If you use JetBrains and you have ever winced at sending every keystroke of a private repo off to a hosted model, here is a small but genuinely useful change. GitHub Copilot for JetBrains can now run models on your own machine.
What changed on 11 August
GitHub’s 11 August 2026 update to Copilot for JetBrains adds Ollama as a BYOK (bring your own key) provider. In plain terms, you can point Copilot at a model running locally through Ollama, with provider configuration and model selection wired throughout the JetBrains experience: IntelliJ IDEA, PyCharm, WebStorm and the rest. The same release adds Copilot memory that persists across chat sessions, so you stop re-explaining your project every time you open a new chat.
There is a pile of smaller fixes too, including automatic Copilot CLI install from the integrated terminal on macOS, Linux and Windows, and restored # file and folder references in the chat box.
Why local models are worth a look
Running a model on your own machine buys you three things:
- Privacy. Prompts and code go to a model on your hardware instead of a hosted endpoint, which is a real difference for sensitive repositories.
- Cost. Local inference does not meter you per token, so heavy, repetitive work gets cheaper to run.
- Control. You choose the model, and you can keep working when the network is flaky.
How to set it up
The rough path, if you want to try it this weekend:
- Install Ollama and pull a coding model, for example a Qwen or Llama variant sized to your RAM.
- In JetBrains, open Copilot’s model settings and add Ollama as a BYOK provider.
- Pick your local model from the model picker and start a chat.
Read the asterisk
Two honest caveats. First, BYOK does not make Copilot free. You still need a Copilot subscription. BYOK changes which model answers, not whether you are paying GitHub. Second, local models are not frontier models. A 7B or 14B model on your laptop will not match a hosted flagship on hard reasoning, and it will lean on your hardware. If your machine is thin on RAM or lacks a decent GPU, expect slower responses and smaller context windows. Use local models for the bread-and-butter work and keep a hosted model on hand for the heavy lifting.
The bigger pattern
This is part of a wider drift in 2026: BYOK and local inference turning up as first-class options inside mainstream tools instead of hacker workarounds. GitHub also folded in enterprise managed settings so admins can control provider access, MCP servers and permission behaviour from a server, which is the boring but important plumbing that lets big companies actually allow this. Local models in your IDE, sanctioned by IT, is a genuinely nice place to land.
Did you know: Ollama began as a way to run Llama models on a Mac and now works as a local model runtime that a growing list of professional IDE tools plug straight into.
Sources
Related on Top Tool Stack: Agent Plugins 1.0: One Plugin That Runs Across Every AI Coding Tool · Claude Code Grows Up: Self-Hosted Runners and Sessions That Talk to Each Other · AI on a Budget: cut your AI costs