Copilot API Proxy

[!WARNING] Reverse-engineered proxy of the GitHub Copilot API. Not supported by GitHub and may break unexpectedly. Excessive automated use may trigger GitHub's abuse detection and risk your Copilot access. Use responsibly.

See: GitHub Acceptable Use Policies · GitHub Copilot Terms

A local proxy that exposes GitHub Copilot as OpenAI-compatible (/v1/chat/completions, /v1/models, /v1/embeddings) and Anthropic-compatible (/v1/messages, /v1/messages/count_tokens) endpoints. Use Copilot as a backend for any tool that speaks either API — including Claude Code.

Prerequisites

• Node.js (for npx) or Bun >= 1.2.x
• GitHub account with Copilot subscription (individual, business, or enterprise)

Quick start (npx)

The fastest way to run the proxy — no clone, no install:

npx copilot-bridge@latest start

Or install globally once and reuse the copilot-bridge (or short copx) binary:

npm i -g copilot-bridge
copilot-bridge start
# or
copx start

On first run you'll be prompted to authenticate with GitHub via device-code flow. The server then listens on http://localhost:4141.

Useful flags:

npx copilot-bridge start --port 8080                  # custom port
npx copilot-bridge start --host 0.0.0.0               # expose on the network (default: loopback only)
npx copilot-bridge start --account-type business      # force business / enterprise (default: auto-detect)
npx copilot-bridge start --rate-limit 30 --wait       # throttle requests
npx copilot-bridge start --manual                     # approve each request
npx copilot-bridge start --github-token ghp_...       # non-interactive auth
npx copilot-bridge auth                               # only mint a GitHub token

From source

If you want to hack on the code:

git clone https://github.com/albanx/copilot-proxy.git
cd copilot-proxy
bun install
bun run dev          # watch mode
# or
bun run build && bun run start

Use with Claude Code

With the proxy running, set these environment variables in the terminal where you launch Claude Code:

export ANTHROPIC_AUTH_TOKEN="llstudio"
export ANTHROPIC_API_KEY=""
export ANTHROPIC_BASE_URL="http://localhost:4141"

Then start Claude Code as usual. It will route all requests through the proxy to Copilot.

Switching models from inside Claude Code

If you only want to use Anthropic-family models (sonnet, opus, haiku), don't set ANTHROPIC_MODEL, ANTHROPIC_DEFAULT_OPUS_MODEL, or ANTHROPIC_DEFAULT_HAIKU_MODEL. With those unset, Claude Code uses its built-in defaults and the /model command works natively — pick sonnet/opus/haiku and the proxy translates the ID to Copilot's matching model (e.g. claude-sonnet-4-5 → claude-sonnet-4.5).

Only pin a model via env vars or settings.json if you want to lock to a specific Copilot model (including non-Anthropic ones like gpt-4.1, gpt-5), since Claude Code's /model picker doesn't list those.

Optional: configure via settings.json

To persist config across sessions, put it in Claude Code's settings file:

• Windows: C:\Users\<USER>\.claude\settings.json
• macOS / Linux: ~/.claude/settings.json

Sample (Anthropic-only, lets /model switch freely):

{
  "env": {
    "ANTHROPIC_AUTH_TOKEN": "lmstudio",
    "ANTHROPIC_BASE_URL": "http://localhost:4141",
    "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1",
    "CLAUDE_CODE_USE_POWERSHELL_TOOL": "1",
    "DISABLE_NON_ESSENTIAL_MODEL_CALLS": "1"
  },
  "forceLoginMethod": "console",
  "alwaysThinkingEnabled": true,
  "effortLevel": "xhigh",
  "autoUpdatesChannel": "latest",
  "skipDangerousModePermissionPrompt": true
}

To pin a specific model instead, add "ANTHROPIC_MODEL": "claude-opus-4.6" (or any ID returned by GET /v1/models) inside env.

