Unreal now exposes an MCP server: a bridge that lets an assistant such as Claude Code genuinely drive the editor — create functions, place nodes, wire events. Here is how to set it up, and what it gives on a concrete mechanic.
On the Unreal side
- 1In Edit ▸ Plugins, search for MCP and enable Unreal MCP — that is what handles the connection.
- 2Enable All Tool Set as well: without it, the assistant connects but has no tools at its disposal.
- 3Restart the editor.
- 4Start the server with the ModelContextProtocol.StartServer command.
- 5Generate the client configuration with ModelContextProtocol.GenerateClientConfig ClaudeCode.
The second command writes a configuration file into the project telling Claude Code which port to connect to.
On the code editor side
- 1Install VS Code, then the Claude Code for VS Code extension from the Extensions tab.
- 2In the Chat section, open the Claude tab and connect your account.
- 3Drag the Unreal project folder into VS Code.
- 4Type /mcp in the prompt window to check the connection.
If /mcp reports nothing connected, it is almost always because the configuration file was generated after the window started. Ctrl + Shift + P, then Reload Window, and check again.
Skills: fewer tokens burnt
Epic publishes a repository of skills that teach the assistant how to use the MCP properly. Optional, but recommended: without them, a good part of the session goes into exploring the available tools.
Installation fits in one sentence: paste the GitHub repository link to Claude Code and ask it to install the skills — saying whether you want them for the project only or for the whole machine.
The /context command shows current consumption. Worth watching on long sessions: it is the best indicator of when to start a fresh conversation.
The test: the triple jump
The mechanic is familiar: the second jump goes higher than the first, the third higher than the second — provided you jump again within a short window after landing. Simple, but it needs state and two distinct events.
A one-line prompt is enough, naming the target Blueprint — here BP_ThirdPersonCharacter. The first request is a little slow: the assistant discovers the tools while it codes. It then remembers them for later features.
The result is sound: a dedicated function for the triple jump, wired to the On Jump event, and On Landed used to measure the delay between landing and the next jump. In-game behaviour is what you would expect, with a reset after the third jump.
The real problem: readability
AI does not see Blueprints in space. It places nodes correctly from a logical standpoint, but stacks them visually — and reading its work back means reformatting by hand every time.
Blueprint Arrange, available on FAB for a few euros, fixes that: Ctrl + R rearranges every node, and wraps the block in an empty comment showing where the code starts and ends.
The real gain comes when you ask the assistant to use it itself. Add to your prompt that it should format its code with Blueprint Arrange: it then produces a tidy, commented graph, readable with no touch-up.