litellm-copilot

A LiteLLM proxy that lets you use GitHub Copilot models via the OpenAI-compatible API. Point any tool that supports ANTHROPIC_BASE_URL or OPENAI_BASE_URL at this proxy to use Copilot models.

Unlike other "copilot-proxy" projects that implement custom proxy servers, this is just a thin wrapper script and config for LiteLLM's built-in github_copilot provider.

Prerequisites

• uv (Python package runner)
• A GitHub account with Copilot access
• A secret manager (1Password CLI or macOS Keychain)

Setup

1. Install uv

macOS:

brew install uv

Linux:

curl -LsSf https://astral.sh/uv/install.sh | sh

2. Create a GitHub Personal Access Token

Create a PAT with the copilot scope:

https://github.com/settings/tokens/new?scopes=copilot&description=LiteLLM%20Proxy

3. Store your token

Store the token in your preferred secret manager.

1Password:

# Store in 1Password (via web UI or CLI), then reference it as:
op://YourVault/YourItem/token

macOS Keychain:

security add-generic-password -a "$USER" -s copilot-token -w "ghp_your_token_here"

4. Configure environment variables

Add to your shell profile (~/.zshrc, ~/.bashrc, etc.):

# Required: Command to retrieve your GitHub PAT
# Pick ONE of these based on your secret manager:

# 1Password:
export COPILOT_PROXY_TOKEN_CMD='op read "op://YourVault/YourItem/token"'

# macOS Keychain:
export COPILOT_PROXY_TOKEN_CMD='security find-generic-password -a "$USER" -s copilot-token -w'

5. Install the script

Clone this repo and add it to your PATH:

git clone https://github.com/jonmagic/litellm-copilot.git ~/.litellm-copilot

Then either symlink the script:

ln -s ~/.litellm-copilot/litellm-copilot /usr/local/bin/litellm-copilot

Or add the directory to your PATH:

export PATH="$HOME/.litellm-copilot:$PATH"

Usage

Start the proxy:

litellm-copilot

To use a custom port:

litellm-copilot --port 8080

To use LiteLLM nightly from GitHub main:

litellm-copilot --nightly

The proxy runs on http://localhost:4000 by default.

Using with AI Agents

Configure your favorite AI tools to use the proxy. Add these to your shell profile or set them before running the tool.

Claude Code

Install:

npm install -g @anthropic-ai/claude-code

Configure:

export ANTHROPIC_BASE_URL="http://localhost:4000"
export ANTHROPIC_AUTH_TOKEN="fake-key"  # Required but not used

Test:

claude -p "What is 2+2?"

OpenCode

Install:

npm install -g opencode@latest

OpenCode has built-in GitHub Copilot support — no proxy needed:

opencode auth login

Gemini CLI

Note: Gemini CLI requires a fix in LiteLLM that forwards extra_headers through the generateContent adapter. This fix is pending in PR #18935. Until it's merged, you can run the proxy using a patched version:

# Clone the patched fork
git clone https://github.com/jonmagic/litellm.git ~/litellm-patched

# Create a venv and install (Python 3.13 required, uvloop doesn't support 3.14+)
python3.13 -m venv ~/litellm-patched/.venv
~/litellm-patched/.venv/bin/pip install -e ~/litellm-patched[proxy]

# Run the proxy using the patched version
GITHUB_TOKEN=$(eval "$COPILOT_PROXY_TOKEN_CMD") ~/litellm-patched/.venv/bin/litellm --config ~/.litellm-copilot/config.yaml --port 4000

Install:

npm install -g @google/gemini-cli@preview

Configure:

export GOOGLE_GEMINI_BASE_URL="http://localhost:4000"
export GEMINI_API_KEY="fake-key"  # Required but not used

Test:

gemini -p "What is 2+2?"

Aider

Install:

pip install aider-chat

Configure:

export AIDER_OPENAI_API_BASE="http://localhost:4000"
export OPENAI_API_KEY="fake-key"  # Required but not used

Test:

aider --model openai/claude-sonnet-4 --message "What is 2+2?"

llm (Simon Willison's LLM CLI)

Install llm:

Homebrew (recommended):

brew install llm

pip:

pip install llm

Install the Azure plugin:

llm install llm-azure

Verify the plugin is installed:

llm plugins
# Should show llm-azure in the list

Configure a dummy API key (llm requires an API key even though the proxy doesn't use it):

llm keys set litellm
# Enter: fake-api-key

Add model configurations to ~/.config/io.datasette.llm/azure/config.yaml:

# Local litellm proxy models
- model_id: copilot-gpt-4.1
  provider: openai
  model_name: gpt-4.1
  endpoint: http://localhost:4000/v1/
  api_key_name: litellm
- model_id: copilot-claude-sonnet-4.5
  provider: openai
  model_name: claude-sonnet-4.5
  endpoint: http://localhost:4000/v1/
  api_key_name: litellm

Note: The copilot- prefix helps distinguish litellm-copilot models from other configured models in llm. You can add any of the models from the Supported Models section below using this pattern.

Set your default model:

llm models default copilot-gpt-4.1

Test:

llm "What is 2+2?"

Supported Models

Last updated: 2026-02-06

Model	Vendor	Category
claude-haiku-4.5	Anthropic	versatile
claude-opus-4.5	Anthropic	powerful
claude-opus-4.6	Anthropic	powerful
claude-sonnet-4	Anthropic	versatile
claude-sonnet-4.5	Anthropic	versatile
gemini-2.5-pro	Google	powerful
gemini-3-flash-preview	Google	lightweight
gemini-3-pro-preview	Google	powerful
gpt-4.1	Azure OpenAI	versatile
gpt-4o	Azure OpenAI	versatile
gpt-5	Azure OpenAI	versatile
gpt-5-codex	OpenAI	powerful
gpt-5-mini	Azure OpenAI	lightweight
gpt-5.1	OpenAI	versatile
gpt-5.1-codex	OpenAI	powerful
gpt-5.1-codex-max	OpenAI	powerful
gpt-5.1-codex-mini	OpenAI	powerful
gpt-5.2	OpenAI	versatile
gpt-5.2-codex	OpenAI	powerful
grok-code-fast-1	xAI	lightweight

License

ISC