A local agent that builds your soul profile to hunt across platforms

What it does

OpenBiliClaw is a locally-run Python agent that tries to understand you—through behavior, feedback, and chat—then actively searches Bilibili, Xiaohongshu, Douyin, YouTube, and the broader web for content it thinks you will like. Everything lives in a local SQLite database and runs against your own LLM API keys, so your psychological portrait never leaves your machine. It delivers recommendations via a browser extension and responsive web interfaces for desktop and mobile.

The interesting bit

Instead of collaborative filtering or simple tag matching, the project constructs a five-layer “soul portrait”—inferring MBTI, cognitive style, and deep psychological needs—to actively guess interests you have never expressed, then ventures into those areas and quietly retreats if wrong. It is an attempt to break the filter bubble by treating recommendation as cross-platform psychoanalysis rather than clickstream optimization.

Key highlights

• Cross-platform discovery from a single local profile, pulling from Bilibili, Xiaohongshu, Douyin, YouTube, and general web sources.
• Recommendations come with friend-style explanations (“why you will like this”) rather than opaque “guess you like” labels.
• Five-layer psychological model (event → preference → awareness → insight → soul) built from usage, feedback, and dialogue.
• Fully local data storage in SQLite; LLM inference uses your own keys, with an experimental option to reuse local Codex CLI OAuth credentials.
• Browser extension handles platform login sessions and in-page interaction, while desktop and mobile web UIs serve the main recommendation feed.

Caveats

• You must supply your own LLM and embedding API keys; the backend validates both before initializing and stops if either is unreachable.
• The browser extension currently relies on manual zip installation in developer mode for most setups, and Firefox temporary loading expires on restart with AMO signing still planned.
• Platform support is uneven: the README lists full discovery for some sources (e.g., Douyin search/hot/feed) and initialization signals for others, with more platforms promised.

Verdict

Worth a look if you are tired of platform silos and want an audacious, privacy-first experiment in personalized discovery. Skip it if you expect a polished, zero-config consumer app—this is a self-hosted agent that requires patience and API keys.