Your First Conversation

This walkthrough takes you from a freshly installed plugin to your first working exchange with the AI agent inside Unreal Engine — using the embedded ClaudusCode terminal, the primary way to use ClaudusBridge.

The terminal runs a full coding agent (a Claude Code build that also drives ChatGPT/Codex and Gemini) headless inside the editor, backed by a real PTY (ConPTY on Windows, forkpty on macOS/Linux). The agent is already wired to the bridge, so it can see and drive the editor through the same 900+ MCP tools an external client would use — no extra setup.

Before you start

• ClaudusBridge is installed and enabled (see Installation).
• A system Node.js install is on your PATH (node --version works). The embedded agent launches via your system node.
• You have a Claude (Anthropic), ChatGPT/Codex (OpenAI), or Gemini (Google) account to sign in with. No API keys are required.

1. Open the ClaudusCode terminal

In the editor's main menu bar, open Tools → ClaudusCode. The menu entry shows the live tool count, for example ClaudusCode (900 tools).

A docked terminal panel appears. This is a real interactive terminal — the bundled agent runs inside it. You can drag the tab anywhere in your editor layout and it will be remembered.

2. Sign in to your provider

The first time the agent runs, it walks you through a normal OAuth login for your chosen provider:

1. Start the agent in the terminal (it launches on first interaction; follow the on-screen prompt if it asks which provider to use).
2. The agent opens — or prints — a sign-in URL. Complete the login in your browser with your own Claude, ChatGPT/Codex, or Gemini account.
3. Return to the editor. The agent is now authenticated.

Everything is 100% local: the only data that leaves your machine is your provider's own login and model calls. There are no API keys and no per-token billing from the plugin — you use your existing account.

Switching providers

The bundled agent can drive Claude, ChatGPT/Codex, and Gemini. If you want to use a different provider later, you can switch and sign in to that account from the terminal.

3. Send your first prompt

Type a request and press Enter. Good first prompts that exercise the bridge:

• What can you do in this editor?
• List the actors in the current level.
• Create a new Material called M_FirstTest and give it a red base color.
• Spawn a cube at the world origin.

When you ask an open-ended "what can you do," the agent answers by querying the live toolset list rather than guessing — so the capabilities you see reflect exactly what's available in your project.

4. What the agent does

Behind the scenes the agent drives the editor through MCP. In tool-search mode (the default) it:

1. Calls list_toolsets to see the available domains — Actors/World, Blueprints, Materials, Niagara, UMG, Sequencer, Assets, Levels, Textures, Multi-User, and the plugin's native toolsets (Asset, Level, Image, Chess).
2. Calls describe_toolset to read the exact tools and parameters in the toolset it needs.
3. Calls call_tool to execute — spawning actors, editing Blueprints, creating materials, and so on.

You'll see these steps stream in the terminal as the agent reasons, calls tools, and reports results. Changes show up live in the editor: new assets in the Content Browser, new actors in the viewport, edited graphs in their respective editors.

If the agent needs permission to make a change, it asks. The agent supports permission modes — default, acceptEdits, bypassPermissions, and plan — so you can let it move freely or keep it on a tighter leash depending on the task.

5. Where the transcript lives

The conversation streams in the ClaudusCode terminal panel. Because it's a real PTY-backed terminal, scrolling back through the panel shows the full session — the agent's messages, the tool calls it made, and their results.

For programmatic, terminal-free control there's also an ACP (Agent Client Protocol) path driven over stdio, exposed through console commands:

ClaudusCode.Acp.Start     # spawn and initialize the agent over ACP
ClaudusCode.Acp.Prompt    # send a prompt to the running agent
ClaudusCode.Acp.Stop      # stop the agent

ClaudusCode.Acp.Prompt will warn no agent running — run ClaudusCode.Acp.Start first if you prompt before starting. This path is handy for automation and for driving the agent from scripts; for day-to-day work, the ClaudusCode terminal is the place to be.

Next steps

• Connecting — attach an external MCP client (Claude Desktop, Cursor, Windsurf, VS Code) to the same bridge.
• Explore the toolsets — ask the agent to list_toolsets and try a Blueprint, Material, or Sequencer task.
• Multi-User — have the agent and a human join the same Concert session as separate participants, each with a presence avatar.