Open-source, local-first, project-centric multi-agent AI productivity desktop workspace with built-in terminal, email, calendar, kanban, and infinite canvas.
OpenLoaf is a desktop AI productivity application developed by Hexems, built around a "project" core unit with a custom three-layer Agent architecture: Secretary Agent for global routing and cross-project coordination, Project Agent with full single-project context awareness, and Worker Agents spawned on-demand for specialized tasks. The system features a three-tier memory system (User / Project / Linked Projects), Markdown-based Skills workflows, MCP tool protocol support, and built-in tools including infinite canvas, terminal, email, calendar, file manager, task kanban, and rich text editor.
Multi-Agent System
- Secretary Agent: Main window global assistant handling calendar, email, project routing, cross-project queries, with automatic task dispatch decisions
- Project Agent: Per-project independent AI with full context awareness of project files, code, docs, terminal, and tasks
- Worker Agents: On-demand spawned specialized sub-agents (exploration/planning/coding/review)
- Intelligent Routing: Simple questions answered directly; project-level tasks routed to corresponding Project Agent; cross-project tasks parallel-dispatched to multiple Agents
Project & Linking Mechanism
- Each project opens in an independent Electron window or browser tab, fully isolated
- User-defined type tags to organize projects (e.g., code / docs / knowledge base)
- Any project can link to others, auto-injecting linked project's memory and skills into current project's AI context
Memory & Skills System
- Three-tier memory: User-level (
~/.openloaf/memory/) → Project-level (<project>/.openloaf/memory/) → Linked Projects (auto-loaded) - Skills system: Reusable Markdown workflows defined by
SKILL.md, supporting global and project-level scopes, dynamically discovered and loaded by AI at runtime
AI Model Capabilities
- Supports 10+ models: OpenAI, Anthropic Claude, Google Gemini, DeepSeek, Qwen, xAI Grok, Ollama, etc.
- BYOK mode: users bring their own API keys, device connects directly to model providers
- Fully air-gapped offline operation via Ollama
- AI感知项目完整上下文(文件结构、文档内容、对话历史)
MCP Support
- Connection via stdio / http / sse
- Global and project-level configuration; can directly import MCP JSON configs from Claude Desktop / Cursor / VS Code / Cline / Windsurf
Built-in Productivity Tools
- Terminal: Full terminal emulator, AI Agent executes commands after approval
- Email: Multi-account IMAP, AI drafting and summarization
- Calendar: macOS / Google Calendar sync, AI scheduling
- File Manager: Grid/list/column views, drag-and-drop, file preview
- Task Kanban: Kanban (To Do → In Progress → Review → Done), priority tags, AI task creation
- Rich Text Editor: Based on Plate.js, supporting LaTeX, tables, code blocks, bidirectional links
- Infinite Canvas: Based on ReactFlow, supporting notes/images/video/drawing/AI image generation/AI video generation/image content understanding
Data Storage & Privacy
- All data stored locally at
~/.openloaf/, SQLite database + filesystem - Zero telemetry, zero uploads, no analytics tracking
Core Dependencies
- Frontend: Next.js 16 / React 19 / Tailwind CSS 4
- Backend: Hono + tRPC / Prisma 7 + SQLite
- Desktop: Electron 40
- AI Integration: Vercel AI SDK
- Collaboration Engine: Yjs
- Canvas Engine: ReactFlow
- Build: Turborepo + pnpm monorepo
- Language breakdown: TypeScript 93.8% / JavaScript 4.8% / Swift 0.6% / CSS 0.3%
Installation & Deployment
- Prerequisites: Node.js >= 20, pnpm >= 10
- From source:
git clone→pnpm install→pnpm run db:migrate→pnpm run dev(Web) orpnpm run desktop - Pre-built: GitHub Releases provides macOS (Apple Silicon / Intel), Windows Installer, Linux AppImage
- Docker: Repository includes
Dockerfile.devanddocker-compose.yml
Use Cases
- Software development: One project per repo, link coding standards project for cross-repo consistent AI behavior
- Research & writing: Create knowledge base project linked to paper projects, AI extracts from curated sources
- Content creation: Canvas brainstorming → AI image generation → editor writing → kanban tracking
- Project management: One project per client, Secretary Agent provides cross-project overview
- Personal knowledge base: Accumulate notes, web clippings, journals, link to work projects for AI-discovered associations
Unconfirmed Information
- No Hugging Face page found; HF integration status unconfirmed
- No associated academic papers found
- Commercial license terms not publicly disclosed (AGPLv3 + commercial dual-license, contact zz@hexems.com)
- Paid credit system (Free/Lite/Pro/Flagship tiers) specific consumption ratios need testing
- Whether Yjs collaboration real-time features are currently enabled is unclear
- Specific native purpose of Swift 0.6% share is unexplained