OpenBiliClaw

Overview

Recommendation systems are middlemen whose weights are set by platforms — OpenBiliClaw flips this: a locally running AI agent first builds a psychological profile from your behavior, feedback, and chats, then proactively hunts content across platforms with that understanding. The project started with Bilibili (Bili + Claw) and now covers Bilibili, Xiaohongshu, Douyin, YouTube, TikTok, X, Zhihu, Reddit, Linux.do, Bangumi, V2EX, Weibo, GitHub, Instagram, and the open web. MIT licensed, latest release v0.3.232 (2026-10-26), 3598 GitHub stars, with Gitee and AtomGit mirrors for mainland China.

Data stays in local SQLite; the profile is built for you alone.

Item	Detail
Positioning	Local-first, open-source, self-evolving cross-platform discovery agent
Stack	Python 3.11+ backend + TypeScript browser extension (Manifest V3)
Data ownership	Behavior, recommendation, and chat data all stay in local SQLite
Platforms	Bilibili, Xiaohongshu, Douyin, YouTube, TikTok, X, Zhihu, Reddit, and more (15+)
License	MIT, with Gitee / AtomGit mirrors for mainland China

How It Works

• Understand you first: infers MBTI, cognitive style, and deep psychological needs from behavior, building a five-layer profile of events, preferences, awareness, insights, and soul
• Proactive exploration: guesses fields you might like but never touched, bridging via psychology to break filter bubbles; correct guesses become formal interests, wrong ones exit quietly
• Feedback loop: likes, dislikes, read-later, and chat feedback all reshape future recommendations, each with a friendly reason
• Optional grounding: add Parallel Search MCP or free anonymous Keenable search to inspiration_search_backends for keyword grounding; the default chain stays unchanged

Deployment Forms

• Browser extension: one-click install from the Chrome Web Store, compatible with Edge, Brave, Arc, Vivaldi, and more, plus Firefox xpi and Safari (macOS) builds; handles in-platform interaction, feedback collection, and login sessions
• Desktop backend: macOS dmg and Windows exe resident in the menu bar/tray; the slim build downloads the bge-m3 embedding model on first launch, while the -with-embedding build bundles the ~1.1GB model for offline use
• Mobile: scan a QR code to open :8420/m/ as an app; the separate OpenBiliClaw-mobile repo offers a Flutter native client (Android, iOS, Web, desktop)
• DSH plugin: installs into DeepSeek Harness with 22 Agent Bridge tools, adding a persistent fourth column (recommendations, library, chat, profile, settings)
• Remote access: Android/iOS native apps embed tsnet, and desktop can enable in-app Tailnet remote access; credentials stay in local memory only, never in config files, API echoes, or logs

Quick Start

1. Install the extension from the Chrome Web Store, or grab the zip from Latest Release for manual install
2. Download and start the desktop package (pick -with-embedding for offline or weak networks), or hand the agent-install.md doc to your AI coding assistant for deployment
3. Log in to Bilibili and other sources in the extension's browser; TikTok and Instagram need manual enabling in settings plus overseas network access
4. Open 127.0.0.1:8420/web for desktop Web, scan the QR code for mobile Web, or install the Flutter client against the same backend

Core behavior, recommendation, and chat data stay in local SQLite; the LLM uses your own API key by default, with no cloud account — profiles and configs never leave your machine.