Skip to main content
Back

OpenBiliClaw Builds a Private AI Agent That Hunts Content Across 12 Platforms

OpenBiliClaw is a local-first agent that builds a psychological profile of you and hunts content across 12 platforms.

Repo, Full Access
  • OpenBiliClaw is a local-first open-source AI agent that recommends content across 12 platforms
  • Builds a five-layer soul profile including MBTI, cognitive style, and deep psychological needs
  • Covers Bilibili, Xiaohongshu, Douyin, YouTube, X, Zhihu, Reddit, Linux.do, Bangumi, V2EX, Weibo, and the open web
  • Data stays in local SQLite; supports Ollama plus bge-m3 for zero-API-key setup
  • Ships a Chrome extension, desktop installer, Docker Compose, and a Flutter mobile client
  • Includes Agent Bridge for Claude Code, Codex CLI, Cursor, and a DeepSeek Harness plugin

Recommendation algorithms optimize for what the platform wants, not what you want. Each platform is also a walled garden: your Bilibili history means nothing to Reddit, and your YouTube taste is invisible to Xiaohongshu. OpenBiliClaw flips the setup by running a personal AI agent locally on your machine that studies you first, then goes hunting across a dozen platforms on your behalf.

The project just crossed 2,600 stars on GitHub and describes itself as a general-purpose personalized content discovery agent that runs on your machine and understands only you. It began as a Bilibili-only tool (hence the name: Bili + Claw) and has grown into a cross-platform recommender covering Bilibili, Xiaohongshu, Douyin, YouTube, X, Zhihu, Reddit, Linux.do, Bangumi, V2EX, Weibo, and the open web.

OpenBiliClaw demo showing local backend, profile generation, and recommendation explanations

A five-layer read on who you are

The core idea is that the agent understands the person before it picks content. Instead of matching tags to a click history, it builds what the authors call a five-layer soul profile going from events to preferences, awareness, insights, and finally a soul-level sketch, inferring MBTI, cognitive style, and deep psychological needs. Recommendations arrive with a friend-like explanation of why you might like something, rather than the usual mystery-meat feed.

On top of that profile, an interest probe runs in the background. It proactively guesses domains you might enjoy but have never touched, promoting confirmed guesses into formal interests and quietly dropping the wrong ones. A collaborative filter would never suggest architecture videos to a mechanical-watch enthusiast because nobody has walked that co-view path yet. This agent will, because it is reasoning about you as a person.