Your workspace. Your GPU. Your AI.
An offline-first, AMD-accelerated AI collaboration workspace for teams and contributors who want private, local AI. Mo cloud, no subscriptions, no data leaving your machine.
RyFlow turns the AMD chip already in your machine into a private, offline AI brain that your entire team shares. Every feature — writing assistance, semantic search, real-time collaboration, voice transcription — runs locally using AMD's open-source ROCm stack and free, quantized language models.
The result: A zero-cost AI workspace that works on air-gapped networks, remembers everything your team has ever created, and gets faster when it detects an AMD GPU.
- Engineering, product, and research teams on LAN or air-gapped networks
- Professionals who need local LLMs without cloud lock-in or usage fees
- Open source contributors extending a self-hosted AI workspace
- Anyone who wants a fully private Notion-style workspace with built-in AI
See ARCHITECTURE.md for the full system design, data model, and contributor guide.
| Feature | Description |
|---|---|
| Local AI | Phi-3 Mini / Gemma 2B via Ollama — runs entirely on-device, no API keys |
| AMD ROCm | Auto-detects AMD GPU and switches to hardware-accelerated inference (up to 10× faster) |
| P2P Collaboration | Real-time co-editing over LAN — documents and code files via Y.js CRDTs + WebRTC; task board via Socket.io broadcast |
| Remote Join Codes | Connect to a host's workspace from another machine over LAN using a generated join code |
| Hybrid Search | Semantic vector search (HNSW ANN) fused with BM25 keyword search (FTS5) and recency scoring |
| Knowledge Graph | Auto-built graph of backlinks, @mentions, and embedding-derived edges; D3 visualisation |
| Multi-Workspace | Create, switch, and manage multiple isolated workspaces from a single session |
| Document Editor | TipTap rich-text editor with AI writing assistance, comments, version history, and backlinks |
| Code Editor | Monaco editor with AI explain, debug, and optimise |
| Canvas | Excalidraw-powered freeform drawing with AI description |
| Task Board | Natural-language task creation — type a sentence, get a Kanban card |
| AI Studio | Persistent RAG chat — answers are grounded in your workspace content |
| Study Guide | Auto-generate summaries, key terms, and quizzes from any document set |
| AI Morning Briefing | Daily LLM-generated standup digest read aloud via browser TTS |
| Command Palette | Ctrl/Cmd+K workspace-wide semantic search and navigation |
| Templates | Built-in and custom workspace templates (Meeting Notes, Project Brief, Sprint Retro, and more) |
| Folder & Git Import | Ingest local .md + code files or clone a Git repo directly into a workspace |
| Voice Input | Offline transcription via Whisper.cpp (Ryzen AI NPU accelerated when available) |
| MCP Server | Expose workspace search and content to any MCP-compatible AI client via npm run mcp |
| Tags | Tag any item across the workspace for filtered views |
| Daily Notes | Auto-created daily scratchpad |
| Workspace Export | Export entire workspace as a .ryflow bundle; import on any device |
| LAN Discovery | Auto-discover peers on the same network via mDNS |
| Image Generation | AI image generation via Pollinations.ai (requires internet — the only non-local AI feature) |
| Category | Technologies |
|---|---|
| Frontend | React 18 + Vite, TailwindCSS, TipTap, D3.js, Monaco Editor, Excalidraw, Framer Motion, Zustand |
| Backend | Node.js + Express, Socket.io, better-sqlite3 (SQLite) |
| AI Layer | Ollama (Phi-3 Mini, nomic-embed-text), Whisper.cpp, Pollinations.ai |
| AMD Stack | ROCm (GPU inference), llama.cpp HIP, Ryzen AI SDK + ONNX Runtime (NPU) |
| Networking | WebRTC (simple-peer), Y.js CRDTs, Bonjour/mDNS |
| Desktop | Electron |
| Container | Docker + Docker Compose |
The easiest way to run RyFlow. Downloads Ollama and the required AI models automatically.
Prerequisites: Docker Desktop
git clone https://github.com/sh4shv4t/RyFlow.git
cd RyFlow
docker compose upThen open: http://localhost:5173
First run downloads phi3:mini (~2.4 GB) and nomic-embed-text (~270 MB). This is a one-time setup; subsequent starts are instant.
