Skip to content

Introduction

kumo is an Unreal Engine 5.7 / 5.8 plugin that gives your AI coding assistant direct access to the editor. Describe a Blueprint, an animation graph, an input action, or a UMG widget in plain English, and the AI builds, edits, or inspects it through validated engine operations — without you ever opening the asset by hand.

It comes two ways: the plugin, docked inside the Unreal editor, and the desktop app, which opens the project wiki and the chat in their own window for teammates who do not run Unreal.

If Unreal Engine can author it, kumo can build it. The plugin covers every asset type the engine can author — 100% of Unreal’s authorable asset types, across the base engine and its plugins — and for each one the AI can create new assets, edit existing ones, and inspect what’s already in the project. Engine-plugin asset types count the moment their plugin is enabled; the few types the engine offers no “create new” path for are import-only, and importing them is covered too. See the Engine Coverage report for the full breakdown.

Rather than reprint a checklist of asset types here, the Operations section in the sidebar has a dedicated page — with copy-paste prompt examples — for each area of the editor: scripting and logic, gameplay systems, animation and motion, user interface, art and assets, world building and cinematics, and profiling and debugging.

Asset types that live in an optional engine plugin (Enhanced Input, Control Rig, State Trees, Common UI, Niagara, the Gameplay Ability System, IK Rig, PCG, Paper2D, Groom, Gameplay Cameras, and more) need that plugin enabled. kumo detects the build state and disables only that area — gracefully, with a clear diagnostic — when its plugin is absent.

When you install kumo, setup.py registers a skill with Claude. The skill tells the AI what operations are available, what parameters they accept, and how to chain them.

When you make a request — say, “create a locomotion state machine with idle, walk, and run states” — the assistant:

  1. Generates a JSON operations plan from the skill
  2. Hands it to the kumo CLI
  3. The CLI validates the plan against the engine’s own schemas
  4. Generates Python and sends it to a running editor (or launches one in headless commandlet mode)
  5. Reports the result, asset paths, or a structured error back to the assistant

The editor does not have to be open. kumo launches Unreal in commandlet mode, performs the operations, and exits.

kumo ships a small markdown skill file plus a CLI; the assistant only loads the schema for the operations it’s actually about to call, instead of embedding full tool schemas into every conversation turn the way an MCP server would. Measured on a live UE 5.8.0 server, an identical “create a Widget Blueprint” task cost ~1,500 tokens with kumo vs ~3,900–9,600 with Epic’s own UE 5.8 MCP plugin (~2.6–6.4× fewer, single-domain — widening with multi-domain work). Against NeoStack AI v1.0’s published 4,000 tok/op, a typical 5–10 op session runs ~9–14× lighter. The CLI also validates every op against Unreal’s own schemas before it runs, so failures return real, typed, verified errors rather than a hallucinated success. Full measurements and reproducible scripts: kumo vs MCP.

The skill points to a curated docs tree the AI reads on demand: a full per-domain API reference, workflow guides, and a knowledge base covering GAS, replication / Iris, motion matching, Mover, Substrate materials, PCG, World Partition, Enhanced Input, and more. The AI consults the relevant guide before it builds, so what lands compiles cleanly and follows engine best practices on the first try — and it only loads the parts it needs, which is why the per-op cost stays in fractions of a cent. See Customizing the AI for the full list.

kumo supports remote control: start a build at your desk, then continue the same conversation from your phone or a browser. Messages stay in sync across every connected device while the session keeps running locally on your machine — so you can kick off a long Blueprint refactor, step away, and keep steering it from the couch.

It works because kumo bundles Anthropic’s official agent tooling and a Claude account sign-in directly into the panel — sign in once, and your in-editor sessions are reachable from anywhere. Every session is bridged by default, so “can I pick this up on the train” is answered before you think to ask; if you would rather bridge sessions one at a time, turn off Start remote control automatically.

On first editor launch, setup.py auto-installs the kumo skill for Claude, and for Codex once you sign in to it:

AssistantSkill locationIn-editor chat panel
Claude.claude/skills/unreal/✅ Bundled
Codex.agents/skills/unreal/✅ Bundled

Claude and OpenAI Codex both drive the bundled in-editor chat panel. First-class integration for Cursor, GitHub Copilot and Windsurf ships next.

Any other CLI-capable tool can already drive kumo directly; see AI Service Setup for the bring-your-own path.

  • Unreal Engine 5.7 and 5.8 (the plugin; the desktop app needs no engine)
  • PythonScriptPlugin — enabled by default in UE 5.x
  • EditorScriptingUtilities — enabled by default in UE 5.x

Optional engine plugins unlock additional asset types and are gracefully disabled when absent: EnhancedInput, ControlRig, StateTree, CommonUI, Niagara, GameplayAbilities, IKRig, PCG, Paper2D, HairStrands (Groom), GameplayCameras, and more.