Unreal Engine MCP

Control Unreal Engine with AI. An MCP server that gives AI assistants (like Claude and Cursor) direct, programmatic access to the Unreal Engine environment to manipulate scenes, spawn objects, and execute commands.

GitHub
23

Documentation

unreal-mcp

Unreal MCP Hero Demo

An MCP server that allows AI agents (Claude, Cursor) to directly control and manipulate Unreal Engine.

This server lets Claude (or any MCP client) read and edit Unreal Engine 5.6/5.8 Blueprints directly, without burning your context window on raw engine JSON, and without needing anything beyond a stock Epic Games Launcher install.

Architecture

The problem this solves

If you've tried pointing an AI assistant at a real Unreal project, you've hit this: Blueprints don't fit in a context window. A single graph dumped as raw engine data is enormous, so either the model never sees enough of the project to have real context, or you spend most of your budget re-explaining what already exists every time you open a new conversation.

This project is built around one idea: the model should never receive a raw engine dump. Every hop between the Unreal Editor and Claude compacts the data: tiered reads, diff-based edits, and a persistent index that's built once and updated incrementally instead of re-scanned on every question.

How it works

Two pieces:

  • UnrealMCPBridge is a C++ editor plugin that runs inside UnrealEditor.exe and exposes a local TCP interface over the engine's own Kismet2/EdGraph/AssetRegistry APIs. Built against a stock launcher install: no engine source required to build or run it.
  • mcp-server is a Node/TypeScript MCP server that translates MCP tool calls into bridge requests, and is responsible for keeping every response cheap: compact field names, capped result sizes, and no re-serializing verbose engine data verbatim.

See ARCHITECTURE.md for the full design.

What's different about this one

There are several Unreal MCP projects on GitHub already, and as of UE 5.8 Epic ships its own experimental first-party MCP plugin (5.8 only, opt-in, requires manually enabling an "Editor Toolset"). Worth being direct about where this project actually differs, rather than just claiming "better":

  • Built around reading, not just writing. Most existing projects are strong at creating and manipulating Blueprints from a prompt, but don't address what happens when the model needs to understand a large, already-built project first. Reading is the first-class citizen here: tiered summaries before full detail, node IDs you can reference without re-fetching.
  • A persistent, incrementally-updated project index. The bridge indexes Blueprints, functions, variables, and cross-asset references once, caches it to disk, and updates it from AssetRegistry delegates as you edit, instead of rescanning the project on every query. find_references answers "what actually uses this Blueprint" without the model having to enumerate the project itself.
  • An optional local-model hook for indexing. If you point UNREAL_MCP_LOCAL_LLM_URL at a local model (Ollama or anything OpenAI-compatible), indexing summaries are generated there instead of spending Claude's tokens on mechanical scanning work. Fully optional. The index works without it.
  • Targets both 5.6 and 5.8 from one codebase, where several existing projects are pinned to a single engine version.

Full survey of the existing ecosystem (licenses, architectures, what each one does well) is in docs/COMPETITIVE_LANDSCAPE.md.

Status

This is being built and verified in public, milestone by milestone. Each milestone's status doc is written honestly, including what's compiled/tested versus what's still unverified:

All four milestones are build-verified, protocol-verified, and live-verified on both engine versions. See docs/LIVE_VERIFICATION.md for the 5.8 session against a real ~20-Blueprint project and docs/UE56_STATUS.md for the 5.6 one: reads returning correct real data, a full create/wire/compile/save write round-trip, and confirmation that the incremental project index actually stays fresh without restarting the editor (M3's core claim).

Both live sessions earned their keep by catching a real bug that no amount of compiling or protocol testing would have surfaced. On 5.8 it was add_node duplicating an already-present override-event node. On 5.6 it was the .uplugin hard-pinning EngineVersion to 5.8.0: every build check passed because UnrealBuildTool ignores that field, but the runtime plugin loader honors it, so the editor stopped on a modal incompatibility dialog and the bridge never started.

Since then, all of that has been exercised live too: remove_node and VariableGet are covered by the node-id and control-flow suites, and add_node now places Branch, Sequence, Cast, and standard-library macros (ForEachLoop, WhileLoop, ...) directly, verified by building and compiling a real conditional graph through the bridge alone. Node ids are persistent GUIDs, and every write is undoable with Ctrl+Z under a named "MCP:" transaction. Still outstanding: CustomEvent/VariableSet node types have not had a dedicated live check, and the M5 catalog covers UFunction-backed nodes; native UK2Node types are placed via the dedicated nodeType values rather than discovered through unreal_find_node.

Quickstart (3-Step Installation)

Ensure you have Node.js 18+ and a UE 5.6 / 5.8 project.

1. Install the Unreal Plugin

Easiest path: download the prebuilt plugin for your engine version and unzip it into your project's Plugins/UnrealMCPBridge/ folder, so you never compile it yourself:

Or build it yourself by copying the UnrealMCPBridge plugin folder to your Unreal project's Plugins/ directory:

# macOS / Linux
mkdir -p "/path/to/YourProject/Plugins" && cp -r UnrealMCPBridge "/path/to/YourProject/Plugins/"

# Windows (PowerShell)
New-Item -ItemType Directory -Force -Path "C:\path\to\YourProject\Plugins"; Copy-Item -Recurse UnrealMCPBridge "C:\path\to\YourProject\Plugins\"

Note: Rebuild/open your Unreal project to compile the plugin, and ensure it is enabled in the editor.

2. Build the MCP Server

Install the node dependencies and compile the typescript codebase:

cd mcp-server && npm install && npm run build

3. Register the Server

Connect the server to your MCP client using the absolute path to mcp-server/dist/index.js:

Claude Code:

claude mcp add unreal -- node "/path/to/unreal-mcp/mcp-server/dist/index.js"

Claude Desktop (claude_desktop_config.json):

{
  "mcpServers": {
    "unreal": {
      "command": "node",
      "args": ["/path/to/unreal-mcp/mcp-server/dist/index.js"]
    }
  }
}

Once registered, open your project in the Unreal Editor and verify the connection via unreal_ping.

For more configuration options and details, see mcp-server/README.md.

Contributing

Issues and PRs welcome. This project is young and moving fast, so check the status docs above before assuming something works end-to-end.

License

MIT. See LICENSE.