// the take
A reverse-engineered proxy that exposes GitHub Copilot's chat completions as OpenAI- and Anthropic-compatible HTTP endpoints, so you can point any OpenAI SDK, Anthropic SDK, or Claude Code itself at your Copilot subscription instead of a direct API key. It's for developers who already pay for Copilot and want to reuse that quota in tools built against the OpenAI or Anthropic wire format.
Translates both the OpenAI chat-completions shape and the Anthropic Messages API (including streaming and count_tokens), which is more than most Copilot proxies attempt. Ships concrete abuse-mitigation knobs — --rate-limit, --wait, and --manual per-request approval — showing the author actually thought about the ban risk rather than hand-waving it. Also includes a /usage endpoint and a small web dashboard for quota visibility, plus a reasonably careful Docker setup (multi-stage build, non-root user, pinned base image, healthcheck).
The entire project sits on a GitHub API that isn't public and that GitHub's own ToS explicitly warns against hammering — the README's abuse-detection warning is not boilerplate, it's the central operational risk, and a GitHub-side change can break this with no notice. Test coverage is thin: three test files cover Anthropic request/response translation and basic chat completions, with nothing exercising the device-code auth flow, token refresh, or rate-limiting logic, which are exactly the parts most likely to rot quietly. It's built around a single persisted GitHub token on local disk with no mention of encryption at rest, so it's not something to run multi-tenant or expose beyond localhost without extra work. And because it depends on scraping/mimicking Copilot's internal client behavior, every GitHub Copilot backend change is a potential breaking change with no SLA or deprecation notice.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →