Unity MCP servers compared

Unity AI Assistant

Unity's own, and the only one here that is not open source. It needs Unity 6.3, which rules it out for any project still on a 2022 or 2021 LTS — that is most shipping projects, because upgrading an LTS mid-production rarely pays. It bills in credits through the Unity Dashboard, and it routes through Unity's relay and Unity Cloud, so it does not work offline.

What you get for that is a single vendor supporting the whole stack, with no third-party package in the project and nobody's open-source release schedule to track.

MCP for Unity (Coplay)

MIT, free, and the most established of the three open-source options: version 10, an active Discord, an ACM Technical Communications paper, and distribution through OpenUPM and the Asset Store. It reaches back to Unity 2021.3 LTS, further than anything else here.

It requires Python 3.10 or newer, installed via uv, and runs the MCP server as a separate Python process. That is a well-trodden design and also a second runtime to install, keep on PATH and debug on every machine and CI image.

Its v10 added AI asset generation — images, audio and models produced from a prompt and imported into the project. KitWright has no equivalent and is not planning one.

It has an execute_code tool built on the same in-memory Roslyn idea as KitWright's, so that is a difference in default rather than in invention: theirs sits in the scripting_ext group, and the server enables only core by default, so it is opt-in.

AI Game Developer (Murzak)

Apache 2.0, and the only one that runs inside a built game. You can ship it in a player and have an LLM drive NPC behaviour or debug a running build. That is a different product category, and if it is what you need the comparison ends there.

Its server is a separate prebuilt binary placed under Library/mcp-server/, so nothing extra to install, but a second process to supervise. Being standalone is what lets it deploy the way any server does: stdio or streamableHttp, an official Docker image, cloud hosting, and none/token/OAuth authentication.

Its reflection-method-find and reflection-method-call pair locates and invokes any C# method in the project, including private ones. KitWright has no direct equivalent. It needs the same Unity 2022.3 KitWright does, so neither has a version advantage over the other.

Where KitWright fits

Nothing to install outside Unity and nothing running beside it — one package, one dependency the Package Manager already resolves, and a C# server on 127.0.0.1:8765 inside the Editor process. That matters most on locked-down machines, on teammates' machines you do not control, and in CI.

execute_code is the primary interface rather than an opt-in extra, and the rest of the design assumes it: implement IKitWrightCommand and everything a snippet creates, modifies or destroys joins editor Undo automatically, with logs, created, modified, destroyed and returnValue in the response. The bet is that an agent which can undo its own mistakes and read the consequences of a change is worth more than a wider set of ways to invoke one.

In play mode it simulates keyboard and mouse input as well as entering and capturing, so an agent can press the button, take the frame, read the console and check whether the score actually changed — rather than reporting that the code compiled.

The ceilings are real and listed above: no runtime support, no remote hosting, no asset generation, and less history behind it than any of the other three.