Gemma 4 powers local VS Code Copilot

This is not a product launch, but it is a useful field report: local coding agents are getting close enough that the limiting factor is now workflow polish, context management, and tool-call reliability.

• –OAI Compatible Provider for Copilot is the key bridge, letting developers reuse GitHub Copilot's VS Code UI with local or alternative model backends
• –Gemma 4 26B-A4B looks viable for coding-agent experiments when paired with careful llama.cpp cache, context, and template settings
• –The reported loops and tool-use failures are the real caveat: local inference speed does not automatically translate into reliable agent behavior
• –For privacy-sensitive or cost-conscious developers, this pattern points toward a practical self-hosted Copilot alternative rather than just a novelty demo