GitHub lets Copilot CLI use BYOK and fully local models without GitHub-hosted routing

In an April 7, 2026 post on X, GitHub Changelog said Copilot CLI no longer has to rely on GitHub-hosted model routing. Instead, teams can point the CLI at their own provider or run a local model while keeping the same terminal-native Copilot workflow. The linked changelog entry makes the scope explicit: GitHub supports Azure OpenAI, Anthropic, and any OpenAI-compatible endpoint through environment variables, while also allowing locally running models such as Ollama, vLLM, and Foundry Local.

That matters because it shifts Copilot CLI from a single hosted path to a more infrastructure-aware tool. GitHub says teams can set COPILOT_OFFLINE=true to stop the CLI from contacting GitHub servers at all. In that mode telemetry is disabled, and when paired with a local model the setup can support fully air-gapped workflows. GitHub also says authentication is optional if you only want to use your own provider credentials, though signing in still unlocks GitHub-native capabilities such as /delegate, Code Search, and the GitHub MCP server.

GitHub is turning Copilot CLI into a bring-your-own-model terminal agent

The more strategic signal is that GitHub is treating model choice, cost control, and network boundaries as first-class requirements for developer agents. The changelog notes that built-in sub-agents like explore, task, and code-review inherit the same provider configuration, so this is not just a fallback chat setting. GitHub also says the provider must support tool calling and streaming, and recommends at least a 128k context window, which shows the company still expects full agent-style terminal workflows rather than lightweight prompt completion. Sources: the X post and GitHub's changelog.