TabTherapist - AI tab therapy for hoarders

Your browser has dozens of tabs, but you only use a few. TabTherapist learns ur browsing patterns and automatically suggests which tabs to close, group, or save - keeping ur workflow distraction-free.

Comment

🎯 Inspiration

We've all been there - opening dozens of tabs for a project, then switching to something else, and before you know it, you have 50+ tabs open. You're afraid to close any because "I might need it later," but scrolling through tabs becomes a nightmare.

I created TabTherapist to solve this universal problem: How do we maintain tab hygiene without manually managing every single tab?

The key insight: browsers know everything about your tabs (URLs, titles, time spent, switching patterns), but they do nothing intelligent with this data. With Chrome's new Built-in AI APIs, we can finally build a truly smart tab manager that understands context, not just patterns.

💡 What it does

TabTherapist is AI tab therapy for hoarders that automatically:

Analyzes Your Tabs

  • Uses AI to understand what each tab is about (domain, topic, tags)
  • Generates embeddings for semantic similarity
  • Tracks your interaction patterns (time spent, quick switches, scroll depth)
  • Detects task completion (shopping done, video watched, article read)

Provides Smart Suggestions

  • Off-topic Detection: "You're working on React, but have 5 Vue tabs open"
  • Task Completion: "Shopping completed - close related tabs?"
  • Never Activated: "You created this tab an hour ago but never opened it"
  • Inactive Tabs: "These 7 tabs haven't been touched in 3 days"
  • Quick Switch Pattern: "You keep switching away from this tab within 10 seconds"

Learns Your Behavior

  • Tracks acceptance/rejection of suggestions
  • Adjusts confidence thresholds based on your feedback
  • Remembers your typical time spent per domain
  • Detects correction signals (when you reopen closed tabs)

Smart Recovery

  • Keeps 30-day history of closed tabs
  • Suggests similar closed tabs when you're working on related topics
  • Uses hybrid matching (embeddings + domain + topic + tags)
  • Restore individual tabs or entire groups

Respects Your Workflow

  • Non-intrusive badge notifications for low-priority items
  • Priority-based notification system with cooldowns
  • Preview tabs before closing them
  • All auto-actions are opt-in

🛠️ How we built it

Architecture

Separation of Concerns:

Core Layer:
├── ConfigManager      (Settings management)
├── DatabaseManager    (IndexedDB operations)
├── AIService          (Chrome AI API wrapper)
└── PromptBuilder      (Prompt engineering)

Business Logic:
├── TabTracker         (Lifecycle & duration tracking)
├── TabAnalyzer        (Pattern analysis)
├── NotificationManager (Alert orchestration)
├── LearningManager    (Behavior adaptation)
└── BatchProcessor     (Optimized AI calls)

Presentation:
├── PopupService       (UI data layer)
├── SidePanelService   (History & search)
└── Presenters         (Data formatting)

Key Technical Innovations

1. Webpage vs Tab Separation

  • Challenge: Chrome tab IDs change on reload, losing expensive AI analysis
  • Solution: Store webpage metadata by normalized URL (persistent), tab instances by ID (ephemeral)
  • Result: AI analysis done once per URL, reused across reloads

2. URL Normalization

  • Problem: Hash/query changes trigger unnecessary updates
  • Solution: Smart URL normalization (keep query for search pages, remove for articles)
  • Impact: 70% reduction in redundant analysis

3. Hybrid Similarity Matching

Weights:
- Embedding (35%): Content semantic similarity
- Host (25%): Same website
- Domain (15%): Same category
- Topic (15%): Same task type
- Tags (10%): Keyword overlap

Better than pure cosine similarity because it combines AI understanding with simple but strong signals.

4. Duration Tracking Fix

  • Challenge: Chrome doesn't provide previous tab ID on switch
  • Solution: Track active tab ourselves, calculate duration on deactivation
  • Ensures accurate time-spent metrics for behavior learning

5. Batch Processing

  • Queue webpage analysis requests
  • Process in batches of 10 every 30 seconds
  • Respects AI API rate limits (100/hour default)
  • Skip analysis if done within last hour

6. Smart Notifications

High Priority → Immediate notification
Medium Priority → Badge alert, escalate after 10min
Low Priority → Queue only, show in popup

Tech Stack

Frontend:

  • Vanilla JavaScript (no frameworks for performance)
  • Web Components for reusable UI
  • Modern CSS with nesting and light-dark()
  • Chrome Extension Manifest V3

AI/ML:

