LLM-powered tab grouping for AI-first browsers.
GroupThink is a Chrome extension that uses Claude (or local Ollama models) to organize your open tabs by what you're doing, not where you are. A YouTube video about React goes in the "Frontend Dev" group, not "YouTube." 64 tabs become 6 meaningful workspaces.
- Press Alt+Shift+G (or Cmd+Shift+K on Mac) to open GroupThink
- The extension reads every open tab's title and URL
- An LLM classifies tabs by topic and intent — grouping semantically, never by domain
- Results render as an interactive zoomable d3 treemap — tiles sized by tab count, color-coded by category
- Adjust the specificity slider (1-10) to control granularity: broad workspace-level groups at 1, fine-grained subgroups at 10
- Refine via chat: type "merge News and Media" or "split React into Components and State" and the LLM restructures on the fly
- Semantic grouping — groups by topic/intent, not by website. "Billing" instead of "AWS + Stripe + QuickBooks"
- Zoomable treemap — click into a group to see subgroups and individual tabs. Breadcrumb navigation back out
- Tab previews — hover any tab tile to see a thumbnail screenshot and metadata
- Natural language refinement — chat bar at the bottom for multi-turn restructuring
- Specificity control — slider from broad (3-5 groups) to fine-grained (nested subgroups for 4+ tabs)
- Dual LLM support — Anthropic Claude (cloud) or Ollama (local/free: qwen2.5, llama3.1, mistral, gemma2)
- Browser context enrichment — optionally uses visit history, bookmarks, and recent sessions for smarter classification
- Cache-first loading — shows cached grouping instantly; regroups only on explicit refresh
- Keyboard shortcut — Alt+Shift+G / Cmd+Shift+K to open from anywhere
popup (browser action)
└── opens app/index.html in new tab
app (React)
├── TreemapView zoomable d3 treemap visualization
│ ├── TreemapNode recursive tile renderer (groups, children, tabs)
│ ├── TabPreview hover popover with screenshot thumbnail
│ └── TreemapBreadcrumb hierarchy navigation
├── SpecificitySlider 1–10 granularity control
├── ChatBar natural language refinement input
└── TabChaos loading animation (tabs fly into groups)
background (service worker)
├── ai.ts LLM client (Anthropic + Ollama)
├── grouping.ts post-LLM enforcement, keyword clustering, singleton merging
├── prompts.ts system prompt + per-call prompt builders
├── browser-context.ts visit history, bookmarks, top sites
└── tabs.ts Chrome tabs API wrappers
All LLM calls happen in the background service worker. The app communicates via chrome.runtime.sendMessage() — ai.ts is never imported in UI code.
| Layer | Technology |
|---|---|
| UI | React 18, d3-hierarchy (treemap layout) |
| Build | Vite 5, @crxjs/vite-plugin (Chrome MV3) |
| AI | Anthropic SDK (Claude), Ollama (local models) |
| Validation | Zod |
| Language | TypeScript |
| Linting | Biome |
| Testing | Vitest |
Prerequisites: Node >= 20, pnpm 8+, Chrome
pnpm install
# Configure LLM provider
cp env.json.example env.json
# Add your Anthropic API key, or configure Ollama URL
pnpm buildLoad the extension:
- Navigate to
chrome://extensions - Enable Developer mode
- Click Load unpacked and select the
dist/directory
After code changes: pnpm build, then click the reload button on the extension card.
Open the extension options page to configure:
- LLM Provider — Anthropic (cloud) or Ollama (local/free)
- Model — Claude Sonnet (default), Claude Opus, Claude Haiku, or local models (qwen2.5, llama3.1, mistral, gemma2)
- Context enrichment — Off (title + URL only), Basic (+ visit frequency), Full (+ bookmarks, sessions)
- Theme — Light, Dark, or Auto (system preference)
MIT