GitHub Copilot CLI Can Now Discover Local Ollama Models

Developers who already run AI models on their own PCs have one less configuration step to worry about. GitHub has updated Copilot CLI so its model picker can discover compatible models from a running local Ollama installation.

Announced on October 7, 2026, the feature begins with Copilot CLI version 1.0.94-0. Instead of leaving the command-line workflow to manually look up every local endpoint, users can open the `/model` picker and review Ollama models alongside their existing configured and cloud options.

It is a convenience feature, not a promise that every downloaded model will become a capable coding agent.

How discovery works

GitHub says Copilot CLI checks a running local Ollama instance for supported models. The CLI does not install Ollama, download model weights or start a missing runtime for you. Those parts must already be working on the machine.

Once discovery finds a compatible model, the picker displays its provider and endpoint. You can then choose Add and use for this session or Add without switching. GitHub says you can start using a selected model in the current session without restarting the CLI.

If the provider cannot be reached, the picker can show an explanation of the connection failure. That is helpful when the real problem is a stopped Ollama server or an incorrect endpoint rather than the model itself.

What you need before trying it

You need a working installation of Ollama with at least one local model already pulled, plus GitHub Copilot CLI at version 1.0.94-0 or later. Crucially, GitHub says discovered models must support tool calling and streaming. A model that can chat normally may still be unsuitable for an agentic coding workflow if it cannot use tools reliably.

A practical starting checklist is to confirm that Ollama can list and run the model on your computer, then open Copilot CLI and enter `/model`. Review the discovered entry before adding it. Do not assume that discovery changes your default model for every future session.

For users running on modest hardware, memory and context size still matter. A large local model may fit only partially in GPU memory, making coding-agent responses slow. The new picker does not increase VRAM, improve a model's tool-use training or guarantee parity with a cloud coding model.

Local model does not automatically mean offline

This is the most important caveat in GitHub's announcement. Selecting a local Ollama model does not automatically enable offline mode or disable GitHub telemetry.

For Copilot CLI, GitHub documents offline mode as an explicit choice using `COPILOT_OFFLINE=true`. That is separate from selecting the model. GitHub also warns that a remote model provider can still receive prompts and code context over the network, even when the CLI is configured for offline mode. In other words, verify the endpoint and provider rather than relying on the word “local” in a menu.

This distinction matters for developers working with private repositories, client information or unpublished code. Before using any agent, check the provider, data-handling policies, repository access and the commands it may run.

What about the GitHub Copilot desktop app?

GitHub points to a related provider workflow in the Copilot app. There, users can add a supported provider through Settings > Model providers. The October CLI improvement specifically adds discovery inside `/model`; the app and CLI should not be treated as having identical setup screens or behavior.

GitHub has separately announced local sandboxing for Copilot CLI, the Copilot app and supported VS Code agent sessions. Sandboxing controls what tools and commands can access; it does not change where a model runs. Local model execution and tool isolation are different settings that can be used together.

Is this worth trying on a Windows gaming PC?

For developers who already use Ollama, yes, particularly as a way to experiment with private, lower-cost local coding assistance. You can compare a compatible small model against your normal cloud model on a contained task such as explaining a function or writing tests for a small module.

But treat it as a workflow option, not a performance upgrade. Tool reliability, memory requirements and security settings remain your responsibility. The official GitHub Copilot CLI announcement and provider documentation are the best places to confirm the exact setup supported by your installed version.