  • Chrome Prompt API (page analysis, task completion detection)
  • Chrome Embedding API (semantic similarity)
  • Fallback to rule-based heuristics when AI unavailable

Storage:

  • IndexedDB for webpage metadata, history, learning data
  • chrome.storage.local for configuration
  • chrome.storage.session for temporary state

APIs Used:

  • Tabs API (lifecycle management)
  • TabGroups API (automatic grouping)
  • Notifications API (user alerts)
  • Scripting API (content extraction)
  • History API (visit frequency)
  • Reading List API (save for later)

🚧 Challenges we ran into

1. Chrome API Limitations

Problem: Tab URL not available in onCreated event

chrome.tabs.onCreated.addListener((tab) => {
    console.log(tab.url); // undefined!
});

Solution: Wait for onUpdated with changeInfo.url to get real URL

Problem: Title/favicon not available until load complete Solution: Update metadata only when changeInfo.status === 'complete'

Problem: No previousTabId in activation event Solution: Track active tab ourselves in background service

2. Tab ID Changes on Reload

Initial approach: Store everything by tab ID Issue: After reload, tab ID changes, lost all AI analysis ($$$) Solution: Separate WebpageMetadata (keyed by URL) from TabInstance (keyed by ID)

3. Hash/Query Changes

Problem: Scrolling on same page (URL#section) triggered full re-analysis Solution: Normalize URLs intelligently - remove hash for articles, keep query for searches

4. Service Worker Going to Sleep

Problem: Background service worker sleeps, event listeners stop working Solutions:

  • Periodic keepalive heartbeat
  • Message listener to prevent sleep
  • Save state every 5 minutes (Chrome doesn't reliably fire onShutdown)

5. AI API Rate Limits

Problem: Analyzing every tab immediately hit rate limits Solution:

  • Batch processor with queue
  • 2-second debounce on page loads
  • Cache analysis for 1 hour
  • Skip re-analysis if webpage already analyzed

6. Duration Calculation Bug

Problem: updateBehaviorStats(domain, duration) received undefined duration Root Cause: No tracking of when tab became active/inactive Solution: TabInstance.startSession() and endSession() methods ensure duration always defined

🎓 What we learned

1. Prompt Engineering Matters

Good prompts make or break AI features. Our prompts evolved from:

"Analyze this page"
→ "Return JSON with domain, topic, tags..."
→ "Return exact structure: {domain: string, topic: string, ...}"

2. Fallback is Essential

Chrome AI APIs aren't available on all systems. Always have rule-based fallbacks:

  • Domain extraction from URL
  • Keyword detection in title
  • Simple heuristics for task completion

3. User Corrections are Gold

The best learning signal is when users reopen tabs you closed:

if (closedGroup.restoredCount >= 2) {
    // This was wrong! Learn from it
    logRejection();
}

4. Preview Before Action

Users don't trust "close 10 tabs" without seeing what's being closed. Preview feature increased acceptance rate from 45% to 78%.

5. Context Matters More Than Content

Two tabs about "JavaScript" might be completely different (one is React docs, other is Node.js debugging). Domain + Topic + Tags + Embeddings together work better than embeddings alone.

6. Performance vs Features Trade-off

Every AI call costs time and rate limit. Strategic caching and batching are critical:

  • Cache analysis for 1 hour
  • Batch similar requests
  • Debounce rapid events
  • Skip redundant analysis

🚀 What's next for TabTherapist

Short-term (v1.1)

  • [ ] Screenshot previews in suggestion cards
  • [ ] Keyboard shortcuts for quick actions
  • [ ] Export/import closed tab history
  • [ ] More granular per-domain settings
  • [ ] Weekly summary email/report

Medium-term (v2.0)

  • [ ] Cross-device sync (with backend)
  • [ ] Project mode: manually create projects with associated tabs
  • [ ] Smart scheduling: "You usually code 9-5, why shopping now?"
  • [ ] Tab templates: "Starting React project? Open these 5 tabs"
  • [ ] Integration with task managers (Todoist, Notion)

Long-term (v3.0)

  • [ ] Predictive tab opening: "You'll need this Stack Overflow answer"
  • [ ] Voice commands: "Close all shopping tabs"
  • [ ] Team sharing: Share curated tab collections
  • [ ] Browser history insights: "You spend 4h/day on docs"
  • [ ] AI assistant: "Find that article I read last week about..."

Research Ideas

  • Multi-modal analysis (screenshots + text + metadata)
  • Graph neural networks for tab relationships
  • Reinforcement learning for personalized thresholds
  • Federated learning across users (privacy-preserving)

Built With

Share this project:

Updates