Turning GitHub Copilot CLI into an AI Agent via ACP

I wanted an AI agent that could check the news, emails, and my calendar. While OpenClaw is currently popular, dealing with its security and permission requirements can be a hassle.

By operating the familiar GitHub Copilot CLI through scheduled execution scripts or Discord, I was able to achieve what I wanted. Receiving execution results periodically is convenient, but being able to ask follow-up questions or give instructions for deeper dives directly is even better.

GitHub Repository: Lunran/acp-client

Step 1: Launch GitHub Copilot CLI as an ACP Server

First, start the GitHub Copilot CLI as an ACP server.

copilot --acp --port 8100

Running it within a sandbox can improve safety.

Reference: https://zenn.dev/lunran/scraps/5105de92cb9687

docker ps -a --format '{{.Names}}' | grep -q "^copilot-acp-container\$" || \
docker start -ai copilot-acp-container || \
docker run -it \
  --name copilot-acp-container \
  -p 8100:8100 \
  -v $(pwd):/workspace \
  -v ./.copilot:/home/agent/.copilot \
  -e GITHUB_TOKEN=$GITHUB_TOKEN \
  copilot-sandbox \
  copilot --acp --port 8100 --autopilot --yolo --model gpt-5-mini

The ACP server seems to expect coding tools as clients, but by following the protocol, you can connect it to any tool.

Copilot CLI ACP server - GitHub Docs

Step 2: Configure Discord and Start the ACP Client

Next, configure the Discord settings and start the ACP client.

git clone https://github.com/Lunran/acp-client.git
cd acp-client
cp .env.example .env

Enter your Discord settings in the .env file.

uv sync
uv run python main.py

Note: Although a model is specified when starting the ACP server, for some reason the default model (Claude Sonnet 4.6) often ends up being used. Therefore, the client includes a process to explicitly set the model during startup.