Your knowledge, structured by AI — capture anything, let agents compile the wiki.
Live Demo · Quick Start · The Story · 中文
git clone https://github.com/Yrzhe/pagefly.git && cd pagefly
python -m src.cli setup # interactive: email, password, API keys
docker compose up -d # → http://localhostThe idea: a knowledge flywheel that grows structured knowledge from the stream of daily life.
PageFly is a self-hosted, private knowledge platform — a structured, automated, API-ready system that turns raw information into compiled knowledge.
You send it raw material (PDFs, markdown, images, voice memos, URLs, Telegram messages), and it:
- Captures — ingests into a structured raw layer with metadata
- Distills — AI classifies, scores relevance, tags temporal type, extracts key claims
- Compiles — agents write and maintain wiki articles (concept pages, summaries, connection maps)
- Serves — REST API, Telegram bot, web frontend, Obsidian-compatible markdown output
You never write the wiki manually — the LLM owns it.
| Feature | Description |
|---|---|
| Multi-format Ingestion | PDF, DOCX, images (OCR), voice (transcription), URLs, plain text |
| AI Distillation | Auto-classification, relevance scoring, temporal tagging, key claim extraction |
| Wiki Compilation | Agents write concept pages, summaries, connection maps with update-first governance |
| Workspace Editor | Tiptap rich-text editor with images, tables, code blocks — draft → finish → ingest to knowledge base |
| Telegram Bot | Send anything via Telegram — text, photos, voice, documents. Inline approval flow |
| Daily Roam | Random knowledge resurfacing with staleness weighting and dedup |
| REST API | ~60 endpoints with multi-token auth (JWT + API tokens + master token) |
| Scheduled Agents | Cron-driven review, compilation, linking, trend analysis, daily roam |
| Knowledge Graph | Interactive force-directed graph of documents and wiki articles |
| Obsidian-Compatible | Wiki output as flat .md files with YAML frontmatter — drop into any PKM tool |
┌─────────────────────────────────────────────────────────────┐
│ Channels │
│ Telegram Bot · REST API · Web Frontend · Scheduler │
└─────────────┬───────────────────────────────────┬───────────┘
│ │
┌──────────▼──────────┐ ┌────────────▼───────────┐
│ Ingest Pipeline │ │ Agent System │
│ │ │ │
│ PDF · DOCX · Image │ │ Compiler (wiki write) │
│ Voice · URL · Text │ │ Linker (connections) │
└──────────┬──────────┘ │ Trend (insights) │
│ │ Review (lint + audit) │
┌──────────▼──────────┐ │ Query (search + chat) │
│ Governance │ └────────────┬───────────┘
│ │ │
│ Classifier (AI) │ ┌────────────▼───────────┐
│ Organizer │ │ Storage │
│ Integrity Checker │ │ │
└─────────────────────┘ │ SQLite (metadata) │
│ Filesystem (documents) │
│ Wiki (markdown) │
└─────────────────────────┘
PageFly was inspired by Andrej Karpathy's LLMWiki — the idea that structured knowledge compilation can be automated.
I saw the tweet and thought: what if we took this further? Not just a wiki, but a complete capture-to-serve pipeline with ingestion, distillation, governance, and API access — plus a workspace where you write alongside AI agents.
Prerequisites: Docker + Docker Compose, an Anthropic API key.
git clone https://github.com/Yrzhe/pagefly.git
cd pagefly
python -m src.cli setup # interactive: email, password, API keys, demo data
docker compose up -dThe setup command generates a valid config.json with a hashed password and — if you accept — seeds a working demo knowledge base so you can see the system in action before adding your own documents.
Already configured? Skip setup and just docker compose up -d.
No config.json needed. Export three env vars and launch:
export ANTHROPIC_API_KEY=sk-ant-...
export PAGEFLY_EMAIL=you@example.com
export PAGEFLY_PASSWORD=your-password
docker compose up -dSet ANTHROPIC_API_KEY, PAGEFLY_EMAIL, and PAGEFLY_PASSWORD in the Railway dashboard. Add a volume mounted at /app/data for persistence.
