Overview
GAIA is AMD's open-source agent factory for creating agent harnesses that run locally by default. It combines coding-agent skills, evaluations, and an auto-fix loop, while the flagship GAIA agent demonstrates conversation, document, data, web, memory, and skill capabilities in one system.
Local inference is designed to keep sensitive workloads on the user's machine, avoid cloud inference fees, and support air-gapped environments. GAIA is optimized for AMD Ryzen AI processors, using NPU and integrated GPU acceleration.
Key Features
- Runs AI models locally for greater control over data and offline deployment.
- Provides an SDK as the foundation for custom agent harnesses.
- Supports reusable skills, agent evaluations, and iterative auto-fixing.
- Offers a desktop application and terminal interface for the flagship agent.
- Includes optional cloud chat through Fireworks AI and AMD LLM Gateway via Lemonade.
Building and Running Agents
Developers can define an agent harness with a system prompt and registered tools, then process user queries through the SDK. Supporting documentation covers composing skills, evaluating agents, and working with the SDK.
Windows users can choose the desktop app, terminal, or both from one installer. Developer installation options include scripts, pip or uv workflows, source installation, and an npm client for embedding the agent in JavaScript or TypeScript applications.
Deployment Considerations
Local models do not require an API subscription, while optional cloud providers apply their own pricing and send conversation history to the selected provider. This distinction lets teams choose between local execution for private or disconnected settings and cloud-backed chat when it suits their workflow.