Jaaz is an open-source AI design agent that lets users create images and videos through a visual, prompt-free canvas—similar to playing with LEGO blocks. It’s designed for creators, marketers, and teams who need to generate high-quality visual content without relying on cloud-based tools or sacrificing data privacy. Unlike Canva or Manus, Jaaz runs locally or in private cloud environments, giving users full control over their assets and workflows.
Built with TypeScript and Python, Jaaz integrates with local AI models like ComfyUI and Ollama, and supports cloud APIs including GPT-4o, VEO 3, Kling, and Sora 2. It offers a cross-platform desktop app for Windows and macOS, with Docker-based enterprise deployment options. The system combines multimodal AI agents, an infinite canvas, and real-time collaboration to enable visual storytelling without text prompts.
What You Get
- Magic Canvas - A visual, prompt-free editing interface where users sketch, draw, or arrange elements, and AI instantly generates or refines content without requiring text prompts.
- Magic Video - Generate AI videos from simple visual instructions on a timeline; users can annotate video frames with arrows or notes and AI executes the sequence without manual prompting.
- Infinite Canvas & Visual Storyboarding - Create multi-scene visual narratives with unlimited canvas space, link layouts, and manage media assets visually with real-time collaboration support.
- Multi-Model AI Agent System - Seamlessly switch between local models (ComfyUI, Ollama) and cloud APIs (GPT-4o, VEO 3, Kling, Sora 2) to generate images and videos with consistent character and style coherence.
- Local-First Privacy & Offline Mode - All processing occurs on-device or in private infrastructure; no data is sent to third-party servers, making it suitable for sensitive or commercial use.
- Built-in Media & Prompt Library - Preloaded assets and templates for quick starts, with support for organizing, reusing, and sharing visual prompts and media files within the app.
Common Use Cases
- Creating marketing visuals for social media - A small business owner uses Jaaz to generate 9:16 product shots with consistent lighting and branding across multiple scenes, without uploading customer data to cloud services.
- Producing viral short-form videos - A content creator draws a simple storyboard on the canvas, adds annotations for camera movement, and Jaaz generates a 30-second AI video using VEO 3 or Kling for TikTok or Instagram Reels.
- Designing brand storyboards for agencies - A design team collaborates in real time on an infinite canvas to map out a 10-scene ad campaign, using local AI to generate variations of characters and environments without relying on external APIs.
- Developing AI-powered design workflows for enterprises - A corporate design department deploys Jaaz via Docker to create internal marketing assets while complying with data sovereignty regulations and avoiding SaaS vendor lock-in.
Under The Hood
Architecture
- Monolithic Electron architecture with tightly coupled frontend and backend components, lacking clear separation of concerns between UI and service layers
- Server-side logic directly exposed via FastAPI endpoints without dependency injection, service abstraction, or modular routing patterns
- React components tightly bound to API calls with no state management or data abstraction layers
- Directory structure lacks domain-based modularity, with frontend, backend, and Electron main process files co-located
- Absence of formal patterns like DI containers, interfaces, or factories, leading to direct class instantiation and tight coupling
Tech Stack
- React with TypeScript and Vite for frontend, bundled into a desktop app via Electron 35.x and electron-builder
- Python FastAPI backend with Pydantic for data validation, supported by Ruff and Black for code formatting
- Playwright with stealth plugins for browser automation, complemented by system-level utilities like 7zip-min and os-proxy-config
- Vitest configured for Node.js testing with forked pools and global utilities, alongside comprehensive ESLint/Prettier rules across multiple file types
- Cross-platform packaging with native installers (NSIS, DMG, AppImage, DEB) and integrated auto-update and notarization workflows
Code Quality
- Testing relies heavily on manual console logging rather than structured assertions or testing frameworks
- Error handling is generic and unstructured, with no custom error types or consistent error propagation
- Mixed naming conventions and Chinese comments in an otherwise English codebase reduce accessibility and maintainability
- No type safety enforcement beyond basic TypeScript usage, with mocked modules lacking interfaces or contracts
- Linting, static analysis, and CI quality gates are absent, leading to inconsistent code standards
What Makes It Unique
- Device-based OAuth flow tightly integrated with Electron enables seamless authentication without traditional login forms
- Dynamic, provider-aware model and tool selection with persistent multi-select capabilities for complex AI agent configurations
- Unified settings system with sidebar-driven, reload-free configuration panels
- Agent Studio implemented as a full workspace environment rather than a simple chat interface, supporting advanced agent orchestration
- Context-aware chat with mutable, savable conversation entities and deep theming/localization embedded directly into UI components