Tokenelli — Private AI Sidebar, 100% On-Device

Summarize, rewrite & chat with a private AI that runs on your own GPU. No cloud, no account, no data ever leaves your device.

Tokenelli is a private AI sidebar that runs entirely on your own computer — no server, no API, no account. Gemma 4 (2.6B parameters) runs directly in your browser via WebGPU. Your prompts and the pages you read are processed on your GPU and never transmitted anywhere. The only network request Tokenelli ever makes is the one-time model download, cached locally after first use. WHAT IT DOES Chat with a local AI in the side panel, with full multi-turn memory Summarize any page, or just your current selection Rewrite, fix grammar, translate, or shorten text in place Inline toolbar for quick actions on selected text Work presets: professional tone, bullet points, action items, and more — plus your own custom prompts Digest all open tabs at once Context menu and keyboard shortcuts (Cmd/Ctrl+Shift+G to open, Cmd/Ctrl+Shift+S to summarize) BEFORE YOU INSTALL — PLEASE READ We'd rather you know this up front than be surprised: The AI model is a ~2 GB one-time download on first use. It's cached afterwards, so this happens once — but on a slow connection it takes a while. While loaded, the model uses roughly 2 GB of GPU memory. Tokenelli frees this automatically after a configurable idle period. You need Chrome or Edge 116+ with WebGPU, and a GPU capable of running it. Most GPUs from the last few years work; older or low-end integrated graphics may be slow or unsupported. This is a 2.6B-parameter model. It's genuinely good at summarizing, rewriting, drafting, and answering questions about the page in front of you. It is not a frontier cloud model, and we won't pretend otherwise. That's the trade: you give up instant setup and peak capability, and you get privacy, zero cost per use, and an assistant that works offline. WHY ON-DEVICE Genuinely private: there's no server to log your prompts or page content — by architecture, not by policy No per-token cost: it runs on your hardware, so the assistant is free, with no usage caps Works offline: once the model is cached, it keeps working with no network at all You control what it sees: per-site permissions with a preview before any page content is read No accounts. No telemetry. No data collection. Your conversations and browsing stay on your machine. Support: https://tokenelli.com/support/index.html Privacy policy: https://tokenelli.com/privacy/index.html