Ue MCP Blueprint

ue-mcp blueprint authoring

The blueprint tool covers reading, authoring, and compiling Blueprints. The default workflow is read → mutate → compile, never fire-and-forget.

Authoring a graph body: prefer the Epic DSL (#711)

When you are authoring the contents of a graph (event graph, function body, macro) - a set of nodes plus their wiring - do not default to node-by-node add_node/connect_pins. Reach for Epic's graph DSL first:

1. blueprint(action="epic_get_graph_dsl_docs") - read the S-expression grammar once.
2. blueprint(action="epic_write_graph_dsl", ...) - author + compile the whole graph body in one call.

The DSL authors and compiles an entire graph in a single round-trip, which is materially faster and more reliable than stitching individual K2Nodes together (one correct pass vs several failed iterations). Use it for graph bodies whenever it is available.

Availability / fallback. The epic_* actions come from Epic's ToolsetRegistry and require UE 5.8+ with the Epic toolset plugins enabled. Check with epic(action="status"). When they are unavailable (pre-5.8, or the registry is off), fall back to the native node path below.

Keep using the native actions for read/discovery, SCS components, CDO/class defaults, interfaces + event dispatchers, structured compile/validate, and anything with no Epic equivalent - the native path adds idempotency and rollback the raw tools lack. See the ue-mcp-epic-routing skill for the full epic-vs-native decision.

Discovery before authoring

For any existing Blueprint:

• blueprint(action="read", assetPath=...) - structure (parent, components, graphs)
• blueprint(action="read_graph_summary", assetPath=..., graphName=...) - lightweight node+edge summary (~10KB) - use this before read_graph (which can be 100KB+)
• blueprint(action="list_graphs", assetPath=...) - every graph in the BP including event graphs, functions, macros, interface impls
• blueprint(action="list_variables" | "list_functions" | "list_local_variables") - the member surface. list_functions covers the same graphs as list_graphs (own functions, parent overrides, interface impls, event graphs and their entry points, macros, collapsed subgraphs), each tagged with kind and source; pass includeInherited: true to also see overridable functions not implemented yet
• blueprint(action="get_execution_flow", assetPath=..., entryPoint=...) - trace exec pins from an entry point
• blueprint(action="get_dependencies", assetPath=..., reverse=false|true) - classes/functions/assets this BP uses, or callers if reverse: true

Mutation recipe

1. Create the skeleton - blueprint(action="create", assetPath, parentClass?).
2. Add variables + components - add_variable, add_component (pass parentComponent for SCS hierarchy).
3. Build graphs - add_node each K2Node, set_node_property for pin defaults, connect_pins to wire exec + data.
4. Compile - blueprint(action="compile", assetPath). Compilation errors come back in the result; fix them before proceeding.

Node wiring fundamentals

• add_node takes nodeClass as a K2Node class short name (K2Node_CallFunction, K2Node_VariableGet, K2Node_DynamicCast, K2Node_IfThenElse, etc.) plus nodeParams for class-specific fields (FunctionReference, VariableReference, TargetType).
• Pin defaults vs linked values: set_node_property writes a literal default onto a pin; connect_pins wires a pin to another node's output. A literal default is ignored once a pin is linked.
• read_node_property reads either a pin default or a reflected node UPROPERTY - use this to verify the pin was actually set before compiling.
• Graphs with duplicate names (rare but possible after rename) can be disambiguated by passing graphIndex alongside graphName.

Components (SCS)

• add_component creates a node in the Simple Construction Script. Pass parentComponent to put the new component under an existing parent - otherwise it becomes a top-level child of the scene root.
• set_component_property writes on the child BP's InheritableComponentHandler override template, not on the parent - the parent stays untouched. This matters when editing inherited components.
• read_component_properties dumps every UPROPERTY on the template, including array contents.
• reparent_component moves an SCS node to a new parent.

CDO (class defaults)

• set_class_default writes a UPROPERTY on the Blueprint CDO (the class default object). For actor tick settings specifically, use set_actor_tick_settings - it handles bCanEverTick, bStartWithTickEnabled, TickInterval in one call.

Interfaces + event dispatchers

• create_interface + add_interface - the implement-side.
• add_event_dispatcher - fires a multicast delegate from the BP.

Behavior Tree TASK Blueprints (#886)

A BTTask Blueprint is an ordinary Blueprint whose parent is UBTTask_BlueprintBase, so the whole graph is authored with override_function, add_node and connect_pins. There is no separate BT-task authoring action, and none is needed. The two entry points are override events, not functions, and the two finishers are ordinary BlueprintCallable calls.

