Index

Agent MCP documentation

한국어 · README

Agent MCP runs an MCP server inside the Unreal Editor so agents can inspect and edit a project, run it, and review the outcome. Start here based on what you want to do.

Getting started

Goal	Read
Install the plugin and connect for the first time	README installation and connection
Configure Claude Code, Codex, authentication, or multiple editors	Connecting clients
Follow task-oriented tool call sequences	Workflows
Look up tool names and access levels	Tool reference
Set ports, tool access, and other project settings	Settings
Understand editing restrictions and approvals	Safety

By feature

Area	Read	Covers
Widget Blueprints and UI	Dungeon UI example	BindWidget, reusable components, data-driven lists, authoring and captures
Visual consistency	UI style system	Theme Data Assets, styled widgets, review
Blueprints, levels, and assets	Tools	Graph inspection/editing, usage search, data and actors
Task guidance	Skills	Plugin and project SKILL.md files, overrides
In-editor AI chat	AI chat	Providers, sign-in, permission modes, approvals
MCP internals	MCP architecture	Transport, Reflection discovery, invocation, approvals and chat integration
Extending the toolset	Writing tools	Adding Reflection-based MCP tools with UFUNCTION
Performance investigation	Investigation tooling and example	Standalone repeated runs, CSV/Insights/GPU diagnosis, variation and Sonheim measurements
Testing	Testbed and smoke tests	Running the included test project

How the pieces fit

AgentMcp connects external Claude Code/Codex clients and the in-editor AI chat to the same Reflection-based tool layer through a local MCP server. Tool execution, access levels, approval gates and Agent Skill loading are separate responsibilities.

See MCP architecture for the request-to-editor call path, or Writing tools to extend the toolset with a UFUNCTION.

Before making changes

• Tool access levels distinguish reads, edits, destructive operations, and editor controls. See Tools and Safety.
• Approval cards in the in-editor AI chat are not the same approval flow as a direct Claude Code or Codex connection. The approval gate depends on the client tag.
• Asset transactions and source-file edits have different recovery paths. source_replace and source_write cannot be reverted with editor_undo; use source control.
• Confirm the target project via editor_get_state before changing it.

Documentation lives in the repository's Docs/ folder. Links here point to those files directly; they do not imply a separate GitHub Wiki.