GitHub lets Copilot CLI run on your own providers and local models

GitHub said in its April 7, 2026 changelog that Copilot CLI can now run against a user’s own model provider or fully local models instead of GitHub-hosted model routing. That sounds like a packaging change, but it materially changes where Copilot’s terminal agent can be deployed. Teams that already pay for Azure OpenAI, Anthropic, or another OpenAI-compatible service can now point the CLI at those endpoints directly, while developers running Ollama, vLLM, or Foundry Local can keep inference on their own machines.

The operational details matter. GitHub says COPILOT_OFFLINE=true prevents the CLI from contacting GitHub’s servers, disables telemetry, and limits the tool to the configured provider. Combined with a local model, that creates a fully air-gapped path for organizations that cannot send prompts or source context to external routing infrastructure. GitHub also made authentication optional for this mode. If a team wants only model access, provider credentials are enough. If a developer signs in as well, they can still combine the external model with GitHub-specific features such as /delegate, GitHub Code Search, and the GitHub MCP server.

GitHub added a few clear constraints alongside the announcement. The selected model needs tool calling and streaming support, and GitHub recommends at least a 128k token context window for best results. Built-in sub-agents such as explore, task, and code-review inherit the same provider configuration, and the CLI will not silently fall back to GitHub-hosted models when a provider setup is invalid. That behavior is important for governance because it means failures stay visible instead of leaking traffic to an unintended endpoint.

The bigger significance is strategic. GitHub is preserving the Copilot CLI interface while loosening the dependency on GitHub as the model router. That gives enterprise teams a way to standardize on one terminal workflow while choosing different models for cost, policy, latency, or data residency reasons. It also makes Copilot CLI more realistic for regulated environments where model access and network boundaries are tightly controlled.

What this does not do is remove the model-quality question. A weak local model will still produce weak agent behavior, and long-context tool use remains demanding even on strong open models. But GitHub has clearly moved Copilot CLI closer to a control plane for agentic terminal work, rather than a thin client for one hosted inference path.