What actually happened
On 23 September 2026 Google said the Antigravity SDK can run agents on the computer in front of you. Two paths: LiteRTAgentConfig with Gemma 4 26B A4B through Google AI Edge LiteRT, or LocalOpenAIAgentConfig against an OpenAI-compatible server such as Ollama, LM Studio or vLLM.
Google also showed a hybrid pattern. Gemini 3.8 Flash plans the work from filenames and task descriptions. Local Gemma agents read the files, propose fixes, criticise those fixes and run tests.
In the recorded run the planner looked at three vulnerable Python modules: auth.py, billing.py and database.py. Google says 97.2% of 3,322 tokens ran locally and 95 tokens went to the cloud. No source left the machine. That split is one workload, not a cost guarantee.
Why a small team cares
Take a five-person software consultancy that audits a client repo under an NDA. Dumping the tree into a cloud model may be a contract problem. Keeping the files local and still using a stronger model as planner is the useful split.
Local inference has no per-token API bill once the box is paid for. That matters if the same review job runs every week. The cloud model still gets the hard planning question. The local model gets the volume.
Google recommends at least 24 GB of VRAM or unified memory for the featured Gemma 4 26B checkpoint, which downloads about 16.8 GB. A recent Mac with enough unified memory, or a workstation GPU, is the realistic floor. A cheap laptop is not.
Hype vs useful
The hype is run everything offline and cancel ChatGPT. Useful is one confidential workflow where you already hesitate to upload.
The SDK talking to Ollama or LM Studio matters more than the Gemma badge. Same orchestration, different local backend. You do not rebuild the agent because the inference box changed.