OpenBiliClaw

OpenBiliClaw is a local, self-hosted content discovery agent. Instead of a platform matching content against your click history, it first builds a psychological profile of you and then goes looking for content you might like across Bilibili, Xiaohongshu, Douyin, YouTube, X, Zhihu, Reddit, Linux.do, Bangumi, V2EX, Weibo, GitHub, Instagram (experimental), and the open web. The project started as a Bilibili-only tool and became a cross-platform agent from v0.3.0.

Three engines do the work. The Soul Engine turns platform activity, feedback, and chat into a five-layer profile (events, preferences, awareness, insights, soul) that infers MBTI, cognitive style, and deeper needs. The Discovery Engine turns that profile into search, trending, related-content, and exploration tasks per source, and a shared evaluator scores every candidate against the profile and recent negative feedback. The Recommendation Engine picks items, explains in plain language why you would like them, and feeds likes, dislikes, and chat feedback into the next round. Interest probes guess fields you have never touched (a mechanical-watch fan might like architecture) and promote a guess to a real interest only when you confirm it.

Everything stays on your machine: behavior, recommendations, and chat live in a local SQLite database, and the profile, credentials, and caches in local files. You bring your own LLM API key. Built-in providers are Gemini, DeepSeek, OpenAI, Claude, OpenRouter, OrcaRouter, Requesty, API Route, Cheaper Inference, and Ollama, plus any OpenAI-compatible service, and OpenAI can experimentally reuse the Codex CLI ChatGPT OAuth credential. Several instances of one provider type can be configured with global and per-module failover chains. Embeddings can run locally through Ollama with bge-m3.

The entry points are a browser extension (Chrome, Edge, Brave, Arc, Firefox, Safari) that shows recommendations in a side panel, collects behavior signals, and holds the platform login sessions, a desktop web UI, a mobile web UI, a Flutter client for Android, iOS, web, and desktop in a separate repository, and a DeepSeek Harness client plugin that adds a fourth panel and 22 Agent Bridge tools. The backend runs as a desktop app, a Docker Compose stack, or from source, and an opt-in embedded tailnet lets the native mobile apps reach it from outside the LAN without a system Tailscale install.

A bundled workspace skill and a versioned Agent Bridge let any skill-capable agent host (OpenClaw, Hermes, WorkBuddy, Claude Code, Codex CLI, Cursor) call the local instance: receive proactive recommendations over WebSocket, answer interest probes, read the profile, pull recommendations by platform, and write feedback back.