Copilot Adapter

Use GitHub Copilot models in Codex and Claude Code through a local adapter. @alialf/copilot-adapter translates the APIs used by those clients into official @github/copilot-sdk sessions.

The adapter runs only on your computer, binds to 127.0.0.1, and limits each process to the project you choose.

Requirements

• Node.js 22.12 or newer
• A GitHub Copilot Business or Enterprise seat
• GitHub Copilot CLI enabled by your organization
• GitHub CLI authentication (gh auth login)
• Codex or Claude Code

Quick start

Install the adapter and check your setup:

npm install --global @alialf/copilot-adapter
copilot-adapter doctor

Start it from the project you want your client to access:

cd /path/to/your/project
copilot-adapter start --cwd "$PWD" --policy read-only

Keep that terminal open. In another terminal, confirm that the adapter is running and list the models available to your Copilot account:

copilot-adapter status --models

Then connect a client.

Codex CLI

copilot-adapter setup codex --scope profile
codex --profile copilot

Use /model and /reasoning inside Codex to change the model and reasoning level.

Codex Desktop or VS Code

copilot-adapter setup codex --scope global

Restart Codex Desktop or reload VS Code, then open the same project used with --cwd. Only local tasks can connect to the adapter.

Claude Code CLI

copilot-adapter setup claude-cli
claude

Run /status to confirm the base URL is http://127.0.0.1:4141. Use /model and /effort to change settings.

Claude Code for VS Code

copilot-adapter setup claude-vscode

Reload VS Code, then open the same project used with --cwd.

For Claude Desktop and detailed setup instructions, see the client setup guide.

Access policies

The policy controls which tools Copilot may use inside the selected project:

• chat: no project tools
• read-only: read, search, and list files (default)
• workspace-write: also create and edit files
• unsafe-shell: also run shell commands
• unsafe-all: also use configured MCP servers

Start with read-only. Use the unsafe-* policies only when you trust the client and understand the additional access.

Useful commands

copilot-adapter status --models   # Check the adapter and available models
copilot-adapter models sync       # Refresh Codex models after a policy change
copilot-adapter token             # Print the local client token
copilot-adapter token --rotate    # Replace the local client token
copilot-adapter --help            # Show all commands

Run a setup command with --remove to undo its changes, for example:

copilot-adapter setup codex --scope profile --remove
copilot-adapter setup claude-cli --remove

Security and limitations

• The adapter accepts connections only from localhost. Do not expose it through a tunnel or public proxy.
• One adapter process serves one project. Use another port and process for a second project.
• The local bearer token protects the adapter; it is not a GitHub token.
• Codex and Claude permission dialogs do not control tools run by the Copilot SDK. The adapter policy does.
• unsafe-shell is not a hardened sandbox.

Documentation

• Client setup
• Command reference
• Architecture
• Testing
• Publishing

License

MIT. The protocol translation layer is based on the MIT-licensed copilot-sdk-proxy.