Guides
What is MCP? A plain-language guide for 3D artists
What is MCP? How the Model Context Protocol lets Claude and ChatGPT use Blender, Roblox Studio and 3D generators, in plain language for artists.
6 min read

Quick answer
What is MCP?
The Model Context Protocol is an open standard that lets an AI assistant use outside tools. A program called an MCP server lists what it can do, such as run Python in Blender or generate a 3D model, and the assistant calls those tools when you ask. Claude, ChatGPT, Codex and other apps all speak it.
The one-paragraph version
An AI chat on its own can only talk. It can explain a Bevel modifier, but it can't add one to your mesh. MCP, the Model Context Protocol, is the standard way to give it hands. Someone writes an MCP server for a tool, say Blender, Roblox Studio or a 3D generator. The server tells the assistant "here's what I can do and what I need". When you ask for something, the assistant picks the right tool and the server does the work. Anthropic released MCP as an open standard in November 2024. OpenAI, Google, Microsoft and others now support it.
If you've ever used a USB device, you already get the idea. One kind of plug, many devices, and they work with any computer that has the socket.
The parts, in 3D terms
The official MCP architecture page names three participants. Here they are, translated:
- Host: the AI app you're typing into. Claude Desktop, ChatGPT, Codex, Claude Code, Cursor.
- Client: a connection the host keeps open to one server. One host, many clients: you can have Blender and a 3D generator connected at once.
- Server: the program that does things. The Blender MCP add-on and its server, Roblox Studio's MCP server, Vi3W's MCP server.
A server can offer three kinds of thing:
- Tools: actions the assistant can take. "Run this Python in Blender." "Generate a model from this image." "Add bones." Most 3D servers are mostly tools.
- Resources: information the assistant can read, like a scene summary or a list of your recent models.
- Prompts: ready-made templates the server suggests for common jobs.
Messages between them use JSON-RPC 2.0. You'll never see that, but it's why any host can talk to any server.
Local and remote servers
This distinction explains most of the setup confusion you'll hit.
Local servers (stdio)
A local server runs on your computer, usually started by the AI app itself, and talks to it over standard input and output (stdio). Blender MCP works this way: your client launches the server, and the server relays to the Blender add-on over a local connection. Local servers can reach things only your machine can see, like your open Blender, your files and Roblox Studio. Typically one client uses one local server.
Remote servers (Streamable HTTP)
A remote server runs on the internet and is reached over HTTPS using the Streamable HTTP transport. Hosted services work this way, including Vi3W's MCP server, and one remote server serves many people. You add it by pasting a URL.
Why this matters
Not every AI app can use both kinds:
- Claude Desktop runs local servers from its config file and takes remote ones as custom connectors (Free through Enterprise plans).
- Claude Code and Codex run local servers from your terminal.
- ChatGPT on the web connects to remote servers only. A local server needs OpenAI's Secure MCP Tunnel. Full MCP with write actions is for Business and Enterprise/Edu, and Pro is read and fetch only.
So "Blender MCP doesn't work in ChatGPT" usually isn't a bug. It's a local server meeting a remote-only client. More on that in can ChatGPT control Blender?.

What MCP looks like in a 3D workflow
Here's a realistic afternoon with two servers connected, a 3D generator and Blender:
- "Make a stylised stone well, game prop, about 4,000 faces." The generator server makes the mesh.
- "Give it a mossy retexture." Same server, retexture tool.
- "Import
well.glbfrom my Downloads folder into the open Blender scene, put it at the village centre, and scale it to 1.6 m tall." You download the GLB; the Blender server runs the import and transform Python. - "Add five barrels around it, randomly rotated." The Blender server again.
- "Screenshot the viewport." Blender server.
Each step is one tool call. You review between them. That rhythm (one request, check, next request) is how MCP work goes well.
Words you'll meet in setup guides
- Connector (Claude) and app or plugin (ChatGPT): the product's name for a connected MCP server.
- Config file: where desktop clients list local servers. Claude Desktop has one under Settings ‣ Developer ‣ Edit Config, with an
mcpServerssection. - uv / uvx: a Python tool many local servers use to install and run themselves. MCP for Blender starts with
uvx mcp-for-blender. - OAuth: the "sign in with…" flow remote servers use, so you never paste a password into the chat.
- Tunnel: a secure link that lets a cloud client like ChatGPT reach a server on your own machine.
- Tool call: one action the assistant takes through a server. Clients usually show each one and often ask before running it.
How to try it today
- Blender: Blender MCP explained, then setup with Claude.
- Roblox Studio: Roblox Studio MCP server.
- ChatGPT: connect an MCP server to ChatGPT.
- Which servers are worth it: best MCP servers for 3D.
What MCP can't do
- Make a model smarter at 3D. MCP gives the assistant tools. It doesn't improve its eye. A bad brief still gets a bad model.
- Make tools safe. A server that runs Python in Blender can do anything Python can. Save your work, use servers from publishers you can identify, and read the tool list when you connect. OpenAI warns that untrusted servers increase exposure to prompt injection.
- Work without the thing it controls. Blender MCP needs Blender open with the add-on's server started. Roblox Studio MCP needs Studio running.
- Be free of costs. Tool descriptions and results go through the model and use your plan's allowance. The service behind the server may charge too.
MCP and skills: the pair to know
You'll see "skills" next to MCP everywhere. They solve different problems. A skill is written instructions the assistant loads when a task matches: how you work. An MCP server is a live connection to a tool: what the assistant can do. A good setup uses both. For example, a skill that says "always apply scale and put the origin at the base before exporting", plus a Blender MCP server to actually do it. See ChatGPT skills vs Claude skills.
Put a 3D generator on the other end
Vi3W's MCP server gives Claude and ChatGPT text-to-3D and image-to-3D, plus Add bones, Make it move, New paint job and Clean shape, with GLB and FBX downloads. It's a remote server, so you add it with a URL and there's nothing to install. Try Vi3W.
Frequently asked questions
what is mcp server
An MCP server is a program that offers tools, data or prompt templates to an AI app over the Model Context Protocol. A Blender MCP server, for example, offers tools to read your scene and run Python in it. The AI app decides when to call them.
mcp server vs api
An API is written for programmers, and every app has to be coded against it. An MCP server describes its tools in a standard way that any MCP-capable assistant can discover and use without custom code. Many MCP servers are a thin layer over an existing API.
does mcp server consume tokens
Yes, indirectly. Tool descriptions, your requests and the results the server sends back all go through the model, and they count towards your usage. Big outputs, like long scene dumps or many screenshots, use the most. The service behind the server may also charge for its own work, such as a model generation.
mcp server vs skills
A skill teaches the assistant how to do something: your export rules, your naming scheme. An MCP server gives it the ability to act: open Blender, generate a mesh. You often want both, with a skill that explains how to use the server's tools well.
Last reviewed

