tc-imba/ue-official-mcp

ue-official-mcp-5.8.0 — pick the right UE 5.8.0 MCP tool

Operational routing for the official Unreal Engine 5.8.0 MCP server (the ModelContextProtocol plugin). This skill is the decision layer + an index into the generated tool catalog under references/. All schemas in this skill are captured from UE 5.8.0 — for other engine versions (including UE 5.8 patch releases) use the matching ue-official-mcp-<X.Y.Z> skill, since the MCP plugin is Experimental and surfaces drift between releases.

With AllToolsets enabled, the server exposes dozens of toolsets / hundreds of tools over HTTP (streamable, default http://127.0.0.1:8000/mcp). It requires the editor running and runs serially on the game thread — never issue overlapping calls. For the exact live counts and the snapshot date this reference was generated from, see references/index.md.

When to use

Before any official-MCP editor-automation call — to choose which toolset/tool fits the task, then read the matching references/*.md for tool names and (in the raw per-toolset file under references/toolsets/) the full input/output JSON schemas. Call directly from those schemas — no describe_toolset pre-flight needed.

How to call

Tool Search is on, so tools/list shows only 3 meta-tools. In a Claude Code session they surface as mcp__<server-key>__*, where <server-key> is the name given to the server in the MCP client's .mcp.json (Epic's setup example uses unreal-mcp → mcp__unreal-mcp__*). The three meta-tools:

1. list_toolsets — names + summaries of the enabled toolsets. Don't call for discovery: use the routing table below.
2. describe_toolset(toolset_name) — the tools (names + schemas) in one toolset. Only call as the recovery path when a call_tool returns a schema-shaped error (missing/unknown field, wrong type), to check whether the live schema has drifted from the file.
3. call_tool(name, arguments) — dispatch a tool; result returns on the same turn.

Flow: identify the domain below → open its references/*.md for the toolset + tool name → open the raw references/toolsets/<Name>.md for the full input schema → call_tool directly. If the call returns a schema-shaped error, re-describe_toolset to identify drift, then retry — and consider flagging the drift so a maintainer can re-sync.

Routing — task domain → reference

Task domain	Reference
Blueprint graphs & logic (incl. graph DSL)	references/blueprint.md
Materials, material instances, textures	references/material.md
Niagara (system/emitter/module/component)	references/niagara.md
Animation, Control Rig, Sequencer	references/animation-sequencer.md
Static/skeletal mesh, primitives, PCG	references/mesh-geometry.md
Scene/level, actors, assets, objects	references/scene-actor.md
GAS (cues, attribute sets, ASC inspect), GameplayTags	references/gas.md
UMG, StateTree, BehaviorTree, WorldConditions, Conversation	references/ui-state.md
Data/Curve/String tables, Data assets, DataRegistry	references/data-tables.md
Editor control (console/PIE/viewport), logs, automation tests, Slate inspect	references/editor-automation.md
GameFeatures, Plugins, Dataflow, Physics, SemanticSearch, AgentSkill	references/project-misc.md

Each domain file links into the raw per-toolset catalog (references/toolsets/), which carries every tool's full input/output JSON schema verbatim from describe_toolset.

Gotchas

• Experimental — APIs/data formats may change between engine versions. The reference is a snapshot from when sync last ran (see references/index.md for the date). Trust the schema optimistically; recover via describe_toolset only if a call_tool returns a schema-shaped error.
• Editor must be running; calls are serial — one at a time.
• AgentSkillToolset.CreateSkill/UpdateSkill require explicit user permission (the tools say so) — don't author AgentSkills unprompted.
• Regenerating this reference: run uv run ue-mcp-skills sync --engine <X.Y.Z> --launch (probes a live editor and rewrites everything under references/). Without --launch, point --endpoint at an already-running editor. The generated files carry a stamp header — do not hand-edit them; edit the domain grouping in scripts/toolset_map.yaml instead.