Exact sequence for a move task:

blueprint(action="create", assetPath="/Game/AI/Tasks/BTT_MoveToTarget",
          parentClass="/Script/AIModule.BTTask_BlueprintBase")

# Entry point. ReceiveExecuteAI is a BlueprintImplementableEvent, so
# override_function places it as an override EVENT in EventGraph and returns
# {"kind": "event", "graphName": "EventGraph", "nodeId": "<guid>"}.
blueprint(action="override_function", assetPath=..., functionName="ReceiveExecuteAI")

# The AI MoveTo node. Its UFUNCTION (UAITask_MoveTo::AIMoveTo) is
# BlueprintInternalUseOnly, so it is NOT a CallFunction: place the K2 node
# class directly.
blueprint(action="add_node", assetPath=..., graphName="EventGraph",
          nodeClass="K2Node_AIMoveTo")

# Finisher. BlueprintCallable on the parent, so a plain CallFunction node.
blueprint(action="add_node", assetPath=..., graphName="EventGraph",
          nodeClass="CallFunction",
          nodeParams={"functionName": "FinishExecute",
                      "className": "/Script/AIModule.BTTask_BlueprintBase"})
blueprint(action="set_node_property", assetPath=..., graphName="EventGraph",
          nodeName="<FinishExecute nodeId>", propertyName="bSuccess", value="true")

blueprint(action="connect_pins", assetPath=..., graphName="EventGraph",
          sourceNode="<ReceiveExecuteAI nodeId>", sourcePin="then",
          targetNode="<AIMoveTo nodeId>", targetPin="execute")
# ... and on to FinishExecute.

# Abort path, same shape.
blueprint(action="override_function", assetPath=..., functionName="ReceiveAbortAI")
blueprint(action="add_node", ..., nodeClass="CallFunction",
          nodeParams={"functionName": "StopMovement",
                      "className": "/Script/Engine.Controller"})
blueprint(action="add_node", ..., nodeClass="CallFunction",
          nodeParams={"functionName": "FinishAbort",
                      "className": "/Script/AIModule.BTTask_BlueprintBase"})

blueprint(action="compile", assetPath=...)

Notes that save a round trip:

• ReceiveExecuteAI / ReceiveAbortAI / ReceiveTickAI take (AAIController* OwnerController, APawn* ControlledPawn). The non-AI variants (ReceiveExecute, ReceiveAbort, ReceiveTick) take (AActor* OwnerActor) instead. Both sets go through override_function.
• blueprint(action="list_overridable_functions") on the task Blueprint lists them with canBeEvent: true, which is the discovery path if you are unsure of a name.
• A task that overrides ReceiveExecuteAI stays active until FinishExecute is called, and one that overrides ReceiveAbortAI stays aborting until FinishAbort is called. Wire both terminators or the task never returns.
• The BehaviorTree ASSET (nodes, decorators, services, the tree structure) is a different surface. This section is only about the task Blueprint's graph.

Reading a level script Blueprint (#942)

read, list_graphs, read_graph, read_graph_summary, get_execution_flow and resolve_graph accept a World/umap path directly: pass /Game/Maps/SomeLevel and the call resolves to that map's level script Blueprint at PersistentLevel.<MapName>. The result carries blueprintPath and isLevelScript so you can see which object answered. A map whose Level Blueprint has never been opened has no level script object yet, and the error says exactly that rather than claiming the Blueprint is missing.

Auditing call sites across a project (#945)

blueprint(action="search_call_sites", functionNames=[...]) finds authored call nodes for named functions across a whole directory, including nested and collapsed graphs, in one call. Narrow it with className (which also lets the Asset Registry rule most packages out before anything is loaded) and directory. Bound the response with limit/offset, or pass dumpToFile to write the full result set to a JSON file the way read_graph does. Reach for it instead of enumerating every Blueprint and reading every graph.

Read stats before trusting an empty result: narrowedByRegistry says whether the dependency filter ran, blueprintsSkippedByRegistry says how much it removed unloaded, and narrowingSkippedReason says why it did not run. Pass narrowByRegistry: false to force a full scan when a result looks too small. Level scripts are not registry assets, so they are only searched when you pass includeLevelScripts, which loads a whole map per level.

Verify before compile

• blueprint(action="validate") runs the compiler diagnostics without saving. Cheaper round-trip when iterating.
• blueprint(action="read_graph_summary") after mutations confirms the graph shape.

Verify a default actually landed (#902 / #931)

blueprint(action="get_variable_default", assetPath, name) reads the RESOLVED value off the generated class CDO, which is the only value worth comparing after a write plus compile. list_variables on its own proves the variable exists, not what it holds.

It also answers whether the value is on DISK. persisted: false (with packageDirty: true) means the write is live in this editor session and has not reached the package, so it will be gone after a restart. Save the asset and read again before calling a write verified. A readback that only echoed the in-memory value would report success for a write that silently reverts.

Pass includeValues: true to list_variables for the same resolution across every variable at once, with one persistence verdict for the listing.