Mullin Works - AgentCraft — AI Editor Agent

Distribution Product Discussion
1

AgentCraft — AI Agent for the Unreal Editor (BYOK, Multi-Provider)

AgentCraft integrates an AI agent directly into the Unreal Editor. Describe the intended change in natural language — for example, "create a worn cyberpunk metal material" or "a pickup coin that spins and disappears on overlap" — and AgentCraft executes the request through native editor tools, displays each tool invocation as it runs, and reads the resulting assets back from the asset registry so that every change can be verified.

Scope and evidence

  • Clearly defined support boundaries. The composer panel organizes capability into domain tabs. Version 1 ships three fully supported domains — Materials, Gameplay (Blueprints), and Lighting — each backed by an automated test suite that passes five consecutive full runs before release. Roadmap domains (World/PCG, Cinematics, Characters) are visible in the interface but explicitly labelled as experimental; they are functional but have not yet passed the v1 release gate.

  • Verified results. Tool calls do not simply report success. AgentCraft reads created assets back — components, parameters, asset paths — and presents this evidence in the conversation.

  • Confirmation and dry-run. Destructive operations request confirmation on a per-call basis. This confirmation can be disabled once a workflow is trusted. Standard source-control practice is recommended for reverting changes.

Provider flexibility (bring your own key)

AgentCraft does not require a subscription. Configure a supported provider with your own API key; usage is billed directly by the provider.

  • HTTP providers: DeepSeek, Qwen (DashScope), GLM (BigModel), and OpenRouter (which provides access to Claude, GPT, Gemini, Llama, and other models through a single key).

  • A per-request cost estimate is displayed before each request is sent.

Built-in MCP server

AgentCraft includes an MCP server (HTTP, specification 2025-11-25). External MCP clients — Claude Desktop, VS Code, MCP Inspector, and others — can connect to the editor and access the complete tool surface through a single local endpoint. A companion npm bridge (agentcraft-mcp-bridge) is available for stdio-only clients.

Engineering standards

  • Pure Slate user interface; no embedded web browser and no CEF dependency.

  • More than 480 automated tests. The supported tool set additionally gates on end-to-end transport tests and five consecutive full-suite runs.

  • Bilingual user interface (English / Simplified Chinese).

Required engine plugins (dependencies)

AgentCraft enables the following engine plugins on install; all ship with Unreal Engine and require no separate download: Editor Scripting Utilities, Enhanced Input, Niagara, IK Rig, Control Rig, Python Editor Script Plugin, StateTree, PCG, Pose Search, Gameplay Abilities, MetaSound, Environment Query Editor, Struct Utils, Audio Capture, Smart Objects, Geometry Processing, Geometry Script, and Modeling Tools Editor Mode.

2

A few corrections and additions to the opening post — it was generated from my Fab listing copy, and two details in it are out of date. Better to fix them here than let anyone hit a surprise after buying.

Correction 1 — per-request cost estimate. The opening post says a cost estimate appears per request. That estimate was removed in a July build: showing a per-message average alongside the running total turned out to be noise. What the panel actually shows is accumulated cost per conversation, in the session list on the left. Apologies for the stale line — the listing copy predates the change.

Correction 2 — provider list. The opening post lists DeepSeek, Qwen (DashScope), GLM (BigModel) and OpenRouter. There are six: those four plus OpenAI and Anthropic as first-class providers. All BYOK, all billed directly by the provider, no subscription.

Two things worth adding that the listing copy doesn’t cover well:

Tools are scoped per tab, not dumped on the model. The panel splits capability into domain tabs, and each supported tab exposes a curated set rather than the whole surface — 41 tools for Materials, 11 for Gameplay (Blueprints), 8 for Lighting. Each set is asserted by a contract test, so it can’t silently drift as tools get added. This matters more than the headline count: a smaller, relevant tool set is why the agent picks the right call instead of guessing among hundreds.

On the headline count — there are 300+ registered tools, and 271 of those are present in any project regardless of which engine plugins you have enabled. The number isn’t padded with optional dependencies.

Bearer auth on the MCP server is on by default. Fail-closed, which is the right default but it does mean the first thing that happens if you follow a generic MCP setup guide is a 401. The token location and how to pass it are both in the docs — worth two minutes before you wire up a client.

Snipaste_2026-07-17_09-47-20
Snipaste_2026-07-17_09-47-201258×1482 142 KB

Fab_EN_03_Tool_Execution
Fab_EN_03_Tool_Execution1920×1080 301 KB

Snipaste_2026-07-29_11-19-44
Snipaste_2026-07-29_11-19-441920×1146 303 KB

Still the thing I’d most like feedback on: which of the experimental domains you’d want promoted first, and what the agent gets wrong on your projects. Read-back verification means a wrong result is usually visible rather than silent — if you catch one, the conversation log will show exactly which tool call did it, and that’s the most useful bug report I can get.

— Mullin