Google has added local-model support to the Antigravity SDK, allowing developers to keep code-intensive agent tasks on their own machines while optionally using a cloud model for high-level planning. The SDK is the programmable agent framework beneath the separate Antigravity coding environment.
Its hybrid “architect-builder” pattern gives one cloud agent a broad view of requirements and dependencies without sending it the source code. That agent creates a plan using filenames and task descriptions. Local builder agents then reproduce vulnerabilities, propose fixes, review patches and run regression tests in isolated environments.
In a demonstration that audited three vulnerable Python modules, Google says only 95 of about 3,400 tokens went to the cloud for planning. The remaining 97.2 percent were processed locally, and no source code was uploaded. The initial configuration runs Gemma 4 26B A4B through LiteRT, while an OpenAI-compatible connector supports local servers including Ollama, LM Studio and vLLM.
Keeping code local can improve privacy and reduce dependence on an external connection, but the token split alone does not establish lower cost or better performance. A community reviewer noted that developers should also measure accepted patches, elapsed time and energy use. The recommended 24GB or more of video or unified memory is a substantial hardware requirement, and organizations still need to secure the local agents and their execution sandboxes.