Unreal_mcp

Connect Claude, Cursor, VS Code or any other MCP client to a running Unreal Editor,
and let it build levels, Blueprints, UI, materials, effects and more, through a native C++ editor plugin.

_{🧭 One tool, nearly 400 capabilities  ·  ⚙️ Native C++ editor plugin  ·  🌐 HTTP or stdio  ·  🔐 Local and token-protected by default  ·  🎮 Unreal Engine 5.0 – 5.8}

🚀 Quick Start  ·  📖 Wiki  ·  📦 Releases  ·  💬 Discussions

📌 Which version is this? This README describes the 0.6 line: the dev branch and npm unreal-engine-mcp-server@beta. The previous stable release, 0.5.30 (npm latest), exposes 23 separate tools instead of one; see Upgrading from 0.5.x.

Contents · What it does · How it works · Quick start · The unreal tool · Configuration · Security · Engine plugins · Docker · Documentation · Development · Community

---

What it does

🏗️ Levels and actors Spawn one actor or hundreds in a single call, place and attach them, find the ones sunk into the floor, and load, stream and save levels.	🧩 Blueprints and UI Create Blueprints, variables and components, build whole event graphs in one batch, and lay out UMG widgets, with a preview image to check them.	🎨 Materials and worlds Material graphs and instances, procedural textures, lighting, landscapes, foliage, Niagara effects and PCG graphs.
🕹️ Gameplay Characters and animation, Gameplay Ability System, AI (Behavior Trees, State Trees, EQS), inventory, networking and Enhanced Input.	🎬 Cinematics and audio Level Sequences, cameras, Movie Render Queue and Take Recorder; Sound Cues and MetaSounds.	🧪 Play and verify Run Play-In-Editor with synthetic input, take screenshots the model can see, read logs, profile, run Python, and package builds.

Nearly 400 capabilities in all, behind a single MCP tool. The assistant finds them by searching in plain words, so nobody has to learn their names. The full list is the generated Action Reference.

_{A platformer level built with the MCP, in Unreal Editor 5.8. Bottom right: the plugin's status, MCP :3000 (1), meaning the native server is running on port 3000 with one client connected.}

How it works

The MCP Automation Bridge plugin runs inside the editor and does all the work, on the editor's game thread. Clients reach it in one of two ways, and both expose the same single tool, unreal:

	🌐 Route A · Native HTTP	🧩 Route B · stdio
Path	Client → the plugin's Streamable HTTP server at http://127.0.0.1:3000/mcp	Client → unreal-engine-mcp-server (Node.js, stdio) → the plugin's WebSocket on 127.0.0.1:8090
Node.js	Not needed	20.19 or later
Capability token	The client sends it in the X-MCP-Capability-Token header	Read from the project, given UE_PROJECT_PATH
Best for	Claude Code, Cursor, VS Code, and several clients sharing one editor	Claude Desktop, and clients that only launch local commands

Everything listens on 127.0.0.1 and requires the project's capability token unless you change it.

Quick start

The Quick Start page walks through this with screenshots. You need Unreal Engine 5.0 to 5.8 and a project with C++ code; a Blueprint-only project can use prebuilt binaries.

1. Add the plugin

Download McpAutomationBridge-plugin-<version>.zip from the newest v0.6 pre-release on the Releases page, and copy the McpAutomationBridge folder it contains into your project:

MyGame/Plugins/McpAutomationBridge/

Or use a clone of this repository: copy plugins/McpAutomationBridge/, or reference the folder from your .uproject with "AdditionalPluginDirectories": ["C:/Path/To/Unreal_mcp/plugins"].

Open the project and let Unreal rebuild the plugin. When it's loaded, the status bar shows MCP off. If you see "Engine modules cannot be compiled at runtime", build the project once in Visual Studio, Rider or Xcode. More in Installation.

https://github.com/user-attachments/assets/d8b86ebc-4364-48c9-9781-de854bf3ef7d

Prebuilt binaries (Blueprint-only projects, teams)

Build the plugin once on a machine with the engine and a compiler, then hand out the zip. No compiler is needed on the target machine:

node scripts/package-plugin.mjs "C:/Program Files/Epic Games/UE_5.7"

This writes build/McpAutomationBridge-v<version>-UE5.7-<Platform>.zip, where <version> is the package.json version (currently 0.6.0-beta-b). Unzip it into YourProject/Plugins/. Binaries only work with the engine minor and platform they were built for: a 5.6 build won't load in 5.5, 5.7 or 5.8.