Reasoning effort

The proxy honors the reasoning effort your client requests, on every routing path:

• Claude Code's "effortLevel" setting (or output_config.effort / a top-level reasoning_effort sent by any client) is forwarded to Copilot.
• The value is clamped to the effort ladder each model actually advertises (e.g. a model without xhigh gets its highest supported level instead of a 400).
• When the client sends nothing, the proxy imitates the VS Code Copilot extension's defaults: xhigh for GPT-5.3+ models, high otherwise.

The effort actually applied is printed on each request log line (effort=…).

Or use the interactive helper, which prompts for models and copies the full launch command to your clipboard:

npx copilot-bridge start --claude-code

CLI options (start)

| Option | Description | Default | Alias | | ---------------- | -------------------------------------------------------------- | ------------ | ----- | | --port | Port to listen on | 4141 | -p | | --host | Interface to bind (0.0.0.0 to expose on the network/Docker) | 127.0.0.1 | | | --verbose | Verbose logging | false | -v | | --account-type | auto, individual, business, or enterprise | auto | -a | | --manual | Approve each request manually | false | | | --rate-limit | Minimum seconds between requests | none | -r | | --wait | Wait on rate limit instead of erroring | false | -w | | --github-token | Provide GitHub token directly | none | -g | | --claude-code | Print/copy a Claude Code launch command after model selection | false | -c | | --show-token | Print GitHub/Copilot tokens on fetch + refresh | false | | | --proxy-env | Read HTTP_PROXY/HTTPS_PROXY from env | false | |

API endpoints

OpenAI-compatible:

| Endpoint | Method | Description | | --------------------------- | ------ | ------------------------------------------- | | /v1/chat/completions | POST | Chat completion (streaming + non-stream) | | /v1/responses | POST | Responses API (streaming + non-stream) | | /v1/models | GET | List available Copilot models | | /v1/embeddings | POST | Create embeddings |

Anthropic-compatible:

| Endpoint | Method | Description | | -------------------------------- | ------ | ------------------------------------------ | | /v1/messages | POST | Messages API (streaming + non-stream) | | /v1/messages/count_tokens | POST | Token count for a message payload |

Notes:

• All endpoints are also mounted without the /v1 prefix.
• Both /v1/chat/completions and /v1/messages accept any model exposed by Copilot — pick a Claude (claude-opus-4.6, claude-sonnet-4.5, …) or non-Claude (gpt-4.1, gpt-5, …) model from /v1/models. The /v1/messages route auto-translates Anthropic-style model IDs (e.g. claude-sonnet-4-5-20250929) into Copilot's dotted form, and passes Copilot-native IDs through unchanged.
• Account type is auto-detected from the Copilot token (individual / business / enterprise). Override with --account-type only if detection is wrong.

Security

This proxy hands out your GitHub Copilot subscription to whoever can reach it — there is no inbound authentication. Two defaults protect you:

• It binds to loopback (127.0.0.1) only. Nothing on your network can reach it. Pass --host 0.0.0.0 to expose it deliberately (Docker, LAN); the server prints a warning when you do, and you should put it behind a reverse proxy that authenticates.
• Browser requests are restricted to loopback origins. Without this, any website you visited while the proxy was running could POST to localhost:4141 from your browser, burn your quota and read the replies. Command-line clients (Claude Code, curl, the OpenAI/Anthropic SDKs) send no Origin header and are unaffected.

Your GitHub token is cached at ~/.local/share/copilot-api/github_token with 0600 permissions. --show-token prints tokens to the console — only use it when debugging, and don't paste the output anywhere.

Tips

• Use --rate-limit <seconds> (optionally with --wait) to avoid tripping Copilot's abuse detection.
• Account type is auto-detected by default. To force a specific plan, pass --account-type individual|business|enterprise. See GitHub's network routing docs.
• If you use opencode, you don't need this proxy — it supports Copilot natively.