whiteguo233/OpenBiliClaw

OpenBiliClaw – A Personal‑AI Recommendation Agent

What it is

• An open‑source, locally‑run AI agent that builds a psychological profile of you from the way you interact with a variety of content platforms (Bilibili, Xiaohongshu, TikTok, YouTube, X/Twitter, Zhihu, Reddit, Linux.do, Bangumi, V2EX, Weibo, GitHub and the open web).
• Using that profile, it actively searches those platforms for items it thinks you’ll like, even if you’ve never seen anything similar before.
• All data (behavior logs, profile, recommendations, chat history) stay on your machine in a SQLite database – no cloud service collects your preferences.

Key components

Component	Role
Browser extension	Injects a sidebar into supported sites, reads your browsing/session cookies, captures feedback (like/dislike, “later”, chat) and forwards the signals to the backend. Works on Chrome‑based browsers, Edge, Brave, Vivaldi, Opera and Safari (via a signed DMG).
Backend server	Written in Go/Python, runs on http://127.0.0.1:8420. It stores data, runs the LLM‑driven reasoning, generates embeddings (default bge‑m3), and serves the web UI. Two install‑options: a lightweight package that downloads the embedding model on first start, or a “‑with‑embedding” bundle that ships the model (~1 GB) for fully offline use.
Web UI (desktop & mobile)	Accessible at /web (large‑screen) or /m/ (mobile‑optimized). Shows recommendation cards, explains why each item was suggested, visualises your “soul portrait” (MBTI, core traits, deeper needs) and lets you chat with the agent to refine its understanding.
Flutter mobile client (separate repo)	Native Android/iOS/Web/desktop app that connects to the same local backend, offering the same UI without needing a browser.
DeepSeek Harness (DSH) plugin	Optional integration that embeds OpenBiliClaw inside the DeepSeek Harness UI, exposing 22 “Agent Bridge” tools for cross‑platform content discovery.

How it works (high‑level)

1. Signal collection – The extension watches your activity on the chosen platforms and records explicit feedback (thumbs‑up, “not interested”, chat messages).
2. Profile building – A chain of LLM prompts turns raw signals into a five‑layer “psychological portrait”: events → preferences → awareness → insight → soul. It also infers MBTI, cognitive style and deeper motivations.
3. Cross‑platform search – Using the portrait, the backend queries the APIs of the supported platforms (or scrapes public pages) to find content that matches inferred interests, even if no other user has taken that path.
4. Recommendation & explanation – Results are ranked, embedded with bge‑m3, and presented with a natural‑language explanation (“Because you enjoy hands‑on mechanical design, you might like this video about watchmaking”).
5. Continuous learning – Every interaction updates the portrait, so the system gradually personalises further.

Why it matters

• Privacy‑first: No server stores your data; you control the LLM API key (OpenAI, Claude, etc.) or can run a local LLM.
• Breaks filter bubbles: Instead of relying on platform‑centric collaborative filtering, it generates novel suggestions based on a deeper model of you.
• Cross‑platform unification: Your interests are no longer siloed per service; the agent aggregates signals across dozens of sites.

Getting started (quick‑start)

1. Install the browser extension from the Chrome Web Store (or download the zip from the latest GitHub release).
2. Download the appropriate desktop installer (.dmg for macOS, .exe for Windows) from the same release page. Choose the “‑with‑embedding” bundle for fully offline use.
3. Log into at least one supported platform (Bilibili is the default) via the extension.
4. Open http://127.0.0.1:8420/web (or scan the QR code for the mobile view) to see recommendations, your profile, and a chat box.

Advanced options

• Tailnet (Tailscale) remote access – Enable the built‑in helper to join a private Tailnet, allowing your phone app to reach the backend across the internet without installing system‑wide Tailscale.
• Docker / HTTPS deployment – A docker-compose.https.yml template with Caddy can expose the API securely behind a custom domain.
• Custom LLM / embeddings – Swap the default LLM by editing config.toml; you can also replace the bge‑m3 embedding model.
• DSH integration – Install the DSH plugin (dsh-openbiliclaw) to embed the agent inside DeepSeek Harness and expose its tools to other agents.

Community & support

• QQ group and Discord server links are provided for user help and feature discussion.
• The project maintains CI builds, releases on GitHub, Gitee mirror, and a Chinese‑focused download mirror on 123 Pan.

---

Bottom line: OpenBiliClaw is a genuine AI‑driven, privacy‑preserving recommendation system that runs entirely on your device, builds a nuanced personal model, and fetches content from a wide range of Chinese and international platforms. It fits squarely within the frontier of personal AI agents and cross‑platform recommendation technology.