2. Connect your client

Route A · Native HTTP (no Node.js)

1. In Edit › Project Settings › Plugins › MCP Automation Bridge, tick Enable Native MCP Server (port 3000 by default), then restart the editor. The status bar now reads MCP :3000 (0).

   

2. Read the capability token the plugin generated: <YourProject>/Saved/MCP/capability-token. Treat it like a password.

3. Add the server to your client and send the token in the X-MCP-Capability-Token header. Claude Code:

   claude mcp add --transport http unreal-engine http://127.0.0.1:3000/mcp --header "X-MCP-Capability-Token: <token>"
   

   Or in a project .mcp.json (Claude Code), reading the token from an environment variable:

   {
     "mcpServers": {
       "unreal-engine": {
         "type": "http",
         "url": "http://127.0.0.1:3000/mcp",
         "headers": { "X-MCP-Capability-Token": "${UNREAL_MCP_TOKEN}" }
       }
     }
   }
   

   Cursor (.cursor/mcp.json) takes the same url and headers without type. VS Code, Windsurf and others: Connecting Clients.

4. Check it: when the client connects, the count in the status bar goes up.

   

Route B · stdio (Node.js 20.19+)

Add this to your client's MCP configuration, for example Claude Desktop's claude_desktop_config.json:

{
  "mcpServers": {
    "unreal-engine": {
      "command": "npx",
      "args": ["-y", "unreal-engine-mcp-server@beta"],
      "env": {
        "UE_PROJECT_PATH": "C:/Path/To/YourProject"
      }
    }
  }
}

UE_PROJECT_PATH (the project folder or its .uproject) is how the server finds the capability token and the plugin's port, so nothing else needs setting. Keep the @beta tag: without it, npm installs 0.5.30, which doesn't match a 0.6 plugin.

3. Try it

With the editor open, ask your assistant to "list the actors in the current level", "spawn a point light 300 units above the origin", or "take a screenshot of the viewport". If it doesn't connect, see Troubleshooting.

The unreal tool

Both routes expose exactly one MCP tool, unreal, with four operations. Only the contract the model is about to use gets loaded, instead of hundreds of tool schemas:

Operation	What it does
search	Finds capabilities from 2-4 plain words, such as spawn actor or save level. Every row carries a ready-to-send nextCall.
describe	Returns one capability's exact contract: parameters, schemas, an example, and the consent grant when one is needed
execute	Runs one capability with validated parameters and returns the data plus a receipt of what changed
configure	Enables or disables groups of internal tools; never touches the editor

A typical exchange:

{ "operation": "search", "query": "spawn actor" }

{ "operation": "describe", "tool": "control_actor", "action": "spawn" }

{
  "operation": "execute",
  "tool": "control_actor",
  "action": "spawn",
  "params": { "classPath": "/Script/Engine.PointLight", "actorName": "KeyLight", "location": [0, 0, 300] }
}

• Parameters are strict. An undeclared name is refused with UNDECLARED_PARAMETER and the list of allowed names; every error carries an executable nextCall.
• Deletes need consent. 62 capabilities (all destructive ones, plus some writes) only run with the consentGrant from describe, passed back as a top-level consent field.
• Names. Both routes accept the tool + action pair; the stdio route also accepts a capability id, such as "capability": "control_actor.spawn".

Full reference with real replies: Using the Gateway.

Migrating from direct tool calls

The single unreal tool is permanent on both routes; there is no opt-out and no 23-tool listing to restore. A client that still calls a canonical tool name directly (tools/call with name: "manage_asset", name: "control_actor", …) receives a bounded, copy-paste-executable DIRECT_TOOL_CALL_REMOVED receipt instead of a routed call. Its nextCall drills exactly one level: { "operation": "search" } for an unknown name, { "operation": "describe", "tool": "<tool>" } when no action was given, or { "operation": "execute", "tool": "<tool>", "action": "<action>", "params": { ... } } when the call already named an action. Run that nextCall through unreal to finish the migration. See Upgrading.

Protocol versions

Both routes negotiate the MCP protocol version at initialize. The native /mcp transport supports 2025-11-25 (latest), 2025-06-18 and 2025-03-26; the stdio server also accepts the legacy 2024-11-05 and 2024-10-07. An unsupported MCP-Protocol-Version header on the native route gets HTTP 400. Details: docs/protocol.md.

The 23 internal tools behind unreal

These route requests inside the gateway; clients never list them. More in the Tools Reference.

