GitHub Copilot (CLI)

GitHub Copilot's CLI supports BYOK (Bring Your Own Key) — you point it at any OpenAI-compatible base URL via env vars, and the CLI routes chat requests through your custom provider instead of GitHub's backend. rapid-mlx is that provider.

Install

$ gh extension install github/gh-copilot
# Requires an authenticated ``gh`` session (``gh auth login``).

Config

Per GitHub's BYOK docs (docs.github.com), the env vars are:

$ export COPILOT_PROVIDER_TYPE=openai
$ export COPILOT_PROVIDER_BASE_URL=http://localhost:8000/v1
$ export COPILOT_PROVIDER_API_KEY=not-needed
$ export COPILOT_MODEL=default

Run

$ rapid-mlx serve qwen3.6-35b-4bit \
    --tool-call-parser qwen3_coder_xml --enable-auto-tool-choice
$ gh copilot suggest "find the largest file in this directory"

Gotchas

• Real-CLI drive is blocked on gh auth login OAuth (interactive TTY, no --no-tty escape hatch as of 2026-07). The integration matrix cell smokes the wire (/v1/chat/completions tool-call round-trip) rather than the vendor binary. A follow-up sibling PR will add a real-CLI subprocess cell once a token-flow harness is agreed.
• DeepSeek R1-Distill XFAILs. Same tool-emission gap as every other function-calling agent (see XFAIL arch).

Empirical

The copilot row of the integration matrix is ✅ on Qwen 3.6, Gemma 4, and gpt-oss (wire-smoke); DeepSeek R1-Distill XFAIL(arch). Source of truth for this page: TestCopilot + BYOK docs at docs.github.com.

See also