Copilot CLI supports custom providers and local models

GitHub Copilot CLI lets developers use remote or local AI models alongside GitHub-provided options. It automatically discovers downloaded models served by Ollama on the same machine, while other providers can be configured manually through providers.json. Supported connections include OpenAI-compatible services, Azure OpenAI Service and Anthropic. Models must support tool calling and streaming, and token limits should match their actual capabilities.

An experimental public preview also lets eligible Windows users install MAI Code 1.1 Flash and Flint NVFP4 directly from the model picker. Both require an NVIDIA GPU and a compatible llama.cpp installation. Available combined RAM requirements are 96 GB for MAI Code 1.1 Flash and 48 GB for Flint NVFP4. Running MAI Code 1.1 Flash locally enables offline use without consuming GitHub AI Credits, and Copilot Fusion can select it for appropriate tasks at the Fusion Standard effort level.

Custom-provider credentials stored in providers.json remain in plain text, so the file should stay private and out of repositories. Setting COPILOT_OFFLINE to true prevents the CLI from contacting GitHub servers, but full network isolation also requires a local or isolated model provider. Users should avoid running a downloaded local model simultaneously across Copilot surfaces because each loads a separate instance into memory.