Category	Tool	Covers
Core	manage_asset	Assets and folders, materials and material graphs, textures and render targets, data tables, structs, enums, source control
Core	manage_blueprint	Blueprints, components, variables, event graphs, UMG widgets, layout, bindings, widget animations
Core	control_actor	Spawning, transforms, attachment, components, materials, tags, placement audits
Core	control_editor	Play-In-Editor, synthetic input, screenshots, viewport camera, undo, editor preferences
Core	manage_level	Create, load, save, stream, import and export levels; world settings; lighting builds
Core	system_control	Console commands, logs, project settings, profiling, builds and packaging, tests, Python
Core	inspect	Read and write any UObject's properties, components and class info
Core	manage_tools	Which internal tools are enabled (through configure)
World	build_environment	Landscapes, foliage, lights and sky, water, weather, splines, procedural terrain
World	manage_level_structure	Sublevels, World Partition, streaming, data layers, HLOD, volumes
World	manage_geometry	Geometry Script meshes: booleans, deformers, UVs, collision, LODs
World	manage_pcg	PCG graphs: create, add and connect nodes, execute
Gameplay	animation_physics	Animation Blueprints, blend spaces, montages, skeletons, Control Rig and IK, ragdolls, cloth, vehicles
Gameplay	manage_character	Character Blueprints, movement, MetaHuman
Gameplay	manage_combat	Weapons, projectiles, hit detection
Gameplay	manage_effect	Niagara systems, emitters and modules, debug shapes
Gameplay	manage_gas	Gameplay Ability System: abilities, attributes, effects
Gameplay	manage_ai	AI controllers, Behavior Trees, EQS, State Trees, Smart Objects, perception, navigation
Gameplay	manage_inventory	Items, loot tables, crafting recipes
Gameplay	manage_interaction	Doors and other interactables
Utility	manage_audio	Sounds, audio components, Sound Cues, MetaSounds, attenuation, mixes
Utility	manage_sequence	Level Sequences, Movie Render Queue, media, Take Recorder, replays
Utility	manage_networking	Replication, RPCs, sessions, game framework classes, Enhanced Input

Configuration

Most setups touch one setting: Enable Native MCP Server for Route A, or UE_PROJECT_PATH for Route B. Everything is listed on the Configuration page.

Plugin settings (Project Settings › Plugins › MCP Automation Bridge, saved in Config/DefaultGame.ini):

Setting	Default
Enable Native MCP Server	off	Serve MCP over HTTP at /mcp
Native MCP Port	3000	The MCP_NATIVE_PORT environment variable of the editor process overrides it
Listen Ports	8090,8091	WebSocket ports for Route B
Require Capability Token	on	Both routes refuse clients without the token
Allow Non Loopback	off	LAN access for both listeners. See Security.

Environment variables (Route B only, in the client's env block):

Variable	Default
UE_PROJECT_PATH	unset	Project folder or .uproject; used to find the token and the port
MCP_AUTOMATION_PORT	the project's first Listen Ports entry, else 8090	Editor WebSocket port
MCP_AUTOMATION_HOST	127.0.0.1	A LAN address also needs MCP_AUTOMATION_ALLOW_NON_LOOPBACK=true
MCP_AUTOMATION_CAPABILITY_TOKEN	read from the token file	Token to present, when the server can't read the project folder
MCP_ADDITIONAL_PATH_PREFIXES	empty	Extra content roots such as /MyPluginContent/, comma-separated; most content-path arguments also accept the mounts the connected editor reports, so this is needed only with no editor connected, for a mount the editor does not report or the server ignores, and for arguments the server treats as files (filePath, outputPath and a few others), even where outputPath names an asset
LOG_LEVEL	info	debug, info, warn or error; logs go to stderr

Security

• Local by default. Both routes listen on 127.0.0.1 only. LAN access needs Allow Non Loopback, and the native server refuses to bind off-loopback unless Require Capability Token is on.
• Capability token. Generated per project at <Project>/Saved/MCP/capability-token (a value typed into Capability Token overrides it) and compared in constant time. Delete the file and restart the editor to rotate it.
• Consent for destructive work. Deletes and some other writes need a per-call consent grant, which the plugin checks itself.
• Guard rails. Asset paths are limited to /Game, /Engine, /Script, /Temp, /Niagara plus configured prefixes, and a content path may also sit under a mount the connected editor reports; arguments the server treats as files (such as filePath, and outputPath even where it names an asset) never use the reported mounts; console commands that chain or quit the editor are blocked; the plugin's own settings are out of reach of automation.

Details: Security. Please report vulnerabilities privately through GitHub security advisories.

Engine plugins

The bridge declares its engine-plugin dependencies, so Unreal enables them together with it.

