Zum Inhalt springen
Medium 1 für Eintrag Asset Context Exporter For AI

Beschreibung

Asset Context Exporter For AI -

Turn your whole Unreal project into text an LLM can read.

If you got tired of pasting Blueprint screenshots into ChatGPT and getting back code that used variable names you never wrote or having to pay suscriptions and burn your tokens only for your AI to just take a look at your project. This plugin is for you.

Claude, ChatGPT and Gemini know Unreal Engine generically; documentation, forums, tutorials. They don't know YOUR project. Ask them about your Blueprint and they guess your variable names, ignore your interfaces, and hand you code that doesn't fit your architecture. Technically valid, contextually useless.

This plugin exports most of your your project, not just blueprints, or one subsystem, into structured JSONL files that any LLM can read directly. Real variable names, real execution flows, real component hierarchies, real skeletons and state machines. Source code and config are exported as Markdown instead, because that is what models read best.

NOTHING LEAVES YOUR MACHINE

No API keys. No account. No subscription. No network calls of any kind. It's a headless commandlet that writes files to a folder you pick. What you do with them afterwards is your call, including feeding them to a fully local LLM. If your studio isn't allowed to send source code to a cloud service, this still works.

UNLIMITED USES
No suscriptions and no extra payments-Buy it once, own it forever.
Export any compatible project as many times as you want.

  • WHAT ACTUALLY GETS EXPORTED

Most exporters dump variable names and call it a day. This one goes further:

▸ BLUEPRINTS

Full execution flow AST. Not just node names, but what calls what, in what order, with what inputs. Pure math nodes (Add, Multiply, Compare) expanded inline. Hard cast count exported as a coupling signal. Component topology, socket data, ChildActor template overrides.
For Widget Blueprints: full widget tree, all properties, UMG animation tracks, and dynamic property bindings.

▸ ANIMATION SYSTEM (rebuilt in 1.1)

Skeletons: complete bone tree with parents and local positions, sockets, virtual bones, slot groups, curves, per-bone retargeting overrides.
Skeletal Meshes: bone hierarchy, skeleton/physics bindings, LODs, morph targets, material slots. Anim Blueprints: AnimGraph node Details plus every State Machine; states, conduits, transitions (from-to) and their rule logic.
Blend Spaces: axes and every sample. IK Rigs and Retargeters, Control Rig RigVM graphs. Physics Assets (ragdolls): per-body shapes and mass/damping, constraint limits, drive parameters, projection. Sequences, Montages and Level Sequences: length, notifies, sections, slots, curve metadata. Raw keyframe dumps are intentionally omitted — you get the structure that matters, not noise.

▸ AI & BEHAVIOR

Behavior Trees exported as full recursive ASTs decorators, services and tasks with all properties. EQS queries embedded inline where referenced. Blackboard keys and types. State Trees with property bindings and editor topology.

▸ MATERIALS

Material domain, blend, shading and all root settings. Material Instances export every scalar, vector and texture parameter override plus the parent path. The layer artists actually tune. Base-material node graphs are traversed where Unreal exposes them to headless extraction; when a graph is held as editor-optional data that can't be resolved offline, the file states so instead of guessing.

▸ WORLD LEVELS

Every actor exported individually: transforms, non-default properties, full component list. World Settings and World Partition metadata. Deep World Load option for full Partition actor resolution.

▸ AUDIO

MetaSound: full DSP graph node structure. SoundCue: recursive node tree.

▸ INPUT

Every Enhanced Input Mapping with modifiers and trigger conditions. Input Actions with their value type AND which Blueprints in your project use them.
▸ DATA

DataTables, CurveTables, StringTables, Primary Data Assets. User Defined Structs and Enums, so the AI knows the exact shape of your types. The full Gameplay Tag dictionary, so it uses your real tags.

▸ MESHES & PROCEDURAL

Static Mesh metadata (LODs, bounds, triangle count, material slots, sockets, collision complexity). Metadata only, never geometry. PCG procedural-generation graphs.

▸ GAMEPLAY ABILITY SYSTEM

Ability, Effect and Attribute Set blueprints detected by parent class and exported with full data through the Blueprints module.

▸ PHYSICS

Physical Materials (surface friction, restitution and density).

▸ SOURCE CODE & CONFIG

Full .cpp / .h files. All .ini configuration. Optional plugin source inclusion.

▸ PROJECT MANIFEST
Master index of every exported asset plus a dependency map. Long asset paths are aliased to short tokens across the export, with the lookup table in the manifest.

  • FOUR EXPORT MODES, ONE CLICK

Context windows are finite. Pick the layout that fits how you'll feed the data:

- STANDARD: one .jsonl per asset, for dragging one specific asset into a chat.

- FRAGMENTED: same, split into parts capped at a Max File Size you set.

- CONSOLIDATED: one file per module, so the AI reads a whole subsystem at once.

- MONOLITHIC: a single project-wide .jsonl for RAG ingestion (the source-code snapshot stays as separate Markdown).

Delta Serialization skips every property still at its class default, so you don't spend tokens on engine boilerplate. Headless CLI flags (-DeepWorld, -MaxRAMMB, -NoAIFilter) give CI automation full parity with the editor UI.

  • THREE WAYS TO USE IT

1~ CHAT CONTEXT

Export, drag the JSONL into Claude, ChatGPT or Gemini, ask specific questions about YOUR project. Use the scope filter to export just /Game/Combat when that's all you need.

2~ RAG PIPELINES

The Monolithic export drops into LangChain, LlamaIndex or any vector database. No retraining.

3~ FINE-TUNING DATA PREP

Structured output ready for dataset preparation.

VS MCP SERVERS AND COPILOTS

You should use both.
They solve different problems. An MCP server gives live access to a handful of assets with the editor open and using commands that burn trough your tokens; a copilot sees what's on screen. Neither hands you 5,000 assets on disk. Vector databases, training sets and nightly CI jobs need bulk offline extraction — that's what this is.

  • ANALYTICS DASHBOARD

Module size and assets-per-module charts, your twenty most complex Blueprints, most referenced and heaviest-in-memory assets, and export efficiency metrics: mode, throughput, delta savings, average size per asset, true physical file count. Loadable from any previous export.

  • LIMITATIONS — PLEASE READ BEFORE BUYING

TEXT DATA ONLY: No .wav, .mp3, .png, .fbx or any binary media. "Materials", "Audio" and "Animation" export structural data and parameters — not the media files, and not raw animation keyframes.

ASSET RESOLUTION: Unsaved, corrupted or orphaned assets may not resolve at export time. An exact 1:1 match with your Content Browser count is not guaranteed.

WORLD LEVELS COST: significantly more export time and memory; Deep World Load multiplies it further.

HARDWARE: NVMe SSDs strongly recommended. Export cannot run while the engine compiles shaders or builds DDC.

LANGUAGE: In V1.1, the editor UI is English-only in this version — the previous multi-language UI was removed and may return in a future update. JSONL keys and folder names are always English for parser compatibility.

WINDOWS 64-BIT ONLY.

  • TECHNICAL SPECIFICATIONS

Supported Platforms: Windows 64-bit (Editor-only plugin).

Supported Engine Versions: 5.4, 5.5, 5.6, 5.7.

Required Plugins: Niagara, MetaSound, Enhanced Input, StateTree.
Includes the Google font "Tomorrow" (SIL Open Font License).

Enthaltene Formate