Add ability to use local LLMs in Copilot

[moffatted]

Select Topic Area

Question

Body

I'm enjoying using the new agentic features of Copilot. I was in the middle of a major refactor yesterday and got rate-limited. I completely understand that is necessary. It would be nice if Copilot would allow me to switch to a local LLM from within Copilot. This seems reasonable if you are trying to ensure fair and equitable access to everyone. I know I could switch to Cline or Continue to use my local LLMs, but I'd prefer to stay within the same user interface. I think what you guys offer for $10.00 a month is very reasonable, but offloading to local LLM's might solve your equitable use issue if developers are getting rate limited.

[notcoderhuman]

Hi moffatted 👋

+1 — this is a really thoughtful suggestion! Rate limits during heavy agentic/refactor sessions are a pain (especially mid-flow), and offloading to a local LLM for overflow or lighter tasks makes total sense for power users. Staying in the same slick Copilot UI while switching to local models would be huge — no context-switching to Continue.dev, Cline, or Ollama directly.

Current state in 2026 (from recent updates & community feedback)

GitHub Copilot does support "bring your own model" (BYOM) / custom LLMs in VS Code (and expanding to other IDEs):

• You can connect to providers like Ollama (local!), OpenAI-compatible endpoints, Anthropic, Gemini, etc., via the model picker in Copilot Chat.
• Setup: In VS Code → Copilot settings → Language Models → Add your Ollama/local endpoint (localhost:11434 usually works).
• This lets you use local models (e.g., DeepSeek-Coder, CodeLlama, Phi-4, Llama 3.1/4 variants fine-tuned for code) for completions/chat/agent tasks.
• But — there's still a catch for full offline/local-only:
  • GitHub login is required to activate Copilot features (even when using local models).
  • Some data (prompts/responses) may route through GitHub/Microsoft for filtering/safety (even with local backend).
  • Full agentic features (like Workspace agents, memory, multi-turn refactoring) are optimized for cloud models and may not work as seamlessly with local ones yet.

So you're halfway there: local models are supported, but not a seamless "switch to local when rate-limited" toggle in the UI, and not 100% offline/no-login.

Why this feature request is spot-on

• Solves rate limits fairly (cloud for premium/fast, local for unlimited/heavy lifts).
• Privacy boost: Local models keep sensitive code fully on-device.
• Cost: $10/mo is fair for cloud access, but local fallback reduces cloud burn.
• Ecosystem fit: Many devs already run Ollama/LM Studio/Tabby for privacy/speed — integrating tighter would rock.

Workarounds while we wait

1. Use BYOM in Copilot (as above) — set Ollama as fallback for chat/completions when rate-limited.
2. Continue.dev extension (free, excellent) — full Copilot-like UI, native local LLM support (Ollama, LM Studio, etc.), context-aware, agents. Many switch here during heavy sessions.
3. Cline or Tabby — similar local-first agents, but Continue is closest to Copilot's feel.
4. Monitor GitHub feedback — similar requests exist (e.g., full offline/local without login, configurable Ollama URLs, self-hosted custom models). Upvote/comment on them:
   • https://github.com/orgs/community/discussions/150243 (self-hosted custom LLMs)
   • Use copilot with local models completely offline microsoft/vscode#246551 (offline local models)
   • Search "local LLM" or "Ollama" in Copilot discussions.

GitHub/Microsoft team — this would be a killer addition for Pro users: a simple "Fallback to local when limited" toggle or deeper local integration (no login for local mode, full agent support). It aligns with privacy trends and helps heavy users stay productive.

Thanks for posting — upvoted and following! What local model/setup are you thinking of using (Ollama + Llama 3.1 70B? DeepSeek?)? Happy to share setup tips if needed.

Keep refactoring! 🚀

[moffatted]

Thanks for the info I wasn't aware you could do a local Ollama.in Copilot I had googled about this and the response was that it couldn't be done. I've been playing with different models including qwen2.5-coder14b and DeepSeek coder 14b. I'm currently limited to 16gb of vram.

[shinybrightstar]

👋 Welcome to the GitHub Community, @moffatted! Congratulations on your first post! 🎉

You are more likely to get a useful response if you are posting your question(s) in the applicable category. Check out this Guide on how to post in the right category. It'll ensure your questions and discussions reach the right people faster!

This particular discussion belongs in the Copilot Conversations category, I've gone ahead and moved it for you. Good luck!

Happy to have you here! 🚀