A few weeks ago I sat down with a simple but slightly insane thought:
What if I could type “Build me a third-person character blueprint with a follow camera and a basic mesh”… and an AI just did it? No ChatGPT spitting out code I have to paste. No manual drag-and-drop in the editor. Just set a goal and walk away while it works inside Unreal.That’s Cipher. It’s not another “AI for game dev” wrapper. It’s an autonomous agent that lives in your Unreal project, reads your high-level goal, plans its own steps, executes them through the Python API, checks its work, and keeps going until the job is done.Why this actually feels differentMost AI tools I’ve tried are fancy autocomplete or chatbots. They hand you a blueprint graph screenshot or a wall of nodes and say “good luck implementing this.” Cipher doesn’t hand you anything—it does the thing.It connects three pieces that talk to each other in a tight loop:AI planner (the brain)
Command system (the translator)
Unreal Engine (the hands)
Right now the brain is Qwen running locally via Ollama—fast, private, zero cost. Later I’ll swap it for Claude when I want deeper reasoning. The whole thing is deliberately modular so the “brain” can be upgraded without touching the rest of the system.How the loop actually works (it’s simpler than it sounds)You give Cipher a goal in plain English:“Create a new Actor blueprint called RotatingCube that has a static mesh component using the default cube and rotates slowly on the Z axis with a point light attached.”
Here’s the pipeline:Goal → Qwen (via Ollama)
Qwen outputs command.json — structured, no hallucinations allowed
Unreal Python script reads the JSON and executes it (create BP, add components, set properties, save asset)
Script writes response.json with what actually happened + current editor state
Back to Qwen: it reads the response, evaluates progress (“70% done, missing rotation logic”), and decides the next command
Loop until the goal is marked complete
The whole thing runs hands-free. I’ve literally started it, gone to grab coffee, and come back to a finished blueprint.Here’s a tiny peek at what a command looks like (real example from a recent run):json
{
"action": "create_blueprint",
"params": {
"name": "RotatingCube",
"parent_class": "Actor",
"components": [
{
"type": "StaticMeshComponent",
"mesh": "/Engine/BasicShapes/Cube.Cube",
"relative_location": [0, 0, 0]
},
{
"type": "PointLightComponent",
"intensity": 5000
}
],
"next_steps": "add rotation logic and test"
}
}
The debugging journey (aka the part where I wanted to throw my keyboard)This was not a smooth “I wrote it in one evening” story. It was messy and I loved it.First big headache: Unreal’s Python API is powerful but extremely picky about editor context and transactions. I spent an embarrassing number of hours chasing “stale callback” errors because I wasn’t properly committing changes or refreshing the asset registry. The AI would happily command something that worked in my head but failed silently in the editor.Then the command pipeline kept breaking in fun ways:AI would invent node names that don’t exist
JSON would get malformed when Qwen got excited
Execution would succeed but the response.json would be empty because I forgot to flush the output
Each time I’d fix one thing, the loop would run a little longer. The first time it successfully created a blueprint and added a camera component without me touching the mouse, I actually said “no way” out loud. The next breakthrough was when it read its own response, realized the rotation wasn’t there yet, and added the timeline + rotation nodes on the next iteration. That moment felt like magic.I kept a little “victory log” in my notes. The best entry:
“2026-03-28 2:14am — it fixed its own mistake. I didn’t even tell it what was wrong.”What Cipher can do right nowCreate brand-new blueprints from scratch (any parent class)
Add and configure components (Camera, StaticMesh, SkeletalMesh, PointLight, etc.)
Modify existing assets (change materials, variables, default values)
Evaluate its own progress and self-correct
Run completely autonomously until the goal is marked “complete”
I’ve used it to spin up quick prototypes: a simple third-person pawn, a rotating collectible with VFX, even a basic door that opens when you overlap a trigger. Nothing production-ready yet, but way beyond what I expected from a local model.What’s nextThe obvious upgrade is swapping the planner for Claude (or whatever frontier model is best when you read this). Claude’s reasoning is noticeably better at long-horizon planning and catching its own logical gaps. I’m also building better memory between runs so it can remember “last time I built a player character this way—let’s reuse that pattern.”Longer term? I want it building entire small scenes or simple gameplay loops without me babysitting every step. The dream is an AI that can take a one-paragraph game design doc and spit out a playable vertical slice while I sleep.The bigger pictureWe’re still early, but the direction is clear: AI isn’t just going to help us build games—it’s going to start building them with us, then eventually for us. Cipher is my small bet on that future.If you’re playing with Unreal Python, AI agents, or just love the idea of autonomous tooling in engines, I’d genuinely love to hear from you. Drop a comment, share your own war stories, or hit me up if you want to poke at the repo (it’s messy but public soon).Building this has been equal parts frustrating and ridiculously fun. Can’t wait to see where it goes next.— Rez
(just a guy who likes making computers do the boring parts)
Top comments (0)