Required (always enabled with the bridge)

Plugin	Used for
Python Editor Script Plugin	Python-backed editor automation, system_control Python execution
Editor Scripting Utilities	Asset and actor subsystem operations
Niagara	Visual effects
Gameplay Abilities	manage_gas
Smart Objects	AI smart objects

Optional (used when present, skipped when not)

Plugin	Used for
Level Sequence Editor, Takes, Movie Render Pipeline, Movie Pipeline Mask Render Pass, Electra Player	manage_sequence: Sequencer, Take Recorder, Movie Render Queue, media playback
Control Rig, RigVM, IK Rig, Animation Data	animation_physics: Control Rig and IK
Chaos Vehicles, Chaos Cloth	animation_physics: vehicles and cloth
Niagara Editor	manage_effect: Niagara authoring
Behavior Tree Editor, Environment Query Editor, StateTree, Mass Gameplay	manage_ai
Geometry Scripting, Geometry Processing, Procedural Mesh Component	manage_geometry
PCG	manage_pcg, compiled in only when the project itself enables the PCG plugin
MetaSound, Synthesis	manage_audio: MetaSound authoring
Enhanced Input, Online Subsystem, Online Subsystem Utils	manage_networking: input mappings, sessions
Interchange, Interchange OpenUSD, Data Validation, StructUtils	Import/export, validation, struct helpers
Fab, Bridge	Fab asset library access (optional McpAutomationBridgeFab module)

The plugin targets every Unreal Engine minor from 5.0 to 5.8. If it fails to build on yours, please open an issue with the build log.

Docker

The image runs the stdio server. It has to reach the editor's WebSocket on 127.0.0.1, so use host networking (Linux), and pass the token because the container can't read your project folder:

docker build -t unreal-mcp .
docker run -i --rm --network host -e MCP_AUTOMATION_CAPABILITY_TOKEN=<token> unreal-mcp

Use -i without -t: a TTY corrupts the MCP stream on stdout.

Documentation

📖 Wiki	Quick Start, Installation, Connecting Clients, Using the Gateway, Configuration, Security, Troubleshooting, FAQ, Upgrading
📋 Action Reference	Every capability with its effect, scope and consent, generated from the capability records
🔌 Gateway client guide	The gateway contract, with the source file behind each claim
📡 Protocol	Transports, version negotiation, cancellation
🔐 Security and receipts	Scopes, consent, path gating, idempotency, refusal codes
🧪 Testing guide	Test suites and how to add cases
🧩 Extending the plugin	Adding an editor action: record, handler, route and tests, plus the rules for plugin code
🗺️ Roadmap	Development roadmap
📝 Changelog	What changed in each release

Development

npm install
npm run build              # clean + compile TypeScript to dist/
npm run dev                # run from source with ts-node
npm run lint               # ESLint (CI fails on any warning)
npm run type-check         # tsc --noEmit, sources and tests
npm run test:unit          # Vitest unit tests (no Unreal required)
npm run test:smoke         # offline mock-mode MCP check (needs built dist/)
npm run test:params        # strict parameter audit
npm run registry:generate  # capability records -> generated contracts, native shards, action reference
npm run registry:check     # fail if generated artifacts drift
npm run manifest:check     # fail if the gateway manifest drifts
npm run eval:check         # search-ranking corpus
npm run version:check      # all version sources agree
npm test                   # integration suite (needs a live Unreal Editor + bridge)

The capability records in src/tools/catalog/capabilities/records/ are the single source of truth; every *.generated.* file, the plugin's MCP/Generated/ shards and the action reference are generated from them. Never hand-edit generated files. How to add a capability: Development.

CI runs, in order: ESLint (npx eslint . --max-warnings=0), type-check, unit tests, registry:check, manifest:check, headers:check, the strict parameter audit (test:params) and eval:check, then a blocking runtime dependency audit (npm audit --omit=dev --audit-level=moderate) and an informational full-tree audit. A Node 20.19 / 26 matrix adds build and test:smoke. Plugin packaging runs only when an Unreal Engine root is configured, because CI runners don't ship an engine. Release archives exclude Binaries/, Intermediate/ and Saved/.

Community

💬 Discussions	Questions, ideas, show and tell
🐞 Issues	Bug reports and feature requests
🗺️ Project board	Roadmap progress and priorities

Contributing: keep pull requests small and focused, with a Conventional Commits title; include reproduction steps for bugs; follow the existing code style. See CONTRIBUTING.md.

License

MIT. See LICENSE.