GPT-4o and Claude-3.7-Sonnet APIs for coding.

Given that GitHub may enforce stricter policies, I want to minimize risks for users. As a result, I am transitioning the project to maintenance mode. The repository will remain available for reference, but complete removal is also a possibility, with no further notice. No major developments will be made. Please conduct your own due diligence and assess the risks before using this tool.

• Constantly hitting 429 (rate limit) can get your GH Copilot subscription suspended. Knowing this is important to protect yourself. Not just copilot-more, people get banned by using LM API too because of this, see reddit thread.
• It seems you can hit 429 quicker with the Claude 3.7 Sonnet models (probably smaller quotas + 3.7 runs faster than 3.5).
• copilot-more can now display token usage stats for your awareness. PR 41
• Highly recommend leveraging the advanced rate limiter features (both token- and request-based) to prevent excessive token usage.

If you want to know more about GH Copilot suspension risk, read this RooVetGit/Roo-Code#1203 (comment)

copilot-more maximizes the value of your GitHub Copilot subscription by exposing models like Claude-3.7-Sonnet for use in agentic coding tools such as Cline, or any tool that supports bring-your-own-model setups. Unlike costly pay-as-you-go APIs, this approach lets you leverage these powerful models affordably.

• Respect the GitHub Copilot terms of service.
• Only use the API for coding tasks.
• Be mindful of the risk of being banned by GitHub Copilot for misuse.

1. Get the refresh token

   A refresh token is used to get the access token. This token should never be shared with anyone :). You can get the refresh token by following the steps below:

   • Run the following command and note down the returned device_code and user_code.:

   # 01ab8ac9400c4e429b23 is the client_id for the VS Code
   curl https://github.com/login/device/code -X POST -d 'client_id=01ab8ac9400c4e429b23&scope=user:email'

   • Open https://github.com/login/device/ and enter the user_code.

   • Replace YOUR_DEVICE_CODE with the device_code obtained earlier and run:

   curl https://github.com/login/oauth/access_token -X POST -d 'client_id=01ab8ac9400c4e429b23&scope=user:email&device_code=YOUR_DEVICE_CODE&grant_type=urn:ietf:params:oauth:grant-type:device_code'

   • Note down the access_token starting with gho_.

2. Install and run copilot_more

• Bare metal installation:

  git clone https://github.com/jjleng/copilot-more.git
  cd copilot-more
  # install dependencies
  poetry install
  # run the server. Replace gho_xxxxx with the refresh token you got in the previous step. Note, you can use any port number you want.
  REFRESH_TOKEN=gho_xxxxx poetry run uvicorn copilot_more.server:app --port 15432

• Docker Compose installation:

  git clone https://github.com/jjleng/copilot-more.git
  cd copilot-more
  # run the server. Ensure you either have the refresh token in the .env file or pass it as an environment variable.
  docker-compose up --build

3. Alternatively, use the refresh-token.sh script to automate the above.

The application allows you to customize behavior through environment variables or a .env file. Available configuration options:

See .env.example for a template configuration file. You can cp .env.example .env and modify the values as needed.

Once you have set up your .env file with all your configuration settings, you can simply run the server without specifying environment variables on the command line:

poetry run uvicorn copilot_more.server:app --port 15432

Rate limiting is optional and only applied to models that you explicitly configure. You can define rate limits for specific models using a rate_limits.json file in the project root directory:

{
  "claude-3.7-sonnet": [
    {
      "window_minutes": 1,
      "total_tokens": 50000,
      "input_tokens": 50000,
      "output_tokens": 5000,
      "requests": 5,
      "behavior": "delay"
    },
    {
      "window_minutes": 60,
      "total_tokens": 500000,
      "input_tokens": 500000,
      "output_tokens": 50000,
      "requests": 100,
      "behavior": "error"
    }
  ]
}

Configuration options:

• window_minutes: Time window in minutes
• total_tokens: Max total tokens in window (optional)
• input_tokens: Max input tokens in window (optional)
• output_tokens: Max output tokens in window (optional)
• requests: Max requests in window (optional)
• behavior: What to do when limit is hit: "delay" or "error"

⚠️ Warning: The default rate_limits.json is just an example and not necessarily suitable for production use. You should adjust these limits based on your actual usage patterns.

Notes:

• Rate limits are only applied to models listed in the configuration file
• Models not listed in the file will have no rate limits
• You must specify at least one of: total_tokens, input_tokens, output_tokens, or requests
• Changes to rate limits require restarting the server to take effect

While rate limits help control usage within time windows, sometimes you may need finer control over request spacing. The SLEEP_BETWEEN_CALLS setting introduces a fixed delay between API calls, which can help prevent burst requests when the API responds very quickly. This is particularly useful when:

• You want to ensure a minimum time gap between requests regardless of response speed
• You need to prevent rapid successive requests that might trigger rate limits
• You want to maintain a more consistent, predictable request pattern

Example: Setting SLEEP_BETWEEN_CALLS=1.0 ensures at least 1 second between each API call, even if the API responds faster.

Now you can connect Cline or any other AI client to http://localhost:15432 and start coding with the power of GPT-4o and Claude-3.5-Sonnet without worrying about the cost. Note, the copilot-more manages the access token, you can use whatever string as API keys if Cline or the AI tools ask for one.

1. Install Cline code --install-extension saoudrizwan.claude-dev
2. Open Cline and go to the settings
3. Set the following:
   • API Provider: OpenAI Compatible
   • API URL: http://localhost:15432
   • API Key: anything
   • Model: gpt-4o, claude-3.7-sonnet, o1, o3-mini

For troubleshooting integration issues, you can enable traffic logging to inspect the API requests and responses.

To enable logging, set the RECORD_TRAFFIC environment variable to true:

RECORD_TRAFFIC=true REFRESH_TOKEN=gho_xxxx poetry run uvicorn copilot_more.server:app --port 15432

Alternatively, you can add RECORD_TRAFFIC=true to your .env file.

All traffic will be logged to files in the current directory with the naming pattern: copilot_traffic_YYYYMMDD_HHMMSS.mitm

Attach this file when reporting issues. Please zip the original file that ends with the '.mitm' extension and upload to the GH issues.

Note: the Authorization header has been redacted, so the refresh token won't be leaked.

The GH Copilot models sit behind an API server that is not fully compatible with the OpenAI API. You cannot pass in a message like this:

    {
      "role": "user",
      "content": [
        {
          "type": "text",
          "text": "<task>\nreview the code\n</task>"
        },
        {
          "type": "text",
          "text": "<task>\nreview the code carefully\n</task>"
        }
      ]
    }

copilot-more takes care of this limitation by converting the message to a format that the GH Copilot API understands. However, without the type, we cannot leverage the models' vision capabilities, so that you cannot do screenshot analysis.