I’ve been building a workflow that lets an AI client drive the Unreal editor directly: a declarative statement goes in, real assets come out inside the editor — blueprints, materials, animation blueprints, Niagara, and so on. The editor-side work is actual editor work, not a simulation: nodes get wired, assets get created, the result compiles and saves. Up to now I’ve only been able to describe that in text. This episode is the first one where the path is on screen, so the six demos below are the point of the post rather than a footnote to it.
Workflow
The client sends a statement to ForgeMindServer over MCP; the server translates it and talks to the running editor through a local bridge; a Companion plugin performs the editor-side operations. From the editor’s point of view nothing exotic happens — the same asset creation and node wiring you would do by hand, just issued as one call. In every recording the small dark window on the left is the instruction stream, one line per tool call with a status dot on the right; the editor on the other side of the frame is responding in the same session, not a separate take stitched in afterwards.
What the demos show
Six separate recordings, one instruction set each: a Blueprint Actor subclass that spawns an Actor from its own graph; a spline component plus a world-space point query on that spline; a material that switches between two colours at a threshold; a dot product between the camera vector and the pixel normal; two DSL payloads in a single dispatch — one blueprint and one material in the same call; and the session check at the end, where the editor reports who is online and what state it is in. Each recording stays open after its script finishes so the last line can be read rather than caught mid-frame.
I want to be precise about what these frames are worth. They are a real execution screen, which is enough to back the claim that the instruction → tool → editor chain actually runs. They are not a verification result: I’m not claiming a pass rate, not quoting a case count, and not asking you to read six recordings as coverage. Several of the cases behind them are still open, and the recordings do not change that.
Measured difference
For the case I keep coming back to — a For Each with Break loop in a Blueprint graph — 20 node builds by hand versus 11 DSL lines, roughly 75% fewer operations. A single call on one blueprint replaces 50+ node-by-node operations. Version coverage is UE 5.0–5.8. Both an OpenSource (Apache-2.0) build and a Commercial build exist; the open build is not the whole product, so please read the licence before assuming otherwise.
Looking for scenarios
I’d rather test against real production pain than my own demo cases. If you have a blueprint task that is pure repetition — dozens of near-identical nodes, a batch of material instances, a stack of animation states — describe it in this thread: the scenario, roughly how many nodes or assets, and how long it takes you today. I’ll run what fits the current coverage and report back honestly, including the cases it cannot do yet.
Source: GitHub GitHub - akimoto-akira/ForgeMind_OpenSource: ForgeMind — a game engine control platform with MCP as its interface: it exposes Unreal Engine project management, asset reading, launch/build/editor operations and other capabilities to AI clients through the MCP protocol. The current implementation targets UE (5.0–5.8), and the platform design is extensible to other engines. · GitHub · Gitee https://gitee.com/hoshinokennji/ForgeMind_OpenSource
