into — those keep their own credentials, so nothing is stored in the browser. Worth remembering that a cloud provider sees whatever context you send it, and will bill you for it.
Only what you write down yourself. There is a Markdown file in your profile folder called soul.md — your name, your time zone, how much detail you want, anything you would otherwise retype every session. Settings shows you exactly where it lives and opens it in your editor. It is yours: nothing rewrites it, and a line you delete stays deleted. By default it is only ever given to a model running on your own machine, so Anthropic, OpenAI and the signed-in CLIs never receive it unless you switch that off. Settings names the provider you are using and says whether the file is being shared or withheld.
Download the Windows installer from the releases page — there is a setup version that adds a Start-menu entry, and a portable single executable that runs without installing. Both already contain the built-in model, so they work offline straight away, and both are around 1.3 GB for that reason. They are not code-signed yet, so Windows will show a SmartScreen warning: choose More info → Run anyway. Checksums are published alongside, and you can verify one with Get-FileHash <file> -Algorithm SHA256. macOS and Linux builds are configured but not yet produced.
Three ways. Drop a folder of JavaScript into the plugins directory and the tools you register show up to the model like the built-in ones. Add an MCP server and its tools join the same list. Or load an unpacked Chrome extension — content scripts, storage, and much of the tabs and webRequest APIs work on the same Chromium engine.