MCP 3D modeling
Mixos is an MCP server with a full 3D suite behind it. Add one URL to Claude or ChatGPT, approve it once, and it can build geometry, assign materials, place lights and cameras, key an animation and path-trace the result — in a tab you have open, on your own plan, with no API key to paste.
420+ commands across Studio, Mold and Paint — the same ones the app's own menus and shortcuts run on. Your assistant is handed the tools of whichever app you connected, so a sculpt session gets brushes and remeshing rather than a scene graph.
Queries for the scene contents, the current selection and a viewport capture come back before anything is changed — and every result reports what actually happened, including when the app clamped or refused, rather than what was asked for.
No local server to run, no addon to install, no Python environment to keep working. The 3D app is a browser tab and the connector is a URL, so the whole setup is two minutes and survives a machine change.
Connecting is a deliberate press in one tab, never automatic. Anything that deletes or overwrites stops in the tab and names the object first. Generating with AI is a separate permission because it spends credits, and access revokes in one click.
Why Mixos
Most 3D MCP servers are a bridge to software you already have to run. You install the desktop application, install an addon into it, start a local server, and keep all three in step — and the assistant can only work while that machine is on. The bridge is the easy part; the standing infrastructure is what makes it a project rather than a setting.
Mixos inverts that. The application is a browser tab, so the connector talks to something that is already running and already signed in. Setup is a URL in your client's connector settings and one approval screen. There is no addon version to match, no port to forward and no Python to keep alive, and the same connection works from a different computer the moment you sign in there.
What it can do is the whole app, not a demo slice: create and edit meshes, boolean and modifier stacks, materials and node graphs, lights, cameras, physics, cloners and particles, rigging and mocap retargeting, a real timeline, and a path-traced render at the end. It reads the scene before it acts and tells you what actually changed — so when the app clamps a value or refuses a command, that is what comes back, not a claim that it worked.
Claude — claude.ai, Claude Desktop and Claude Code — and ChatGPT as a custom connector on Plus, Pro, Business, Enterprise and Edu. One connection serves both: they are separate grants with separate permissions, and revoking one leaves the other working.
No. The model runs on the plan you already pay for in your own client, so nothing is metered here and there is no key to paste. Mixos never sees your conversation.
For changes, yes — commands run in a tab you have pressed Connect in, because your document lives in your browser. With nothing open the connector still answers what Mixos can do and searches every command across all three apps, and it tells you no tab is connected rather than guessing.
Not silently. Any command that deletes or overwrites part of a scene stops in the Mixos tab and asks first, with the object named — that prompt is ours, and it happens whatever your client decides to confirm on its side. Every change is a normal undo step.
Same commands, opposite billing. The in-app assistant uses a key you paste; the connector uses the client you already pay for, with its own memory and files alongside your scene.
Read next
Vertices, edges and faces — the box-modelling fundamentals.
Which of the two approaches fits the thing you're making.
Which format keeps materials and which throws them away.
Why the same texture reads bigger on one model than another.
Keep exploring
Add the connector in Claude or ChatGPT, approve it once, and start describing what you want built.