Work through forms
“Open the support form, choose Billing, and enter the issue details I gave you. Pause before submitting.”
Give Nabsun a task. It can navigate websites, click controls, fill forms, and work through multiple steps in your open tabs. You approve actions by default. Use the included local model, connect Codex or Claude, or bring your own AI provider.
Windows x64 and Apple-silicon Mac · About 1.4 GB, including the model
All downloads and installation detailsFree and open source. No account needed for the built-in model. Cloud providers have their own account requirements and charges.
Page: Service request
Open form → fill fields → review → approve submission
Illustrated examples, not a live assistant. Results depend on your model and the page.
Nabsun can use page controls and check what changed after each action. These are examples to try on supported sites; completion depends on the model, the page, and the permissions you allow.
“Open the support form, choose Billing, and enter the issue details I gave you. Pause before submitting.”
“Search for desk lamps, set the price filter to under $50, and open three options in separate tabs.”
“Find the notification settings and turn off the weekly digest.” Nabsun can follow links, change controls, and check the resulting page.
Keep control through approval prompts, allow specific tools, or enable Autopilot for write actions. You handle sign-in, CAPTCHA checks, and protected credential fields.
The first-run guide walks you through setup. You can reopen it in Settings → Getting started.
Choose your platform’s download. The local model is included, so there is no separate model setup.
Try the built-in model immediately, or click Connect for Codex or Claude in Extensions → Assistants. Nabsun installs the official CLI if needed and opens provider sign-in. No terminal or npm commands required.
Open a website and the assistant. Tell it what to do, such as “Open the contact form and fill in these details.” Review the requested actions and approve them to continue.
Switch models above the chat or in Settings → Model. Codex and Claude have a searchable picker, provider defaults, and custom model IDs. Each provider remembers its selection.
| Option | What you need | Where it runs | Useful for |
|---|---|---|---|
| Built-in Qwen3 1.7B Included | No account or API key | On your CPU | Simple questions and basic browser tasks |
| Codex / Claude Code | Connect in the app and sign in to an eligible provider account | Through the provider’s CLI and service | Tasks that benefit from a more capable model |
| OpenAI / Anthropic API | Your API key; provider usage charges apply | At the selected cloud provider | Choosing models through your own API account |
| Ollama / custom local model | Your Ollama server, or a compatible local GGUF model and engine | On the configured machine or server | Using a local model you already run |
Local inference can work offline. Websites and cloud assistants still need internet access. Cloud providers receive the prompts and page context sent to them; model access depends on your account.
Recorded on an earlier release. The current version adds onboarding, model selection, themes, and update notifications.
The YouTube player loads when you click. Open on YouTube ↗
Choose the built-in assistant, connect a provider, or skip setup and start browsing.
Find Codex and Claude models, use the provider default, or enter a custom model ID.
Classic, Playful, Blossom, and Ocean styles, each with light, dark, or system mode.
Checks at startup and every 24 hours. Check manually in Settings → App updates. Installation is manual.
Also improved: address-bar editing, keyboard navigation, and feedback when settings cannot be saved. Full release notes ↗
Downloads include the local model and are about 1.4 GB. Verify files against the release’s SHA256 checksums.
Run the setup executable to install Nabsun. The portable build runs without installation.
Windows installer ↗Portable executable ↗Unsigned build. Windows may show a SmartScreen warning.
Download and extract the ZIP, then move Nabsun to Applications. For Macs with Apple M-series chips.
Mac ZIP ↗Ad-hoc signed, not notarized. macOS may block the first launch. Review the release notes before opening it.
No Intel Mac or Linux installers in this release. All published release assets ↗
Nabsun is a developer preview. Review the assistant’s output and actions, especially before submitting forms or changing account data.
Reading pages is automatic. Clicking, typing, navigation, and page code require approval by default. You can allow individual tools or enable Autopilot for write actions. Autopilot stays enabled until you turn it off; page code has its own permission.
Yes. Nabsun exposes browser tools through MCP, a protocol that lets AI tools connect to other applications. Copy the connection configuration from Settings → Integrations → Connect an external agent. Calls use the same approval settings as the assistant sidebar.
The bundled model runs locally, so its inference does not send your prompts to a cloud model provider. Normal browsing still contacts websites, and update checks contact GitHub. If you choose a cloud assistant, it receives the context sent to it. CLI credentials remain in the provider CLI’s store.
You can edit soul.md with preferences such as your time zone or how you like answers. By default this file is shared only with local models. Settings shows its location and whether it is shared with the selected provider.
Open a GitHub issue with your Nabsun version, operating system, and steps to reproduce the problem. Remove passwords, API keys, and other private information from screenshots or logs.
Requires Node.js and npm. These commands install dependencies and fetch the local engine and model. Packaged downloads above do not require these steps.
git clone https://github.com/naveenalavilli/nabsun.git
cd nabsun
npm ci
npm run fetch:model
npm startIf Electron did not download, run npm run setup. Linux requires manual engine setup. See the development instructions.