Why Trove AI?

You save 1000 articles. You re-read 5.

The problem isn't that you have too much — it's that your tools treat "save" and "read" as the same action. WeChat 收藏 buries them. 收藏 ≠ reading. That gap is the entire problem.

Pocket shut down in 2024. Omnivore shut down too. Their users' carefully curated libraries vanished overnight.

Trove AI is a self-hostable, AI-powered second brain that turns "save for later" back into "actually read & remember." Built first-class for the Chinese internet (WeChat 公众号 / 知乎 / 抖音 / 小红书 / B 站 / 头条 / 掘金 / CSDN), with WeChat Bot ingress, automatic knowledge graph, and one-way Obsidian sync as built-in features.

It's yours to host. It's yours to keep.

Highlights

Screenshots

⚠️ Add screenshots in docs/screenshots/ — see open issues for placeholders.

Who is it for?

• Product managers and researchers drowning in saved-but-never-read articles
• Engineers and lifelong learners who want their weekly tabs to compound into knowledge
• Privacy-conscious users who don't want their reading habits living in some startup's database
• People burned by Pocket / Omnivore shutdowns wanting data sovereignty
• Content curators building structured personal knowledge bases
• Self-hosters who run their own infrastructure for fun and survival
• Cross-device readers who switch between phone (commute) → iPad (couch) → laptop (desk) and want all three to feel native

Compared to alternatives

Trove AI is the only option that combines: deep Chinese platform support · WeChat Bot · auto knowledge graph · Obsidian sync · self-host · full-device responsive UI. If any one of those is critical to you, the alternatives don't cover it.

Quick Start (5 minutes)

Prerequisites

• Docker ≥ 24.0 with Compose v2 (docker compose ..., not docker-compose)
• ~ 4 GB RAM free
• ~ 5 GB disk

That's it. No Python or Node required on the host.

Steps

# 1. Clone
git clone https://github.com/weaiw/trove-ai.git
cd trove-ai

# 2. Configure secrets
cp .env.example .env
# Edit .env, at minimum:
#   POSTGRES_PASSWORD=$(openssl rand -base64 24)
#   SECRET_KEY=$(openssl rand -base64 48)

# 3. (Optional) Pre-fill LLM keys (or skip — configure via web UI later)
cp backend/app/config_store.example.json backend/app/config_store.json

# 4. Run
docker compose up -d

# 5. Open
open http://localhost

First-time setup creates an admin user. The credentials appear in backend logs:

docker compose logs backend | grep -i admin

Full self-host guide with troubleshooting: docs/SELF_HOST.md.

Cloud deployment

Any Docker-capable VM works. Battle-tested on:

• 腾讯云 Lighthouse / CVM (recommended for China users)
• AWS EC2 t3.medium
• DigitalOcean 4GB droplet
• Hetzner CX22

Bring your own reverse proxy (Caddy / Traefik / Nginx) for HTTPS, or use Cloudflare Tunnel.

Architecture

        ┌──────────────────────────────────────────────────────┐
        │       Any device — PC · iPad · mobile · browser       │
        │   • Web app   • WeChat Bot   • Obsidian plugin        │
        └─────────────────────────┬────────────────────────────┘
                                  │
                          ┌───────▼───────┐
                          │  Nginx :80    │
                          └───┬────────┬──┘
                              │        │
                  ┌───────────▼──┐  ┌──▼────────────┐
                  │  Frontend    │  │  Backend       │
                  │  (Next.js 14)│  │  (FastAPI)     │
                  │  Responsive  │  │  async         │
                  └──────────────┘  └───┬────────────┘
                                        │
                  ┌─────────────────────┼─────────────────────┐
                  │                     │                     │
        ┌─────────▼──────────┐ ┌────────▼───────┐  ┌──────────▼────────┐
        │  PostgreSQL 16     │ │  Redis 7       │  │ External APIs     │
        │  + pgvector        │ │  (cache)       │  │ LLM + embedding   │
        │  • articles        │ │                │  │ • DeepSeek        │
        │  • embeddings 1024d│ │                │  │ • 讯飞 / OpenAI   │
        │  • knowledge_edges │ │                │  │ • SiliconFlow     │
        │  • users + tokens  │ │                │  │ • any compatible  │
        └────────────────────┘ └────────────────┘  └───────────────────┘

Tech stack

Configuration

Everything user-facing is configurable via the web UI:

Settings page → AI 对话模型 / 嵌入模型 / 缓存

Environment variables

See .env.example for the complete template with comments.

API endpoints (high-level)

Full OpenAPI spec at http://localhost/api/docs once running.

Obsidian Sync — companion plugin

Plugin repo: weaiw/trove-sync-obsidian (MIT)

One-shot snapshot to your local vault. Never overwrites your edits. Your data survives any future shutdown.

Setup:

1. Web app → Personal Settings → Obsidian Backup → Generate Sync Token
2. Download plugin from Releases
3. Drop into <your-vault>/.obsidian/plugins/trove-sync/
4. In Obsidian → Community plugins → enable Trove AI Sync
5. Paste token + server URL → click Sync Now

The plugin auto-detects already-synced articles via dual-OR (sync_state.json ∪ frontmatter scan), so it's safe to lose either side of the index.

Documentation

• docs/SELF_HOST.md — Full self-host guide with troubleshooting
• CONTRIBUTING.md — How to contribute
• API docs at /api/docs (auto-generated from FastAPI)

Roadmap

v1.1 — current

• ✅ WeChat Channels (视频号) capture
• ✅ Smart generic extraction (trafilatura → headless render → BeautifulSoup)
• ✅ Article-scoped Q&A (📄 this-article / 📚 whole-library toggle)

v1.0

• ✅ Multi-platform capture (8+ sources)
• ✅ AI processing pipeline (summary / key-points / tags / embedding / mind-map)
• ✅ RAG Q&A + semantic search
• ✅ Auto knowledge graph + learning paths
• ✅ WeChat Bot ingress
• ✅ Obsidian sync plugin
• ✅ Multi-tenant + revocable sync tokens
• ✅ Self-host via Docker
• ✅ Responsive UI for PC / pad / mobile

v1.1

• 🔜 Browser extension (one-click clip from any tab)
• 🔜 Image local download (offline-safe backup)
• 🔜 Pocket / Omnivore import
• 🔜 Better article deduplication
• 🔜 PWA support for "add to home screen" on mobile

v1.2

• More LLM providers (Claude, Gemini, Doubao native)
• Per-user theme & language preferences
• Bulk re-process articles with new AI prompts
• Article version history

v2 — research

• Obsidian community marketplace submission
• Multi-vault Obsidian sync
• Notion / Logseq / Reflect export
• Audio podcast generation from saved articles
• Daily / weekly digest emails

FAQ

Contributing

See CONTRIBUTING.md.

Especially welcome:

• New platform parsers (parser_service.py)
• Translations (English, 日本語, others)
• UI polish and accessibility
• Bug reports with reproduction
• Comparison tests with other LLM providers

Acknowledgements

• Built with Hermes AI coding agent + DeepSeek as the LLM brain — over 2.7 billion tokens spent in vibe-coding, 0 lines of human-written code
• Backend: FastAPI · SQLAlchemy · pgvector · Playwright · curl_cffi
• Frontend: Next.js · Tailwind · lucide-react · react-flow
• Embedding: BAAI/bge-m3 · fastembed
• Inspired by Pocket, Omnivore, Readwise — and frustrated by the first two shutting down

License

Core: AGPL-3.0. Obsidian plugin: MIT.

For commercial closed-source SaaS deployment, contact the maintainer for a separate commercial license.