Package Exports

@ascdong/copilot-proxy
@ascdong/copilot-proxy/dist/index.js

This package does not declare an exports field, so the exports above have been automatically detected and optimized by JSPM instead. If any package subpath is missing, it is recommended to post an issue to the original package (@ascdong/copilot-proxy) to support the "exports" field. If that is not possible, create a JSPM override to customize the exports field for this package.

Readme

copilot-proxy

Expose your GitHub Copilot subscription as OpenAI-, Anthropic-, and Gemini-compatible HTTP endpoints, so any tool that speaks those APIs — Claude Code, Codex CLI, Gemini CLI, custom scripts — can route through Copilot.

Runs on Node.js ≥ 20 (or Bun in development).

Features

OpenAI-compatible: POST /v1/chat/completions, /chat/completions, /v1/responses (streaming + non-streaming), plus GET /v1/models
Anthropic-compatible: POST /v1/messages, /v1/messages/count_tokens (streaming + non-streaming)
Gemini-compatible: POST /v1beta/models/{model}:generateContent / :streamGenerateContent / :countTokens, plus GET /v1beta/models for CLI preflight
Format translation: Anthropic ↔ OpenAI, Gemini ↔ OpenAI, and Responses ↔ Chat Completions for models that don't support /v1/responses natively (e.g. Claude via Copilot)
Web search fallback via Tavily — when a model rejects Anthropic's web_search tool, the proxy runs the query and injects synthetic server_tool_use / web_search_tool_result blocks
Vision passthrough — image inputs are forwarded to vision-capable Copilot models
GitHub Device Flow OAuth — one-time login, tokens persisted locally
Auto-refreshing Copilot token — the short-lived upstream token is refreshed in the background
Model aliasing — claude-opus-4-6 ↔ claude-opus-4.6, etc.
5-category token tracking — every request logged with input / cache-creation / cache-read / output / reasoning tokens, duration, status
One-shot client configuration for Claude Code, Codex, and Gemini CLIs
systemd user service — install/uninstall as a background service on Linux/WSL

Getting started

Use this from inside WSL. Install the package in your WSL distro, log in once, then install it as a background service. The proxy listens on 127.0.0.1:8989 and serves both WSL and Windows clients on that port — there is no Windows-side install needed.

1. Open a WSL shell

wsl                 # or launch your distro from the Start menu

Make sure Node.js ≥ 20 is available inside WSL (node -v).

2. Install the package

npm install -g @ascdong/copilot-proxy

copilot-proxy login

A device-flow code is printed; open the URL in your browser and confirm. Tokens are persisted under ~/.copilot-proxy/.

4. Install as a background service

copilot-proxy service install

This registers a user-level systemd unit that starts the proxy on login and restarts it on failure. Verify it's up:

systemctl --user status copilot-proxy
curl http://127.0.0.1:8989/v1/models   # should return JSON

From here, any OpenAI / Anthropic / Gemini client — running in WSL or in Windows — can point at http://127.0.0.1:8989. Any API key will do; the proxy uses your Copilot session, not the client-provided key.

5. (Optional) wire up a client CLI

copilot-proxy config claude    # writes ~/.claude/settings.json (+ Windows path under WSL)
copilot-proxy config codex     # writes ~/.codex/config.toml
copilot-proxy config gemini    # writes ~/.gemini/.env + ~/.gemini/settings.json

Existing settings are merged, not overwritten.

Service management

copilot-proxy service reinstall   # apply package updates
copilot-proxy service uninstall
journalctl --user -u copilot-proxy -f

Configuration options

Optional YAML at ~/.copilot-proxy/config.yaml (auto-generated on first run). Defaults are fine for most users; override port / address / log level as needed. Environment variable GITHUB_TOKEN can be used in place of the interactive login.

Web search via Tavily

When a model rejects Anthropic's web_search tool, the proxy can transparently run the query through Tavily and synthesize server_tool_use / web_search_tool_result blocks so the client still gets a normal Anthropic-shaped response.

web_search:
  enabled: true
  provider: "tavily"
  tavily_api_key: "tvly-..."

Restart the proxy after editing. With enabled: false (the default) or a blank key, web_search requests are passed through unchanged.

CLI reference

copilot-proxy <command> [options]

Commands:
  start [-p <port>] [-H <host>]        start the proxy server (foreground)
  login                                sign in to GitHub via Device Flow OAuth
  config <claude|codex|gemini>         configure a client tool
         [-o <path>]                   override the settings output path
  usage [-m <YYYY-MM>]                 show token usage statistics
  logs  [-l <n>] [-e] [--model <m>]    show recent request logs
        [-d <YYYY-MM-DD>]
  service <install|uninstall|reinstall>
                                       manage the systemd user service (Linux/WSL)

Run copilot-proxy --help for the full listing with examples.

`logs` — inspect requests

[13:34:22] 200  claude-opus-4.7          /v1/messages       1 in   12 cc   84 cr  938 out   0 r  21.2s
 │          │    │                        │                  │      │       │      │         └─ reasoning tokens
 │          │    │                        │                  │      │       │      └─ output tokens
 │          │    │                        │                  │      │       └─ cache read tokens
 │          │    │                        │                  │      └─ cache creation tokens
 │          │    │                        │                  └─ fresh input tokens
 │          │    │                        └─ endpoint
 │          │    └─ model (translated name when aliased)
 │          └─ HTTP status
 └─ request time

Request logs live at ~/.copilot-proxy/logs/requests/YYYY-MM-DD.jsonl (180-day retention by default; cleanup runs on server startup).

References

License

MIT

@ascdong/copilot-proxy