Google Antigravity SDK adds support for local AI agents

Google has announced that its Antigravity SDK now supports local model execution, allowing developers to run AI coding agents directly on their own machines. This update provides two primary paths for local workflows: using Google AI Edge LiteRT with models like Gemma 4 26B A4B, or utilizing OpenAI-compatible local servers such as Ollama, LM Studio, and vLLM.

By enabling local execution, developers can achieve greater cost efficiency by avoiding per-token cloud inference charges and rate limits. The setup also offers enhanced privacy, as source code and prompts can remain on the local machine, and provides offline resiliency for environments without stable internet connections.

Additionally, the SDK supports hybrid workflows, where developers can configure a system to use cloud-based models for high-level planning while delegating code-heavy tasks to local agents. For the featured Gemma 4 26B setup, Google recommends hardware with at least 24GB of VRAM or unified memory. The LiteRT runtime supports various hardware backends, including Apple Silicon Metal, NVIDIA CUDA, and NPU support.