[FEATURE] Add support for Github Copilot SDK

bump

@pranjalsinha99 @HeegonJin you don't need to wait on first-party SDK support for this - there's a way today that does exactly what you're after: use your Copilot sub (incl. its Claude models) inside Claude Code, no separate Anthropic key.

I maintain a fork that's a local proxy exposing Copilot as a native Anthropic endpoint. point Claude Code at it:

npx copilot-api-node20@latest start

then set ANTHROPIC_BASE_URL=http://localhost:4141 (+ any non-empty ANTHROPIC_API_KEY) and run claude as usual. requests get forwarded straight to Copilot's native /v1/messages, so thinking, tool use and usage reporting all come through afaict.

full disclosure it's my own fork, not anything official - been running it ~9 months for a few folks on my team. you're welcome to give it a try: https://www.npmjs.com/package/copilot-api-node20

You can already do this without a proxy: Copilot exposes an Anthropic-compatible /v1/messages endpoint, and Claude Code can talk to it directly. I've switched from copilot-api to this setup (Copilot Enterprise, Claude Code 2.1.285).

Setup

1. Log in with the GitHub CLI (gh auth login) using an account that has a Copilot seat. Copilot's API accepts the gh token directly as a Bearer token.
2. ~/.claude/settings.json:

{
  "apiKeyHelper": "gh auth token",
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.enterprise.githubcopilot.com",
    "CLAUDE_CODE_ATTRIBUTION_HEADER": "0",
    "ENABLE_TOOL_SEARCH": "true"
  }
}

• ANTHROPIC_BASE_URL: this is the Copilot Enterprise host. For other plans the host is probably https://api.githubcopilot.com, but I haven't tested that.
• CLAUDE_CODE_ATTRIBUTION_HEADER=0: needed for prompt caching. Without it, cache_read stays at 0 and every turn is billed as fresh input.
• ENABLE_TOOL_SEARCH=true: optional. It's off by default with a custom base URL, but works on Copilot and cut my cached prompt prefix from about 38k to 27k tokens.

WebSearch

Copilot rejects Anthropic's server-side web search tool.

Workaround: a PreToolUse hook on WebSearch that runs the query through Copilot's /responses endpoint with its web_search tool on a GPT model (that endpoint only accepts OpenAI models; the same gh token works there), then returns the answer and sources to Claude as the hook's deny reason. Claude uses it as the search result.

WebFetch works as-is, because Claude Code fetches pages client-side.