Github CoPilot as a Model Provider

[elasticdotventures]

Aider has a Copilot interface
https://aider.chat/docs/llms/github.html

Github Copilot offers a $10/mo subscription that gives 300 'pro' (claude sonnet 4, gemini 2.5, et al) and UNLIMITED OpenAI GPT-4.1 and 4o with 65k context size. These tools work well for a variety of small scoped tasks from an orchestrator.

There is also a free tier which current grants ~2,000 free requests per month, so literally anybody with a github account that's got to be 100% of RooCodes user base.

It appears possible/straightforward to access the Copilot LSP via JSONRPC and initiate a github signin request, that is the device auth flow where they give you an 8 digit code, and that returns an auth token that plugins are supposed to hidden inside the secret enclave by the plugin. Currently the easiest way to short-cut this is to a working demo is use the Aider backdoor -- install ANY IntelliJ/jetbrains editors (pycharm, rust rover, etc.) and it the editor will write a file with a cleartext token, that can today be put into Roo as an OpenAPI compatible endpoint with the https://api.githubcopilot.com OPENAI_BASE_URL and it just works, out of the box. It doesn't need to be refreshed unless you remove the token, you can then uninstall the jetbrains ide.

My request for information & documentation on the Github VsCode Copilot LSP is documented here:
github/docs#39398

** since the Xcode implementation/version is open source I think that's sufficient, it'll take a ton of coordination for Redmond to break that interface and disable (alienate) users of neo/vim, xcode, intellij etc. -- but I am suggesting taking a more formal route once this approved and approaching Microsoft directly and not have another repeat of the gemini-cli.

In terms of specific features I'd like to see in the provider would include:

• autodection/enablement of plugin & models if Github Copilot LSP is setup in VSCODE
• ability to filter models by free / pro.
• ability to use pro models (ex: Gemini 2.5) for context compression
• ability to use free models & fail over to pro models when context size is insufficient

Delegating to code 4o or 41 is effective, especially if treated as a smol agent -- total grunt work. Writing/checking tests, anything <64k -- especially for situations where multiple agents are reviewing code each with their own mission/focus for separation of concerns. I.e. cognitive linting, looking for bad patterns, running tests that should pass (not necessarily fixing the code).

Note: I am currently developing my own library/extension and will share that code/library when it's finished. Github Copilot is particularly interesting because it's free & generous - github accounts are free - so it's a great onboarding for students, etc.

(I am not an emloyee, but I've got friends in Redmon, who can hopefully bring awareness/engagement once these features are on the Roo roadmap! 😉💖🦘)

[SannidhyaSah]

Its been already there for months. It's called VS code Lm provider .
Check this .https://docs.roocode.com/providers/vscode-lm