Guides
Remote vs local MCP servers for 3D: which one does your workflow need?
Local MCP servers drive Blender and files on disk; remote ones run cloud 3D tools. How each works in Claude and ChatGPT, and which to pick.
7 min read

Quick answer
Should I use a remote or a local MCP server for 3D?
A local server runs on your own computer beside an app like Blender and talks to the AI client over stdio. A remote server runs on the internet and is reached over HTTPS. Use local to drive desktop software and files on disk; use remote for cloud services such as AI 3D generation, rigging and retexturing.
Two kinds of MCP server, one protocol
The Model Context Protocol (MCP) has one job: let an AI app such as Claude or ChatGPT call tools that live somewhere else. The protocol itself doesn't care where those tools live. The MCP architecture docs define two transports, and they are the whole difference between "local" and "remote":
- stdio for a server that runs as a program on the same machine as the AI app. The app starts it and talks to it through standard input and output. There's no network in between, and a local server usually serves exactly one client.
- Streamable HTTP for a server that runs somewhere on the internet. The app sends HTTPS requests to a URL, and one remote server serves many users at once.
Everything else, the tools, the JSON-RPC messages, the way Claude decides to call a tool, is the same in both. If you've never met MCP before, start with what MCP is, explained for 3D artists and come back.
What a local MCP server is good at in 3D
A local server is the only way to drive software that runs on your desk. The best-known 3D example is Blender. Both the official Blender Lab server and the community MCP for Blender are local: an add-on inside Blender opens a connection, a small MCP program relays tool calls to it, and Claude can then list your objects, apply transforms, or run Python inside your open scene. Setup is covered in Blender MCP with Claude.
Local makes sense when the work touches:
- An app with a live scene. Blender, or a game engine editor with its own local server, as in Roblox Studio's MCP server.
- Files on your disk. Importing
chest.glbfrom your Downloads folder, renaming a folder of FBX exports, or checking which textures are missing. - Private work that shouldn't leave the machine. With a local model, you can even keep the whole loop offline; see Blender MCP with a local LLM.
The cost is setup. You install the server (often with uv or npm), add an entry to your client's config file, restart the client, and keep the 3D app open while you work. In Claude Desktop that config lives under Settings, Developer, Edit Config, in the mcpServers section.
What a remote MCP server is good at in 3D
A remote server wraps a service that already runs in the cloud. You don't install anything. You paste one URL into your AI client, sign in, and the tools appear.
For 3D, that's the natural shape for anything that needs a GPU farm rather than your laptop:
- Generating meshes from a text prompt or an image.
- Rigging and animating a character on someone else's hardware.
- Retexturing a model with new PBR maps.
- Retopology or file conversion that would otherwise need a desktop tool.
Vi3W's MCP server is remote for exactly this reason. Generation, Add bones, Make it move, New paint job and Clean shape all run on Vi3W's servers, so the AI client only sends a request and gets back a model to download. There's nothing to keep running on your machine, and it works from a browser.
The trade-offs are the usual ones for cloud tools. You need an account with the service, your inputs (prompts, reference images, models) are uploaded to it, and the service decides what it charges. A remote server can't see your local Blender scene or the files on your disk unless you upload them.
How Claude and ChatGPT handle each type
The two big chat apps don't treat local and remote servers the same way, and this decides a lot.
Claude
- Remote servers are added as custom connectors. On Pro and Max that's Customize, Connectors, Add custom connector, then the server's URL. Custom connectors are available on Free, Pro, Max, Ultra, Team and Enterprise plans, per Anthropic's connector guide. Claude reaches the server from Anthropic's cloud, not from your computer.
- Local servers go in Claude Desktop's config file, and Claude Desktop starts them itself. They only work in the desktop app, because claude.ai in a browser has no way to start a program on your machine. Claude Code reads local servers from its own MCP configuration.
ChatGPT
- Remote servers only. According to OpenAI's help article, ChatGPT connects to remote MCP servers over a public HTTPS endpoint, on ChatGPT web, not mobile. Full MCP (including write actions) is for Business and Enterprise/Edu workspaces; Pro users can connect custom servers with read and fetch permissions.
- No direct local servers. To reach a server on a private network or a developer machine, OpenAI points to its Secure MCP Tunnel.
So if your plan is "ChatGPT drives my Blender", you're in tunnel territory and should read whether ChatGPT can control Blender first. If your plan is "ChatGPT generates a model for me", a remote server is the direct route.
A quick way to decide
Ask one question about each tool you want: where does the work actually happen?
- On my machine, in an open app or on my disk. You need a local server. Use Claude Desktop, Claude Code, or another client that can start local programs.
- On a cloud service I'd normally use in a browser. You need that service's remote server. Any client that supports remote MCP works, including claude.ai in a browser.
- Both. That's common in 3D and it's fine. Connect a remote server for generation and a local one for Blender in the same client, and let the AI pass files between them. The handoff is covered in chaining MCP servers into a 3D pipeline.
A typical combined setup looks like this. Claude calls Vi3W's remote server to make a stylised lantern from a prompt and add a new paint job. You download the GLB. Then Claude calls the local Blender server to import the GLB, scale it to 0.4 m, and place twelve copies along a path. Each server does the part it's good at.

Common mistakes when mixing the two
- Pasting a local command into a remote connector field. A custom connector wants an
https://URL. A line likeuvx mcp-for-blenderbelongs in Claude Desktop's config file, not in Customize, Connectors. - Expecting a remote server to see local files. A cloud tool can't read
C:\Assets\barrel.fbx. Upload the file, or let a local server do the file work. - Running two clients against one local server. Blender MCP setups work best with one client at a time. Close Cursor before you try Claude Desktop on the same Blender.
- Forgetting to restart. Claude Desktop only reads its config at start-up. Quit it fully after editing.
When a local connection refuses to come up, the step-by-step checks in Blender MCP not working cover most cases.
What neither type can do
MCP moves requests and results. It doesn't make the underlying tool smarter. A local Blender server lets Claude run Python, but Claude still writes that Python, and complex organic shapes built from code tend to look like primitives glued together. A remote generation server makes real meshes, but it can't arrange your level for you or know your studio's naming rules unless you tell it. And no MCP server, local or remote, removes the need to check the result: open the model, look at the topology and the scale, and decide whether it's good enough to ship.
Start with the remote side
If you mostly want models, not scene automation, a remote server is the shortest path: no installs, and it works in a browser. Connect Vi3W to Claude using the Claude Desktop setup guide or to ChatGPT using the ChatGPT setup guide, or try Vi3W in the browser first.
Frequently asked questions
local mcp server vs remote mcp server
A local server is a program on your machine that the AI client starts and talks to over stdio, usually for one client. A remote server is a web service reached over Streamable HTTP that serves many users. Local suits Blender and files on disk; remote suits cloud tools.
are local mcp servers safe
They are as safe as the code you install. A local server runs with your user's permissions, so it can read files and, in Blender's case, run Python. Install servers only from sources you trust, save your work before long sessions, and keep any network socket on localhost.
local mcp server chatgpt
ChatGPT doesn't connect to local servers directly. It talks to remote MCP servers over HTTPS, and OpenAI's Secure MCP Tunnel is the supported way to reach a server on a private network or a developer machine.
Last reviewed

