An open-source tool called Siftly launched March 4, 2026, transforming Twitter/X bookmarks into a searchable, categorized knowledge base that runs entirely on users' local machines. Created by developer @viperr, the project gained 1,230 GitHub stars and 96 forks within five days, demonstrating strong community appetite for privacy-first bookmark management.
Four-Stage AI Pipeline Processes Bookmarks Without Cloud Services
Siftly runs a local AI pipeline using Claude (via Anthropic API or CLI) that processes bookmarks through four distinct stages. First, entity extraction mines hashtags, URLs, mentions, and over 100 recognized tool names from stored tweet JSON without requiring API calls. Second, vision analysis examines images, GIFs, and video thumbnails for OCR text, objects, scenes, and mood, generating 30-40 visual tags per image.
The third stage performs semantic tagging, producing 25-35 searchable tags per bookmark by combining tweet text and image context. Finally, the categorization stage assigns one to three categories per bookmark with confidence scores. All data is stored in a local SQLite database with FTS5 full-text search indexing.
The tool requires only Node.js 18+ and an AI API to function—no browser extensions, cloud services, or subscriptions. New Anthropic accounts include $5 free credit, sufficient for processing thousands of bookmarks at Haiku pricing.
Interactive Mindmap and Natural Language Search Replace Manual Organization
Siftly's standout feature is an interactive force-directed graph visualization that organizes bookmarks by category. Users can expand and collapse groups, click nodes to open original tweets, and navigate their bookmark collection spatially. The interface also offers traditional grid and list views with filtering by category, media type, and date.
Natural language search lets users query across tweet text, image OCR, visual tags, and semantic metadata with phrases like "funny meme about crypto crashing." A command palette accessible via Cmd+K (Mac) or Ctrl+K provides global search functionality. Results can be exported as CSV, JSON, or ZIP archives with media files and manifests.
Customizable Categories and Privacy-First Architecture
Users can create custom categories with names, colors, and descriptions. The description field directly influences AI categorization—more specific descriptions yield more accurate classifications. Default categories include Funny Memes, AI Resources, Dev Tools, Design, Finance & Crypto, Productivity, News, and General.
The privacy model keeps all data in a local SQLite file. External calls go only to the configured AI provider (sending tweet text and image data). The tool includes no telemetry, tracking, or account requirements. Import is handled via bookmarklet (drag to bookmark bar, visit X bookmarks, auto-scroll to capture all) or console script, with automatic deduplication on re-import.
Technical Stack Built on Next.js and Prisma
The project uses Next.js 16, TypeScript, Prisma ORM, SQLite with FTS5 full-text search, Tailwind CSS v4, and the Anthropic SDK. It's released under MIT license with 44 commits, 9 active pull requests, and includes CLAUDE.md documentation for Claude Code IDE integration. The local-first architecture with optional Claude CLI detection eliminates API key friction for existing users, while the four-stage pipeline is incremental—interruptions don't require restarting from scratch.
Key Takeaways
- Siftly launched March 4, 2026, gaining 1,230 GitHub stars and 96 forks within five days
- Four-stage AI pipeline processes bookmarks locally: entity extraction, vision analysis, semantic tagging, and categorization
- Interactive mindmap visualization organizes bookmarks as force-directed graph with expandable category nodes
- All data stored in local SQLite database with no cloud services, telemetry, or subscriptions required
- Natural language search queries across tweet text, image OCR, visual tags, and semantic metadata using Claude reranking