01
When to use LocalOpenAIAgentConfig
Use LocalOpenAIAgentConfig when you already run a model behind a local OpenAI-compatible endpoint or want backend flexibility outside the SDK-managed LiteRT workflow. You operate the inference server; Antigravity connects to it.
The surrounding agent layer can continue to manage tools, workspaces, policies, hooks and triggers, but model behavior remains backend-dependent. API-shape compatibility does not make all local models equally capable.
02
Connect Antigravity to Ollama
Google's repository documents Ollama as an external OpenAI-compatible option and provides localhost:11434/v1 as an example base URL. The model identifier must match the model exposed by your Ollama installation.
No OpenAI-hosted API key is required merely because the local server uses an OpenAI-compatible protocol. Authentication, model availability and hardware requirements are properties of your local server configuration.
03
LM Studio and vLLM
LM Studio can expose a local OpenAI-compatible server. Start the server, load the intended model and point Antigravity at the endpoint and model identifier actually shown by LM Studio instead of assuming an unrelated port.
Google's launch post also names vLLM. Run vLLM's OpenAI-compatible server with a model it supports, then configure Antigravity against that endpoint. GPU requirements come from the chosen model and vLLM, not from Antigravity alone.
04
External server versus LiteRTAgentConfig
Use LiteRTAgentConfig for Google's SDK-managed LiteRT route, currently best documented with Gemma 4 26B A4B. Use LocalOpenAIAgentConfig when the inference server has its own lifecycle outside Antigravity.
Google warns against using LocalOpenAIAgentConfig to connect to litert-lm serve because LiteRTAgentConfig is designed to manage the LiteRT server itself.
05
Privacy, compatibility and performance limits
A localhost inference endpoint can keep model requests on the machine, but remote MCP servers, web APIs and cloud services attached to the same agent can still transmit data. Verify the full tool graph before claiming an offline workflow.
There is no justified universal fastest backend. Performance depends on the model, quantization, runtime, hardware, context and task. Test tool calling and instruction following as carefully as raw tokens-per-second.
Sources
Primary and supporting sources
Use the linked documentation to check current pricing, access and product behaviour. The dates above describe this article; they do not guarantee that every vendor claim remains current.