AI Agents

Run AI agents directly inside the Unreal Editor. Each agent is a CLI-based AI that has full access to all plugin tools via the MCP protocol.

Supported Agents

Claude Code (Anthropic)

• Install: npm install -g @anthropic-ai/claude-code
• Auth: claude in terminal, follow the login flow
• Requires: Node.js 18+, Anthropic account (Claude Max subscription or API credits)
• Models: claude-sonnet-4-6, claude-opus-4-6, claude-haiku-4-5
• Thinking levels: Low, Medium, High

Codex (OpenAI)

• Install: npm install -g @openai/codex
• Auth: Set your OpenAI API key
• Requires: Node.js 18+, OpenAI API key
• Models: Various OpenAI models

GitHub Copilot

• Install: npm install -g @github/copilot then pip install github-copilot-sdk
• Auth: gh auth login
• Requires: Node.js 18+, GitHub Copilot subscription, GitHub CLI
• Note: Both the npm CLI and Python SDK are required

Gemini CLI (Google)

• Install: npm install -g @google/gemini-cli
• Auth: gemini in terminal, follow the Google login flow
• Requires: Node.js 18+, Google account (AI Pro or AI Ultra subscription for best models)

Setup in the Plugin

1. Go to Settings > AI Models
2. Click the “AI Agent” tab
3. Select the agent you want to use (Claude Agent, Codex, Copilot, Gemini Agent)
4. If the agent shows “Not found”, click “Install” to install the CLI automatically
5. If authentication is needed, follow the instructions shown
6. Select your preferred model and thinking level
7. Click “Save Slot”

How Agents Work

When you send a message in the Architect view with an agent selected:

1. The plugin spawns a Python agent runner process
2. The agent connects to the MCP server (port 9877)
3. Your message is sent to the agent
4. The agent reasons about your request and calls tools to execute actions in the editor
5. Tool results stream back in real-time, visible as bubbles in the chat
6. The agent writes its final response

Agent Features

• Thinking blocks: See the agent’s reasoning process (displayed as collapsible “Internal reasoning” sections)
• Ask Before Edit: Agents respect the interaction mode. In “Ask Before Edit”, you confirm each tool call
• Concurrent agents: Multiple agent instances can run simultaneously on different conversations
• Per-agent model selection: Each agent type has its own model dropdown with available options
• Thinking effort control: Adjust how deeply the agent reasons (Low/Medium/High)

Troubleshooting

Agent shows “Not found”

The CLI tool isn’t installed or isn’t in your system PATH. Click “Install” or install manually in a terminal.

Agent shows “Not authenticated”

Run the agent’s auth command in a terminal (e.g., claude for Claude Code, gh auth login for Copilot).

Mac: Agent not detected despite being installed in terminal

UE launched from Finder doesn’t inherit your terminal’s PATH. The plugin probes common install locations automatically. If your agent is installed in a non-standard location, try launching UE from the terminal instead.