Kimi K3 + Unity CLI: 10 Games Built by an AI Agent

Tools used in this article

UnityKimi K3Unreal EngineVibeUE

Can an AI agent build a working game in Unity on its own? Yes, and it is now boring enough to measure. Stefan handed Kimi K3 ten written briefs and Unity’s new command-line interface, then walked away. Eight came back as playable Unity builds: 22 prompts, about 23 hours of unattended agent time and $103.34 of API tokens in total. Below: the setup, the per-game numbers, and the part where the same model on Unreal Engine lost badly.

A model that writes C# is not new. It has been able to do that for two years. What changed in July 2026 is that Unity gave it a way to run the C#, look at the running game and fix its own mistakes, without a human relaying console output back and forth.

{
  "mcpServers": {
    "unitycli": {
      "command": "C:/Users/<you>/AppData/Local/Unity/bin/unity.exe",
      "args": ["mcp", "--project-path", "C:/Users/<you>/kimi-test"],
      "env": { "NO_PROXY": "*", "UNITY_NO_BANNER": "1" }
    }
  }
}

Watch all ten tests

Every brief, every result and every number, on YouTube.

1. What the Unity CLI actually changes

Unity shipped the CLI on 20 July 2026. It is a single self-contained unity binary that installs editors, resolves project versions, manages modules and handles auth from the terminal. Useful for CI on its own. The part that matters for agents is the layer above it.

Two properties turn that into an agent loop. The output is structured, so a model gets parseable results instead of scraped console text. And the Editor is self-describing: run unity command with no arguments and it reports the operations it currently exposes, so the agent discovers its own capabilities at runtime instead of working from a hardcoded list.

2. Install it in about ten minutes

$env:UNITY_CLI_CHANNEL='beta'; irm https://public-cdn.cloud.unity3d.com/hub/prod/cli/install.ps1 | iex

3. Ten briefs, eight Unity builds, and the numbers

Each test was one written brief, no reference project, no starter template, primitives only and no imported assets. Every brief ended with an instruction to verify the result before reporting back, along the lines of “actually play it before you say it works” or “run a few waves yourself before telling me it works”. The follow-up prompts were visual passes and bug fixes, not rebuilds.

Costs are recalculated at Kimi K3 API rates ($3 in, $0.30 cached, $15 out per million tokens). Stefan ran the tests on a subscription, which works out cheaper; the API figures exist so the numbers are comparable.

The ones that worked

Stack Tower was the first test and set the tone: two prompts, one to build and one to upgrade the visuals, and the result is playable and fun. Slabs slide back and forth, the overhang gets sliced off and falls away as debris, colours cycle as the tower grows. The camera sits slightly left of centre and the stacking animation is plain, but nobody had to touch a line of it.

Spider Walker is the one to steal for your own project. A six-legged walker with no animation clips and no Animator: every leg is posed in code with inverse kinematics, feet plant on the ground and stay planted while the body moves, and the gait alternates in a tripod pattern. Setting that up by hand for a custom character is an afternoon of bone work in any engine. Here it was three prompts and 1 hour 49 minutes, and two of those prompts were visual polish.

Ping Pong is the plainest brief of the ten and the one that carries the most weight, because Stefan gave the identical prompt to a second agent on a different engine so the two could be cut side by side. The Unity version came back in two prompts and 1 hour 34 minutes: neon on dark, one accent colour per side, a trail on the ball, a pop and a small shake on impact, first to five. A square artefact still flickers on some hits. Cheapest but one, at $8.57.

Shader Garden asked for five effects hand-written in CG with no Shader Graph: dissolve, hologram, lava, toon and waves, plus a plain lit sphere as the control. Four of the five landed. The toon shader is the miss, and a sphere is a poor subject for a toon shader, so the brief carries some of that. The agent’s own test was a good one: every effect that failed to compile would have shown up as Unity’s magenta error material.

Cross Road is the most complete build of the set and the most expensive Unity one. Endless recycled lanes, a chicken that hops one grid square per key press, cars, rivers with rideable logs, a rail lane with a train that warns before it arrives, and a camera that creeps forward so falling behind kills you. It was playable on the first pass, which is why it earned two visual passes and a bug fix on top: four prompts, 5 hours 8 minutes, $22.38.

NavMesh Defense is the least glamorous build here and the most useful. Enemies pathfind on a live NavMesh that towers carve at runtime, with no scripted waypoints, and the agent worked out the shortest route through the winding walls without being told anything about routing. The balance is a mess, cannon damage is odd and the health bars are two pixels tall, but the systems layer is correct.

The ones that fell short

Suika Merge is the honest failure. The fruit-merging jar looks fine in 3D and everything moves, but fruit of the same tier sometimes refuses to merge, and three prompts across 4 hours 37 minutes did not fix it. This is the cheapest concept in the whole set and it took the second-longest Unity run.

Beat Pulse synthesised every voice in code with no audio files in the project: kick, hat, bass and pad, at 92 BPM, with a visualiser reacting to each. The music is structured rather than random noise, which is the bar the brief set, but the visuals stayed basic and the mix has no limiter. Cheapest run of the ten at $7.98.

4. Where Unreal Engine lost

Two of the ten ran in Unreal through its MCP instead, with the same model and the same style of brief. They are the two most expensive rows in the table, and they are the reason the other eight stayed in Unity.

The cleanest evidence is not in that table at all. Ping Pong was run twice from the same prompt: once through the Unity CLI, once through VibeUE in Unreal. Same model, same brief, same day. Unity took two prompts. The Unreal counterpart needed far more time and far more revisions to reach the same place, played a little laggier, and lost the look the brief asked for: no glow anywhere, a score you cannot read at 1080p, win text sliced by the centre line, paddles down to thin slivers.

The demolition range was meant to be a Niagara test: a cannon, a brick wall, a stone tower and a wooden shed, with a heavy ball that knocks things apart and slow motion on a key press. The destruction physics came out fine. The Niagara effects came out so dense you cannot see the cannonball. It took about seven hours.

Fireworks Night is the single most expensive run in the video: 4 prompts, 8 hours 38 minutes, 928k tokens in and $29.99. It ran overnight, crashed the editor three times and had to be resumed. The fireworks themselves look good. The lake reflection is not a reflection at all, just duplicated particles falling the wrong way, which is exactly the sort of thing a screenshot-based review pass fails to notice.

What this actually costs

None of these ten is a shippable game. Balance is off, HUDs are ugly, one merge rule never got fixed and the tower defense needs its numbers redone by hand before anyone would call it playable. What the numbers do say is that the prototype stage of a small 3D idea now costs a tenner and an afternoon you spend elsewhere, and that the engine you pick changes that price by a factor of three. Judgement, taste and balance stay yours. Typing the systems layer no longer has to be.

More on agents inside engines: Claude Code plus Godot MCP, Unreal Engine 5.8’s native MCP, running a local LLM against an engine, and a full UE5 level in one day.

Frequently asked questions