Tarn — Private Local AI Assistant Autor: BroMind

Chat with the page you're on — using the models already running on your machine.

Tarn connects your browser to a local model server you install and puts it in a side panel on every page.

• Works out of the box with your local server. Tarn finds it, sets up the Ollama connection automatically, and can download a starter model suited to your hardware. If you already run Ollama, you're chatting in under two minutes.

• Reads the page — honestly. Ask about the page you're on. Tarn shows how much of it fits your model's context window, and condenses long pages in visible steps — you see each part it summarizes — instead of silently truncating. When a page can't be read (browser pages, canvas apps), it tells you why.

• Chat with PDFs and images. Attach a PDF to ask about its contents, or attach an image and let a vision-capable model describe it.

• Selection actions everywhere. Explain, summarize, translate, or rewrite selected text from the right-click menu — or from a floating toolbar that can rewrite text in place in any text field (on sites you allow).

• Built for real work. Streaming replies you can stop or regenerate, model switching mid-conversation, a full-tab view for long sessions, a prompt library you insert with /, edit-and-resend for any of your messages, searchable local history with visible token budgeting, per-chat system prompts and settings, JSON export/import, and copy-as-Markdown.

• Private by architecture. No account. No cloud. No telemetry. Prompts and page content go only to the model server you configure — normally your own machine — and one click wipes all data. The only host access granted at install is localhost/127.0.0.1; access to any website is granted by you, per site, at runtime. Check it in your network inspector: besides talking to your own server, the only traffic is what you explicitly request during setup (the starter-model list and model downloads).

Requires a local model server (free — e.g. Ollama from ollama.com).