Copilot support?

[marcinkubica]

Hi!

Wondering what's your thoughts on adding Github Copilot support.

Currently available only in opencode and a vibes copilot-api

[Nayjest]

Hello,
👌
Investigating GitHub Copilot's possibilities on using alternative models / APIs

[Nayjest]

1. I tried GitHub Copilot in PyCharm → I simply don’t have the option to switch the model as described in the documentation 🤔: https://docs.github.com/en/enterprise-cloud@latest/copilot/how-tos/use-ai-models/change-the-chat-model?tool=jetbrains

2. I tried I tried GitHub Copilot in fresh VS Code installation →
https://docs.github.com/en/copilot/how-tos/use-ai-models/change-the-chat-model?tool=vscode
I successfully configured LM-proxy via AI Toolkit, as it's described in Copilot's documentation and it works via AI toolkit playground.
One nuance: AI Toolkit for VS Code requires entering the chat completions endpoint instead of the API base URL — i.e., you should enter something like http://127.0.0.1:8000/v1/chat/completions.

However, the configured model does not appear in the Copilot Chat UI model selector as described in the documentation. 😳

It feels like Copilot’s documentation is AI-generated and doesn’t reflect the actual features. 😵

I will try to return to this later.

@marcinkubica Meanwhile, Could you clarify which environment you’re using GitHub Copilot in
and where you see the option to use custom models / LLM APIs?

[marcinkubica]

@Nayjest I do use Copilot with VSCode, but also now terminal (hence opencode etc).

About custom models - I'd leave that bit for the time being.

Try curl https://api.githubcopilot.com/models -H "Authorization: Bearer $(gh auth token)"

[Nayjest]

@marcinkubica
I've got it working!
Was able to connect GitHub Copilot via LM-Proxy to any model.
For educational purposes only, of course.
Polishing the required code changes now.

[Nayjest]

[Nayjest]

@marcinkubica the following setup works for GitHub Copilot in JetBrains.

Please note that this software is not intended to be used with random language models.
Use the example setup below for educational purposes only.

1. Redirect requests to original LLM (for example, api.business.githubcopilot.com) to the target host (127.0.0.1) using the OS hosts file.
2. Install LM-Proxy v2.1.0 or later.
3. Create a certificate for the target domain and place cert.pem and key.pem in the destination specified in the LM-Proxy configuration (ssl_certfile, ssl_keyfile). In this example, they are placed in the same folder as the LM-Proxy configuration.
4. Edit and save the configuration shown below as your_config.py, specifying your LLM connection, environment variable with the API key, etc.
5. Start LM-Proxy:

lm_proxy --config your_config.py

import os
from dotenv import load_dotenv

from lm_proxy.config import Config, Group
from lm_proxy.bootstrap import env

load_dotenv(".env")

config = Config(
    port=443,
    ssl_certfile="cert.pem",
    ssl_keyfile="key.pem",
    api_prefix="",  # this is important, removes "/v1" from URLs
    api_key_check="lm_proxy.api_key_check.AllowAll",  # ignore API Key used by Copilot
    connections=dict(  # LLM of your choice, or python function for pre-processing prompts / post-processing LLM response
        anthropic={
            "api_type": "anthropic",
            "api_key": os.getenv("ANTHROPIC_API_KEY"),
            "api_base": "https://api.anthropic.com",
            "model": "claude-opus-4-1",
        },
    ),
    routing={"gpt-4.1": "anthropic.claude-opus-4-1"},
    # "capabilities" section of model-info is required for Copilot
    model_info={
        "gpt-4.1": {
            "capabilities": {
                "family": "gpt-4.1",
                "limits": {
                    "max_context_window_tokens": 128000,
                    "max_output_tokens": 16384,
                    "max_prompt_tokens": 128000,
                    "vision": {
                        "max_prompt_image_size": 3145728,
                        "max_prompt_images": 1,
                        "supported_media_types": [
                            "image/jpeg",
                            "image/png",
                            "image/webp",
                            "image/gif"
                        ]
                    }
                },
                "object": "model_capabilities",
                "supports": {
                    "parallel_tool_calls": True,
                    "streaming": True,
                    "structured_outputs": True,
                    "tool_calls": True
                },
                "tokenizer": "o200k_base",
                "type": "chat"
            }
        }
    }
)

Additionally here is example showing how you can override system prompt:

# use it in connections instead of dictionary with LLM connection details
import microcore as mc
async def custom_llm(*args, **kwargs):
    kwargs.pop("max_tokens", None) # remove max_tokens to avoid errors with Opus model
    kwargs.pop("top_p", None)
    args[0][0] = mc.SysMsg(  # replace system message
        "ALWAYS ANSWER IN CAPITAL LETTERS"
    )
    return await env.connections["anthropic"](*args, **kwargs)

[Nayjest]

See https://github.com/Nayjest/lm-proxy/releases/tag/v2.1.0