Introduction
Running your own inference platform is only half the fun if nobody can use it from their editor. With Bring your own Key (BYOK), VS Code lets you plug any OpenAI-compatible endpoint, like a LiteLLM gateway, directly into the Copilot Chat. This post shows the setup in a few steps.
Info: The screenshots were taken with VS Code version 1.147.0. The UI might look slightly different in other versions.
Step 1: Open the Chat and add a model
After starting VS Code, the Chat view is shown on the right. In the box "Choose how you want to use Copilot" you can either sign in to use the GitHub Copilot models or add your own. Click on Add Models.
Hint: If you don't see the Copilot Chat, press Ctrl + Shift + I (German keyboard: Strg + Shift + I) to open it.
Step 2: Choose a provider
The Language Models dialog opens. Click on Add Models again and you get a list of supported providers like Anthropic, Azure, Google or OpenAI. For a self-hosted platform select Custom Endpoint.
Step 3: Name the group
Next, VS Code asks for a Group Name. This is just the display name under which your models are grouped, e.g. Private AI Platform. Confirm with Enter.
Step 4: Enter the API key
Now enter the API key for your endpoint, e.g. the LiteLLM virtual key you created for your user, and confirm with Enter. The key is stored as a secret by VS Code and is not written to disk in plain text.
Step 5: Configure the models
Finally VS Code opens the chatLanguageModels.json file in your user profile. The provider entry with the name, vendor customendpoint and a reference to the stored secret is already there. You only have to fill in the model details:
[
{
"name": "Private AI Platform",
"vendor": "customendpoint",
"apiKey": "${input:chat.lm.secret.c45e343}",
"apiType": "chat-completions",
"models": [
{
"id": "<model id as exposed by your endpoint>",
"name": "<display name in VS Code>",
"url": "https://<your-endpoint>/v1/chat/completions",
"toolCalling": true,
"vision": true,
"maxInputTokens": 128000,
"maxOutputTokens": 16000
}
]
}
]
- id: the model name your endpoint expects in the request
- name: how the model is shown in the model picker
- url: the endpoint of your platform
- toolCalling / vision: set according to what your model supports. Tool calling is required for Agent mode
- maxInputTokens / maxOutputTokens: should match the context window of your deployment
After saving the file, the model shows up in the model picker of the Chat and is ready to use.
Conclusion
With just a few clicks and a small JSON file, VS Code talks to your own inference platform. Your code and prompts never leave your infrastructure, and you can still use the full Copilot Chat experience.





Top comments (0)