- Web UI:
http://localhost(or your configured domain) - API:
http://localhost:8000/api— all endpoints documented - Telegram: Message your bot to start ingesting (if configured)
python -m src.cli load-demo # adds 3 sample docs + 5 wiki articles
python -m src.cli clear-demo # removes them| Category | Endpoints | Description |
|---|---|---|
| Knowledge | 14 | Ingest, list, read, update, delete, download documents |
| Wiki | 2 | List and read compiled wiki articles |
| Workspace | 8 | Rich-text documents with status flow (draft → finished → ingest) |
| Search & Query | 3 | Full-text search, agent Q&A, knowledge graph |
| Chat | 3 | Shared conversation (synced with Telegram) |
| Schedules | 6 | Cron task management with run history |
| Activity | 5 | Desktop capture events and audio upload |
| Roam | 1 | Random knowledge resurfacing |
| Categories | 3 | Dynamic category management |
| System | 7 | Stats, trends, tokens, health, demo data |
Auth: Bearer token (JWT from login, API token, or master token).
The server runs on its own — the clients below are optional add-ons that capture content into your PageFly instance.
One-click clip the page you're reading into your knowledge base.
Path: browser-extension/ (Manifest V3, unpacked load).
1. Open chrome://extensions → enable "Developer mode"
2. "Load unpacked" → pick the browser-extension/ folder
3. Click icon → set Server URL + API token
4. On any page: click icon → "Clip this page"
Menu-bar app that captures your active app + window context every few seconds and records meeting audio that gets transcribed server-side.
Path: desktop-capture/ (Swift / SwiftUI, Xcode 15+).
cd desktop-capture && ./scripts/package-local.sh
# → dist/PageflyCapture-<version>.dmgOpen PageflyCapture → menu bar icon → Preferences → enter server URL + API token → grant Accessibility + Microphone.
| Layer | Choice |
|---|---|
| Runtime | Python 3.11+ |
| API | FastAPI |
| Database | SQLite (WAL mode) |
| AI Agents | Claude Agent SDK (Anthropic) |
| Scheduler | APScheduler |
| Bot | python-telegram-bot |
| Layer | Choice |
|---|---|
| Framework | React 19 + Vite + TypeScript |
| Editor | Tiptap (ProseMirror) |
| Styling | Tailwind CSS v4 |
| Router | react-router-dom v6 |
| Icons | Lucide React |
| Task | Model |
|---|---|
| Classification & Agents | Claude (Anthropic) |
| Voice Transcription | gpt-4o-transcribe (OpenAI) |
| Image OCR | mistral-ocr-latest + mistral-small-latest |
pagefly/
├── src/
│ ├── agents/ # Compiler, Linker, Trend, Query, Review (Claude SDK)
│ ├── channels/ # Telegram bot, REST API
│ ├── governance/ # Classifier, Organizer, Integrity checker
│ ├── ingest/ # Pipeline + converters (PDF, DOCX, voice, image, URL)
│ ├── scheduler/ # Cron jobs, inbox watcher
│ ├── shared/ # Config, roam, indexer, types
│ ├── storage/ # SQLite DB, deletion logic
│ └── auth/ # JWT, TOTP 2FA, email verification
├── config/
│ ├── SCHEMA.md # Wiki conventions (injected into agent prompts)
│ └── skills/ # Agent skill definitions
├── frontend/ # React + Vite + Tailwind + Tiptap
├── browser-extension/ # Chrome extension (Manifest V3)
├── desktop-capture/ # macOS menu bar app (Swift)
├── data/ # Runtime data (not tracked)
│ ├── raw/ # Ingested documents
│ ├── knowledge/ # Classified & organized
│ ├── wiki/ # Compiled articles
│ └── workspace/ # Editor images
├── docker-compose.yml
└── Dockerfile # Multi-stage (Python + Node frontend build)
- Author: @yrzhe_top
- Live: pagefly.ink
- Inspired by: Karpathy's LLMWiki
MIT — do whatever you want with it.