Testing

Testbed and smoke test

한국어

The repository root is a small Unreal Engine 5.5 project that builds and tests the plugin.

Path	Contents
Plugins/AgentMcp	The plugin, with its skills in Plugins/AgentMcp/Skills
Source/AgentMcpTestbed	Row structs, data assets, widget bases, game mode and the C++ classes of the UI sample
Source/AgentMcpTestbedEditor	Test-only tools and sample_show_widget
Content/Samples/DungeonUi	Widget Blueprints, theme data asset and item table built through the tools
Content/Samples/Pickup	A spinning, bobbing pickup Blueprint (variables, a pure function, construction script and event graph) and its demo level, built through the blueprint tools
Art/Requests	Art requests for the sample
Docs	Case studies, the UI style system of the sample, and experiments
.mcp.json, .codex/config.toml	Claude Code and Codex connection settings for testbed port 18766
.claude/skills, .agents/skills	Bridges that let clients begin the served plugin skills
Config	Testbed port 18766, the smoke test's skill folder in SkillDirectories, and the UI sample's theme in DefaultGame.ini
Tools/mcp_smoke.py	Smoke test using Python 3 standard library only
Tools/mcp_call.py	Calls one tool from the command line

1. Build the AgentMcpTestbedEditor target.
   <UE>\Engine\Build\BatchFiles\Build.bat AgentMcpTestbedEditor Win64 Development -Project=<path>\AgentMcpTestbed.uproject -WaitMutex
2. Open AgentMcpTestbed.uproject and wait for Agent MCP server listening on http://127.0.0.1:18766/mcp.
3. Run python Tools/mcp_smoke.py --out Saved/MCP/smoke.json.

The smoke test covers MCP transport and errors, every tool, undo and rollback, request cancellation, Play In Editor, viewport capture with game UI, Live Coding, nested Widget Blueprint editing, Blueprint graph editing, skills, data-asset/DataTable creation, texture and mesh import, level authoring and animation assets.

P9 checks the source tools (ranges, regular expressions, CP949 and CRLF, refusals), the navigation tools and the approval gate for tagged clients. It needs no testbed content, so it also runs against another project; skip the sections that need the testbed:

python Tools/mcp_smoke.py --url http://127.0.0.1:<port>/mcp --expect-project <Project> --skip-slice1 --skip-p1 --skip-p2 --skip-p3 --skip-p4 --skip-p5 --skip-p6 --skip-p7 --skip-p8

P9 writes two text files at the top of the first readable folder (Config by default) and removes them again.

P11 checks the perf tools against CSV captures and a log it writes under Saved/MCP and removes again.

To call a single tool:

python Tools/mcp_call.py editor_get_state --url http://127.0.0.1:18766/mcp --expect-project AgentMcpTestbed