Copilot CLI finds local Ollama models, but local isn't offline

GitHub Copilot CLI can now discover local Ollama models on its own. Start Ollama, type /model, and your local models show up next to Copilot's cloud models. Handy. But GitHub's changelog is clear on one point many teams will miss: choosing a local model does not turn on offline mode and does not turn off GitHub telemetry.

What actually happened

Per GitHub's October 7 changelog:

• Version: Copilot CLI 1.0.94-0 and later.
• Discovery, not auto-install. /model lists models from a running local Ollama instance. Nothing gets added until you pick one, review its provider and endpoint, and choose "Add and use for this session" or "Add without switching."
• No restart. You can switch to the local model mid-session.
• Requirements. Ollama and the model must already be installed. The model must support tool calling and streaming, or it won't work as an agent.
• Offline is separate. Offline mode stays opt-in through COPILOT_OFFLINE=true. And even in offline mode, a remote provider endpoint can still receive prompts and code context.
• Only Ollama is named. The post does not mention LM Studio or other runtimes, and it says nothing about plan eligibility or billing.

Why it matters for your business

Plenty of small teams want a coding agent on code they can't send to a cloud model: a client's repo under NDA, a pricing engine, a customer database schema. Running the model locally is the right instinct. Assuming that makes the whole session private is the wrong one.

There are two separate questions. Where does inference run? Local, if you picked an Ollama model. What else leaves the machine? That depends on telemetry and offline settings, which the model picker does not change.

What we'd set up:

1. Set COPILOT_OFFLINE=true in the shell profile on any machine that touches sensitive repos. Do it in your dotfiles or device management, not by memory.
2. Check the endpoint. When you add a discovered model, confirm it points at localhost, not a shared box on the network.
3. Pick a model that does tool calls well. A small model that fumbles tool calls will waste more time than it saves. Test on a real task first.
4. Write it down. One line in your security notes: which repos are local-only, and how that's enforced.

Sources: GitHub Changelog: Discover local models in GitHub Copilot CLI, GitHub Docs: Use BYOK models in Copilot CLI.