Google Antigravity SDK now runs agents fully on the local machine

Running an agent locally used to mean giving up the model you wanted. Google updated its Antigravity SDK this week so an agent can run entirely on the developer's own machine, using Gemma 4 26B through Google's LiteRT runtime or any OpenAI-compatible local server such as Ollama, LM Studio or vLLM.

The SDK's new LiteRTAgentConfig class spins up and manages a local loopback server backed by a .litertlm checkpoint. The SDK detects the hardware backend at startup: GPU on Apple Silicon through Metal or on NVIDIA through CUDA, or NPU where one is available. Google's docs recommend 24 GB of VRAM or unified memory to run the Gemma 4 26B A4B checkpoint, which downloads about 16.8 GB. The checkpoint is pulled from Hugging Face through litert-lm and registered at ~/.litert-lm/models/.

What this changes

Antigravity's SDK was already a Python library for building AI agents with tools, policies and workspaces. The local option means the model the agent calls to plan, read code and write patches can also run on the laptop. The agent still carries the SDK's standard fields: tools, policies, hooks, MCP servers, subagents, workspaces and system instructions. The only new thing is where the model actually runs.

InfoQ, which first covered the update, reports a worked example from Google in which an agent auditing three vulnerable Python modules used 95 of roughly 3,400 tokens on a cloud model and 97.2 percent locally, with no source code sent to the cloud. Google supplied those figures; no independent benchmark has reproduced them, and the SDK docs state only the hardware requirement and the model names.

The second path, through an OpenAI-compatible server

For teams already running a local model behind an OpenAI-compatible API, there is a second class, LocalOpenAIAgentConfig, that lets Antigravity call it. Google's docs name Ollama, LM Studio and vLLM as supported servers. The config takes a model name, a base_url, and an optional .lightweight() call that applies tool and prompt optimizations for smaller local models. This is the path for a team that already uses ollama serve on an internal machine and wants Antigravity to call it instead of Google's cloud.

Separately, Google ships LiteRTAgentConfig for its own LiteRT backend, where the SDK handles the local server's lifecycle. The docs say not to point LocalOpenAIAgentConfig at litert-lm serve: use the dedicated class so the SDK can own the process.

The hardware floor

The 24 GB floor is the only hard limit named in the docs, and the 16.8 GB download runs on anything that clears it. Apple Silicon with 32 GB of unified memory meets this. Most mainstream laptop graphics cards do not; the compact Gemma 4 variants and smaller models from the OpenAI-compatible side are the alternative. For teams that run a client's code and are not allowed to send it to any cloud, that constraint is still cheaper than losing the engagement.

Source

• Primary: Google, Antigravity SDK local models documentation
• Reporting: Sergio De Simone, "Google Antigravity SDK Adds Local Model Support for Greater Token Efficiency and Code Privacy" on InfoQ