Claude Code skills for Unreal Engine

A Claude Code skill is a folder with a SKILL.md file: YAML frontmatter (a description is recommended, plus optional fields such as allowed-tools) followed by Markdown instructions. Claude Code loads a skill when your request matches its description, or when you type /skill-name.

• Project skills live in .claude/skills/<name>/SKILL.md. Commit them and your team gets them too.
• Personal skills live in ~/.claude/skills/<name>/SKILL.md.
• Plugin skills come with a plugin you install through /plugin, and you call them as /plugin-name:skill-name.

Anthropic's documentation suggests keeping a SKILL.md under 500 lines. A skill's description is cut at 1,536 characters in the skill list, and after the conversation is compacted each skill keeps only its first 5,000 tokens. Edits to a skill are picked up while Claude Code runs.

Epic publishes unreal-engine-skills-for-claude-code-plugin on GitHub under the MIT licence. Its README describes it as a way to control Unreal Editor from Claude Code over MCP, with 30+ toolsets behind it: actors, Blueprints, materials, Niagara, Sequencer and automation testing among them.

• The unreal-mcp skill: how to use Unreal MCP from Claude Code, with reference pages for setup and day-to-day operations.
• A SessionStart hook (unreal-context.sh) that tells Claude Code the repository is an Unreal project.
• Two companion skills: create-toolset, and unreal-skill for creating and reviewing skills.

The README gives these commands. Run them inside Claude Code:

/plugin install unreal-engine-skills-for-claude-code@claude-plugins-official

# if the marketplace is missing:
/plugin marketplace add anthropics/claude-plugins-official

To turn it on for a whole team, the README adds "unreal-engine-skills-for-claude-code@claude-plugins-official": true under enabledPlugins in the project's .claude/settings.json.

Before it does anything useful you need:

• Unreal Editor with the Unreal MCP (ModelContextProtocol) and All Toolsets plugins enabled.
• The editor running with the MCP server started (it auto-starts by default, or run ModelContextProtocol.StartServer).
• bash on your PATH for the SessionStart hook. On Windows, use Git Bash or WSL.

Connecting Claude Code to the editor is covered step by step in how to set up Unreal Engine 5.8 MCP with Claude Code or Cursor.

The unreal-mcp skill triggers when you ask for an action in a live project (spawn actors, edit Blueprints, compile with Live Coding) and stays out of the way for conceptual questions. With tool search on, Claude Code discovers tools with list_toolsets and describe_toolset, then runs them with call_tool.

• Save before and after bulk changes: MCP edits are not always undoable.
• Wait for compiles to finish, and compile C++ through Live Coding.
• Run dependent calls one after another, not in parallel.
• Stop on any result that is not an explicit success.
• Check whether Play in Editor is running when results look odd.

Engine-side skills are a different thing

Unreal has its own idea of a skill. In the 5.8.0 source release, a skill is an engine object with a description and instructions, served through two Unreal MCP tools, ListSkills and GetSkills. Epic's unreal-mcp skill tells Claude Code to call ListSkills at the start of unfamiliar work and to follow a project's skills over generic defaults.

The source release ships engine skills for Blueprint basics, material basics, default outdoor lighting, Niagara (fundamentals, renderers, optimization and more) and PCG graph generation. Unreal MCP also has an Instructions setting (Editor Preferences > Agentic Workflows > Model Context Protocol; the project version is shared through DefaultGame.ini) that sends a hint to every client.

• The README warns: "Localhost is not a trust boundary." Unreal MCP has no authentication of its own.
• execute_tool_script runs arbitrary Python with full editor access. The README advises against --dangerously-skip-permissions.
• Commit or shelve your work before a long session; some edits cannot be undone in the editor.
• Unreal MCP is Experimental, and the README does not name a supported engine version.
• The skill tells Claude Code how to act carefully, but checking that the result works in play is still on you.

Claude Code without the plugin

Without Unreal MCP, Claude Code still handles C++, config files and builds from the terminal, but it cannot see inside Blueprint assets or the running editor. More on that split: Claude Code for Unreal Engine vs UE-native AI.

Dola is not part of this stack. It is a desktop app plus its own Unreal Editor plugin, built for Unreal Engine 5.7 on Windows and Linux. Its agents run Claude through Dola's backend, so you do not install Claude Code, a plugin marketplace entry or an MCP config.

• Review before you keep: each edit waits in a Changes panel. Open the diff in Unreal, then accept it, or reject it to restore the old asset.
• QA by playing: a QA agent reads the graph and plays the change in Play in Editor before the run is called done.
• Pick by setup: on a 5.8 project where you already use Claude Code, Epic's plugin is the direct route. On 5.7, Dola is packaged and checks its own work.

A real Dola run on 9 November 2026: one prompt, "make BP_Lamp blink", checked in PIE, then held for review.

The whole run, step by step: see Dola work.