This project is no longer maintained. (Due to changes introduced by Copilot, the repository will only function with the chat window, not the Edit window.)

Copilot Proxy

Make VS Code GitHub Copilot work with any open weight models, Llama3, DeepSeek-Coder, StarCoder, etc. This should work even you are not a Copilot subscriber.

🌟 Motivations

• I'm already familiar with and enjoy using the GitHub Copilot extension (yes, I know there are other awesome extensions, such as Continue.).
• Copilot may not always utilize the latest GPT models. It currently uses models like gpt-4-0125-preview, gpt-3.5-turbo and others.
• Transferring code from the editor to ChatGPT to use GPT-4o is inconvenient.
• I'm interested in using alternative models such as Llama3, DeepSeek-Coder, StarCoder, and Sonnet 3.5.
• I have subscriptions to both ChatGPT and Copilot but would like to cancel my Copilot subscription.

🏃‍♂️ How to Run

1. Install copilot_proxy

   git clone https://github.com/jjleng/copilot-proxy.git
   cd copilot-proxy
   # install dependencies
   poetry install
   # build and install copilot_proxy as a runnable script
   poetry build && pip uninstall copilot_proxy -y && ls dist/*.whl | sort -V | tail -n 1 | xargs pip install
   # verify that the script has been installed successfully
   copilot_proxy --version

2. Run copilot_proxy with Ollama models, OpenRouter models or any OpenAI API compatible endpoints

    # Ollama
    MODEL_URL="http://localhost:11434/v1/chat/completions" MODEL_NAME="llama3:instruct" MODEL_API_KEY="whatever" copilot_proxy start
    # OpenRouter
    MODEL_URL="https://openrouter.ai/api/v1/chat/completions" MODEL_NAME="deepseek/deepseek-coder" MODEL_API_KEY="YOUR_KEY" copilot_proxy start

3. Install mitmproxy certificate

   copilot_proxy uses mitmproxy to proxy the Copilot traffic. You need to install the mitmproxy certificate. After first run, you should find cert files under ~/.mitmproxy. See https://docs.mitmproxy.org/stable/concepts-certificates/ for details.

   Quick guides to import the CA certificate to your operating system.

   Windows (not verified)

   • Open the Windows Start menu and type mmc to open the Microsoft Management Console.
   • Go to File -> Add/Remove Snap-in....
   • Select Certificates and click Add.
   • Choose Computer account and click Next.
   • Select Local computer and click Finish.
   • Click OK to close the snap-in window.
   • In the console, expand Certificates (Local Computer) -> Trusted Root Certification Authorities -> Certificates.
   • Right-click on Certificates and choose All Tasks -> Import.
   • Follow the import wizard, select Browse, and navigate to the mitmproxy-ca-cert.pem file. Make sure to select All Files to see the .pem file.
   • Finish the import process.

   Mac (verified)

   • Open Keychain Access from Applications -> Utilities.
   • Import it into the System keychain.
   • In the File menu, select Import Items.
   • Navigate to ~/.mitmproxy and select the mitmproxy-ca-cert.pem file.
   • Find the imported certificate in the list, right-click it, and select Get Info.
   • Expand the Trust section and set When using this certificate to Always Trust.
   • Close the info window and authenticate with your password to apply the changes.

   Linux (not verified)

   • Copy the mitmproxy-ca-cert.pem file to the system certificate directory:

     sudo cp ~/.mitmproxy/mitmproxy-ca-cert.pem /usr/local/share/ca-certificates/mitmproxy-ca-cert.crt

   • Update the certificate store:

     sudo update-ca-certificates

   Restart copilot_proxy

4. Setup the proxy in VS Code

   Follow the guide here: https://docs.github.com/en/copilot/managing-copilot/configure-personal-settings/configuring-network-settings-for-github-copilot

   

✨ Code Completion

For code completions, the best result will be calling the model through the FIM (fill-in-the-middle) template. This can be done by calling the /completion endpoint with the suffix parameter. If your inference server is compatible with OpenAI /completion endpoint and supports the suffix parameter, you can get better completions by using the code in the completion branch. This setup will produce the best result.

NOTICE: Ollama doesn't support the /completion endpoint and vLLM doesn't support the suffix parameter.

Code from the main branch works well with Copilot chat, but might not produce high quality completions. However, it is agnostic about the inference servers.

🤔 Why not go the simpler route?

In VS Code settings, we can

{
  "github.copilot.advanced": {
    "debug.testOverrideProxyUrl": "http://localhost:11434",
    "debug.overrideProxyUrl": "http://localhost:11434"
  }
}

This did not work for me.

• Only code completion triggers the endpoint, chatbox does not.
• Seems I need to keep the Copilot subscription?
• Flexibility.