Thilko

About

Highlight any web article, ask Dabbis (an ai agent) about it, remember every note. A reading companion for the people who actually read.

### Detailed description
```
Thilko is a reading companion for the web.

Highlight any sentence on any article or PDF. Ask Dabbis — your small scholarly AI assistant — to explain it, push back on it, or sit with it. Save notes that survive page reloads. Every highlight, comment, and conversation is stored in a personal memory backend so you can come back to them, search them, and bundle them up to share with another LLM in one click.

Built for the way people actually read: select a sentence that struck you, ask a question, move on. The window stays out of the way of the text.

What you get:
• A floating selection toolbar that appears when you highlight text. Click Comment to save a note, click Ask Dabbis-AI to start a conversation.
• Highlights that survive page reloads, with smart re-anchoring for articles that get edited.
• A draggable, frosted-glass popover that anchors to the right edge of the viewport — out of the way of the text you're reading.
• Streaming AI replies with full markdown, including code blocks.
• PDF support: any *.pdf link opens in Thilko's viewer, with the same toolbar and the same memory.
• A side library with semantic search across every highlight, note, and AI conversation you've ever made.
• "Open with Claude": bundle a highlight, an article, or a whole topic into a markdown summary and paste it into claude.com with one click.
• Lazy-by-default persistence: highlights only get saved when you actually attach a note or a question. Your memory stays clean.
• Local-first telemetry: a 100-entry ring buffer of recent operations, viewable in Options. No remote logging.

How it works:
Thilko talks to a proxy server you configure (either run your own, or use one a friend hosts). The proxy holds your secrets and forwards inference requests to OpenAI's Codex API and storage requests to Supermemory. The extension never holds an API key. Your data is scoped to a "slot" namespace you choose.

The code is open source under MIT — see the GitHub repo linked in the support URL.