Vi3W
Skip to content

Guides

Can you chain MCP servers? Building a Vi3W → Blender → game engine pipeline in one chat

Connect several MCP servers to one assistant and run a 3D pipeline: generate in Vi3W, clean in Blender, import to a game engine. Steps and prompts.

6 min read

A glowing cable plug connecting a glass chat bubble to a small 3D scene on a turntable
Illustration. One assistant, several connections, each doing one job.

Quick answer

Can I chain MCP servers for a 3D pipeline?

Yes. An AI client such as Claude can hold several connections at once and pass results from one to the next: generate a model on a cloud service, import and clean it in Blender through a local server, then hand the file to a game engine. The assistant coordinates; each server does one job.

Why one server is rarely enough for 3D

A 3D asset passes through several tools before it ships: something makes the mesh, something fixes and arranges it, and something puts it in the game. Each of those has, or can have, its own MCP server. Because one AI client can hold several MCP connections at once, you can ask the assistant to move a model through all of them in a single conversation.

The assistant is the glue. It calls a tool on server A, reads the result, and uses it as input for a tool on server B. No server talks to another server directly; each only talks to the assistant. That keeps every server simple and lets you swap one out without breaking the rest.

If local and remote servers are new to you, read remote vs local MCP servers first, because a 3D chain almost always mixes both.

A cloud server makes the mesh. With Vi3W's MCP server connected, the assistant can generate from a text prompt or a reference image, then run Add bones, Make it move, New paint job or Clean shape on the result. The output is a model file (GLB or FBX, with OBJ, STL, PLY and glTF available) that you or the assistant download.

A Blender MCP server, either the official Blender Lab server or the community MCP for Blender, runs inside your open Blender. The assistant imports the downloaded file, applies scale, sets the origin, renames objects to your convention, places copies in the scene, and exports. Setup is in Blender MCP with Claude.

Some editors ship their own local MCP server; Roblox Studio is one, covered in the Roblox Studio MCP server guide. Where your engine has none, this link is a manual import of the exported file, which is still only one step. Engine import settings are covered in the AI 3D game asset pipeline.

Three platforms: a sketch card, a grey model and a finished mushroom house
Illustration. Prompt, base mesh, finished asset: each stage can sit behind its own server.

Setting it up in Claude

Claude Desktop is the most practical host for a mixed chain, because it can run local servers and use remote connectors in the same chat.

  1. Add the remote server. In Claude, open Customize, Connectors, Add custom connector, name it Vi3W, and paste Vi3W's MCP server URL from the Vi3W app. Full steps: Vi3W MCP in Claude Desktop.
  2. Add the local Blender server. Either enable Claude's Blender connector or add the community server to Claude Desktop's config (Settings, Developer, Edit Config, an entry under mcpServers). Start the server inside Blender.
  3. Restart Claude Desktop fully so it picks up the config.
  4. Check both are live. Ask: "Which tools do you have from Vi3W and from Blender? Don't run anything yet." You should see both sets.

ChatGPT can hold several remote servers, but it doesn't connect to local servers directly, so the Blender link needs OpenAI's Secure MCP Tunnel or a manual step. Details are in Vi3W MCP in ChatGPT.

A first chained session, prompt by prompt

Keep each step small and check it before moving on.

  1. "Using Vi3W, make a stylised wooden market stall with a striped canvas roof, game-ready, about 2.5 m tall. Show me the result before doing anything else."
  2. "Good. Run Clean shape at a lighter model size, then give me the GLB download link."
  3. Download the file to a known folder, for example ~/Downloads/market-stall.glb.
  4. "In Blender, import ~/Downloads/market-stall.glb, apply all transforms, set the origin to the bottom centre and scale it so it's 2.5 m tall."
  5. "Rename the object SM_MarketStall_01 and its material M_MarketStall_01."
  6. "Place three copies along the X axis, 4 m apart, then export the selection as FBX to ~/Exports/market-stalls.fbx with Apply Transform on."
  7. Import the FBX into your engine, or ask the engine's MCP server to do it if it has one.

That's one asset through three tools with no manual clicking in between, except the download. Naming rules are worth writing down once; AI 3D asset naming conventions has a template you can paste into your first message or a Claude skill.

Rules that keep a chain from breaking

  • Name the server in the prompt. "Using Vi3W..." and "In Blender..." stop the assistant guessing which tool to use when two servers could both plausibly answer.
  • Pass files by path. A remote server can't see your disk and a local server can't see the cloud. The handoff is a file in a folder both you and the local server know about.
  • Confirm before paid steps. Generation and the 3D tools cost XP, shown before each run. Ask the assistant to wait for your OK before each one.
  • One link at a time. Long multi-step prompts ("generate, rig, import, place, export") fail in the middle and leave you guessing which step broke. Short prompts are easier to retry.
  • Check after each link. Look at the model after generation, and at the scene after import. Catching a wrong scale at step 4 is cheaper than at step 7.

Chains worth trying

  • Prop set for a level. Generate ten props in Vi3W, import them all into Blender, apply your naming, lay them out on a grid for review. The batch side is in batch 3D props from one chat.
  • Character to animation test. Generate a character, run Add bones, then Make it move with Walk, and import the animated FBX into Blender to check the deformation on the timeline.
  • Retexture variants. Run New paint job three times on one model (wood, stone, gold), import all three into Blender side by side, and render a comparison.
  • Print prep. Generate a figurine, convert it to STL, import it into Blender and run the 3D-Print Toolbox checks before slicing.

What a chain can't do

Chaining moves work between tools; it doesn't add judgement. The assistant can't tell whether a stall reads well from your game camera, whether a rig deforms acceptably at the shoulders, or whether a texture matches the rest of your art. It also can't recover gracefully from every failure: if an import fails halfway, it may try something odd to fix it. Keep backups, keep steps short, and look at the result yourself at each link.

Start the chain

Connect the first link today: Vi3W in Claude, then Blender MCP as the second. Or make the first model in the browser and hand it to Blender yourself.

Frequently asked questions

mcp server for blender

There are two main ones: the official Blender Lab MCP server, which Claude's Blender connector is built on, and the community MCP for Blender. Both are local servers that run Python inside your open Blender scene.

mcp for 3d modeling

For 3D modelling you usually combine two kinds: a generation server that makes meshes from text or images, and a DCC server (such as Blender's) that edits, arranges and exports them.

best mcp for 3d modeling

It depends on the job. Blender's MCP servers are best for editing and scene work; a generation service's server is best for making new meshes, rigging and retexturing. Most 3D pipelines use one of each.

Last reviewed