FAQ
Short answers. Click a question to open it.
What is MCP?
The Model Context Protocol is an open standard that lets AI applications use external tools. This project is an MCP server for the Unreal Editor, so any MCP client can use it.
Which AI clients work?
Any MCP client. Claude Code, Claude Desktop, Cursor and VS Code are covered step by step in 🔌 Connecting Clients. A client needs either Streamable HTTP with custom headers (native route) or the ability to launch a local command (stdio route).
Which models work?
Any model your client runs that can call tools. The single-tool design is meant to help smaller models too: they search for what they need, and every reply includes the exact next call to make. Stronger models still make fewer mistakes on long, multi-step jobs.
Why is there only one tool?
Listing about 400 actions, each with a full schema, would use up much of a model's context before it did anything, and would push many clients past their tool limits. With one tool, the model searches, loads the one contract it needs, and runs it. See 🧭 Using the Gateway.
Is it free?
Yes, under the MIT license. You pay only for whatever your AI client or model costs.
Which Unreal Engine versions and platforms are supported?
UE 5.0 to 5.8 on Windows, macOS and Linux. The plugin is editor-only.
Can it be used in a shipped game?
No. It is an editor plugin, and its modules are never packaged into your game.
Does it work with Blueprint-only projects?
The plugin compiles from source, which needs a C++ project. Add any C++ class to convert the project, or build prebuilt binaries once and share them.
Do I need Node.js?
Only for the stdio route. The native HTTP route runs entirely inside the editor.
Native HTTP or stdio: which should I use?
Native HTTP if your client supports Streamable HTTP with headers (Claude Code, Cursor, VS Code): no Node.js, and several clients can share one editor. stdio for Claude Desktop, or any client that only launches local commands; it also picks up the token by itself. Both expose exactly the same tool. Comparison: 🔌 Connecting Clients.
Do the plugin and the npm package versions match?
Yes. They are released together under the same number (for example 0.6.0-beta-b), and the stdio route expects matching versions. The plugin's version is in Edit › Plugins and in McpAutomationBridge.uplugin.
Can the assistant see the editor?
Yes. control_editor.screenshot returns the viewport image to the client, so a model that accepts images can look at its own work. manage_blueprint.preview_widget renders a UMG widget to an image the same way.
Can several clients, or several editors, be used at once?
Yes. One editor serves up to 16 native HTTP sessions, for example Claude Code and Cursor together. Two editors need two ports: set MCP_NATIVE_PORT per editor, or give each project its own Listen Ports on the stdio route (details).
Can it run Python?
Yes, through system_control.execute_python, with explicit consent per call. The Python Editor Script Plugin is enabled together with the bridge. Most tasks have a dedicated capability, which is safer and returns structured results, so the assistant should prefer those.
Where is the complete list of what it can do?
The generated Action Reference lists every capability with its effect, scope and consent. 🧰 Tools Reference gives the overview.
Can it wreck my project?
It can change and delete things; that is the point. The safeguards:
- deletes need a per-call consent grant;
-
control_editor.undosteps back through editor transactions; - partial failures are reported, not hidden.
💡 Use source control anyway, and commit before letting an assistant loose on a big task.
Does it send my project anywhere?
The plugin and the server talk only to each other and to your MCP client, over 127.0.0.1 unless you enable LAN access. Nothing sends telemetry. The exceptions are capabilities that exist to use an online service: Fab and MetaHuman calls go through the engine's own Epic services, and only when you run them.
Your AI client does send tool results (asset names, property values, screenshots) to its model provider, as it does with any tool.
I'm on 0.5.30. Should I move to 0.6?
0.6 is still a beta, but it is where development happens: new capabilities, fixes, and the single-tool gateway. ⬆️ Upgrading lists what changes.
How do I report a bug or ask for a feature?
Issues for bugs and feature requests, Discussions for questions and ideas. Security problems go through private vulnerability reporting. 🩺 Troubleshooting lists what to include in a report.