Most platform recommender systems optimize platform-side goals and keep interests siloed; that makes cross-platform interests fragmented and surprises rare. OpenBiliClaw flips that model: a local-first AI agent that builds a persistent psychological portrait of a single user and proactively searches multiple sources for content it predicts you'd like, while keeping data on your machine.
What Sets It Apart
- Cross-platform, local-first discovery: unifies signals from Bilibili, Xiaohongshu, Douyin, YouTube, X, Zhihu, Reddit, Bangumi, V2EX, Linux.do, Weibo and arbitrary Web pages, but keeps core data (profiles, recommendations, feedback) in a local SQLite store.
- Agent-driven exploration rather than passive matching: uses an LLM + embedding pipeline and a "five-layer soul" portrait (events → preferences → awareness → insight → soul) to actively probe for latent interests and surface low-overlap "surprise" recommendations.
- Integrations and UX: browser extension collects consented signals and acts as a task bridge; a local backend exposes Web UI (desktop/mobile), optional desktop installers bundle local embeddings (bge-m3) and Ollama support, and a Flutter mobile client is available in a related repo.
- Transparent, trainable loop: likes/dislikes and Socratic-style chats are written back locally to refine the model's hypotheses and future discovery strategies.
How it works (high level)
OpenBiliClaw ingests platform signals via an installed browser extension and per-source adapters, normalizes candidates into a shared evaluation pipeline, scores them against the local portrait using embeddings and LLM judgments, and serves mixed-platform recommendation pools with per-topic quotas and de-duplication. Embedding can be local (Ollama + bge-m3) or remote; LLM provider is configurable. The system is designed for editability: you can swap models, edit portraits, or integrate the repo into agent hosts via an Agent Bridge.
Who it's for & trade-offs
Great fit if you want a private, researchable content-recommendation agent on your machine that links interests across multiple platforms and accepts conversational tuning. It suits technically comfortable users who can install a browser extension and run a local backend or desktop package.
Look elsewhere if you need a fully managed cloud service, enterprise-scale multi-user deployment, or a turnkey consumer app without any local installation: the project is local-first, experimental in places (pre-release desktop installers and platform-specific adapters), and some sources require you to be logged into the platform in the same browser for richer initialization signals.
Privacy note: by default the system keeps recommendation data, portraits and caches on-disk; sending content to cloud LLM/embedding providers depends on your configuration.
