Hi Product Hunt 👋 I'm Naveen, and I built Nabsun. Most of my work already happens in a browser, but the AI part of it always lived somewhere else — a separate tab, a chat window, copy-pasting context back and forth. I wanted the assistant to just be in the browser, working in the tabs I already had open, behind the logins I was already signed into. Two decisions shaped it: 🔹 It ships a model. Qwen3 1.7B is packaged inside the app and runs on your CPU — no account, no API key, and it keeps working with the network off. When you need more, point it at Ollama, an Anthropic/OpenAI key, or the Claude Code / Codex CLI you're already signed into. 🔹 Reading is free, acting asks. The assistant reads pages on its own, but anything that changes something — a click, a keystroke, a navigation — stops and asks you first. Running JavaScript on a page is a separate setting that's off by default, and it can't touch password or one-time-code fields at all. It's also an MCP server in both directions, so Claude Code or Codex in your editor can drive your real, signed-in browser through the same tools and the same approval prompts. Being upfront about where it's at: this is a v0.1.5 developer preview. Windows only for now (macOS is configured but not yet built), builds aren't code-signed yet so SmartScreen will warn you, and the installer is ~1.4GB because the model is inside it. Checksums are published with every release, and you can always build from source instead. It's MIT licensed — clone it, read it, change whatever you'd rather have work differently: github.com/naveenalavilli/nabsun I'd love feedback on the permission model especially. Does asking-before-acting land in the right place, or does it get in the way? And if you manage to get it to take a write action without prompting you, please open an issue — I genuinely want to know. Thanks for taking a look 🙏

Nabsun

Open-source AI browser with a local model built in

Open-source AI browser with a local model built in