Prerequisites:
- Node.js 18+
- Ollama installed
git clone https://github.com/sh4shv4t/RyFlow.git
cd RyFlow
# Install all dependencies
npm run install:all
# Start backend, frontend, and Ollama health check together
npm run devnpm run dev auto-checks Ollama, starts it if possible, and pulls required models on first run.
If Ollama cannot be auto-started, run it manually first:
ollama serve
ollama pull phi3:mini
ollama pull nomic-embed-text# After completing the Option 2 setup:
npm run electronIf you have an AMD GPU, install ROCm for up to 10× faster AI inference:
- Linux: ROCm Installation Guide
- Windows: ROCm support via WSL2
Ollama detects and uses ROCm automatically. RyFlow shows ⚡ AMD in the top bar when active.
git clone https://github.com/ggerganov/whisper.cpp
cd whisper.cpp
make
bash ./models/download-ggml-model.sh baseSet the binary and model paths in your .env (see .env.example).
# Run backend and frontend together (with Ollama check)
npm run dev
# Backend only (port 3001)
npm run dev:backend
# Frontend only (port 5173)
npm run dev:frontend
# Frontend unit tests
npm --prefix frontend test -- --run
# Backend integration tests
npm --prefix backend run test:integration
# MCP server (requires an active workspace)
npm run mcp
# Build for production
npm --prefix frontend run buildRyFlow runs Phi-3 Mini via Ollama entirely on your device — no data leaves your machine. Every chat, summarisation, and task parse happens locally. When an AMD GPU with ROCm is detected, inference is automatically offloaded to the GPU.
Every document, task, code file, canvas, and chat is converted into a 768-dimension vector embedding using nomic-embed-text and stored in SQLite. Retrieval uses a hybrid pipeline: approximate nearest-neighbour search (HNSW) is fused with BM25 full-text search (FTS5) via Reciprocal Rank Fusion, then re-ranked by recency. The result is semantic + keyword search across your entire workspace history.
Peers on the same LAN are discovered automatically via mDNS. Three collaboration channels are active in remote (joined) mode:
- Documents and code files — Y.js CRDTs over WebRTC with a self-hosted signaling server at
ws://host:3001/yjs. No external relay. - Task board — Socket.io broadcast events (
task-created,task-updated,task-deleted) relayed through the host. Creates, status moves, edits, and deletes appear live on all peers' boards. - Canvas — not yet synced live; changes reflect on next load.
Remote workspaces on different machines connect via join codes, with the host acting as a reverse proxy.
Export your entire workspace as a .ryflow file (a zip of the SQLite database and uploads). Import it on any device to continue where you left off. Workspace knowledge persists across devices, migrations, and team handoffs.
Run npm run mcp to expose your active workspace to any Model Context Protocol-compatible client (Claude Desktop, Cursor, etc.). Available tools: search_workspace, get_document, get_neighbours, list_workspace_contents.
Contributions of all sizes are welcome. For anything beyond a small fix, please open an issue first so we can align on approach.
frontend/ # React + Vite SPA — pages, editors, Zustand store, Vitest tests
backend/ # Express API, SQLite, Ollama/Whisper services, P2P discovery, MCP server
electron/ # Desktop shell — spawns backend/frontend, system tray, electron-builder
scripts/ # Ollama bootstrap and model-pull helpers
See ARCHITECTURE.md for a full system map, data model, and notes on each subsystem. COLLABORATION_AUDIT.md covers the Y.js and WebRTC layer in depth.
# Fork, then clone your fork
git clone https://github.com/your-username/RyFlow.git
cd RyFlow
npm run install:all
git checkout -b feat/your-feature-name
# Make changes, then run tests
npm --prefix frontend test -- --run
npm --prefix backend run test:integration
git commit -m "feat: describe your change"
git push origin feat/your-feature-name
# Open a Pull RequestFollow conventional commit prefixes: feat:, fix:, docs:, refactor:, test:.
MIT — see LICENSE for details.
RyFlow — Don't rent intelligence. Run it.