Guides
Retopology from a chat: clean up AI models with the Vi3W MCP server
Retopologize AI 3D models from Claude or ChatGPT with the Vi3W MCP server: Clean shape at 500 to 20,000 faces, baked textures, and export to GLB or FBX.
5 min read

Quick answer
Can I retopologize a 3D model from Claude or ChatGPT?
Yes, by connecting the assistant to Vi3W's MCP server. It can then run Clean shape on a generated model, rebuilding the mesh at a size you choose between 500 and 20,000 faces and baking the original texture onto it, and run Make it lighter to shrink the file, before you download GLB, FBX, OBJ or STL.
The problem it solves
AI 3D generators are good at shape and bad at meshes. A generated barrel or chest often arrives as tens of thousands of uneven triangles. That's fine for a thumbnail. In a game it's heavy, in Blender it's awkward to edit, and in a web viewer the file is slow to load.

Retopology rebuilds that mesh with fewer, more even faces, and baking keeps the look in the texture. Vi3W does this on its 3D screen with Clean shape. Through Vi3W's MCP server, Claude and ChatGPT can run it for you from a chat, as one step in a longer job.
What the tools do
- Clean shape rebuilds the model at a Model size you choose, from 500 to 20,000 faces, and bakes the original texture onto the new mesh. The silhouette and the look carry over. The face count drops to your target.
- Make it lighter ("Smaller file, same look") compresses the file without changing the shape. It's free.
- Download gives you GLB, FBX, OBJ, STL, PLY or glTF.
Clean shape is a paid tool and costs XP. Vi3W shows the cost before it runs. Make it lighter and every download format are free.
One honest note on topology type: Clean shape produces an even triangle mesh at your target count. That's what game engines render anyway, and it's light and clean. If you need all-quad topology for sculpting or subdivision in Blender, run Blender's Quad remesher (QuadriFlow) afterwards. Claude can do that over Blender MCP, as shown in can Claude retopologize a mesh?
Connect Vi3W to your assistant
You need Vi3W's MCP server URL from the Vi3W app.
- Claude: Customize ‣ Connectors ‣ + Add ‣ Add custom connector. Paste the URL, sign in to Vi3W, and click Add. Custom connectors work on all Claude plans. Team and Enterprise members use the connector their Owner added. Anthropic's guide.
- ChatGPT: Plugins ‣ plus button ‣ Add custom MCP server. Paste the HTTPS endpoint, set up authentication, accept the risk warning and select Create as a plugin. Then type @ in a chat to select it. Full MCP (tools that change things) is on Business and Enterprise/Edu, on the web. OpenAI's guide.
Picking a face count
Tell the assistant a number, not "low poly". Rough starting points:
- Small background props (bottles, rocks, crates in the distance): 500 to 2,000.
- Props players look at (chests, weapons, furniture): 2,000 to 8,000.
- Characters and hero props: 8,000 to 20,000.
- 3D printing: pick by detail, not by budget. Printers don't care about face counts, but slicers and file sizes do.
- Web and AR: lower is better, and Make it lighter helps as much as the face count.
The reasoning behind these numbers: what is a good poly count for game assets.
A chat session, start to finish
- "Generate a wooden treasure chest with iron bands, stylised, for a mobile game."
- "Run Clean shape on it at 3,000 faces."
- "Make it lighter, then give me the GLB."
Or, for a batch:
"Here are five prop ideas: a lantern, a barrel, a stool, a sack of grain and a wooden sign. Generate each, run Clean shape at 1,500 faces, and give me FBX downloads named after the props."
The assistant handles the sequence and naming. Vi3W does the mesh work. Look at each preview before moving on. The assistant reports what the tools returned, but it can't judge whether the lantern's handle survived.
If a result loses detail you care about, ask for a higher face count. A thin handle at 1,500 faces may need 3,000.
Before rigging and animation
Rigging a dense mesh is slow and deforms badly. A cleaner, lighter mesh rigs better. The order that works:
- Generate.
- Clean shape at a character-friendly count (8,000 or more for a humanoid).
- Add bones, then Make it move.
That flow is covered in rig and animate from chat with the Vi3W MCP. For why joints need even topology, see AI character topology for animation.
Clean shape, Decimate or hand retopology?
- Clean shape (Vi3W): AI-generated models, in the browser or from a chat. It rebuilds the mesh, bakes the texture and needs no Blender.
- Decimate (Blender): quick reductions on static meshes you already have in Blender. It keeps the existing texture only if UVs survive.
- Quad remesh (Blender): all-quad meshes for further modelling or subdivision.
- Hand retopology: characters whose faces and joints must deform well, and hero assets where edge flow matters.
The trade-offs: decimate vs retopology and Blender AI retopology.
What it can't do
- No hand-placed edge loops. Clean shape is automatic. Faces and shoulders for close-up animation still benefit from manual retopology.
- No all-quad output. It produces an even triangle mesh. Use Blender's Quad remesh if you need quads.
- It works on Vi3W models. The tools run on models you generated or opened in Vi3W, not on arbitrary files elsewhere on your machine.
- Very low targets lose detail. At 500 faces a complex model becomes a silhouette. Pick the count for how close the camera gets.
- The assistant can't see quality. Check every preview yourself.
Then refine in Blender
Download FBX or GLB and keep working: adjust UVs, add LODs or hand-fix a problem area. For reductions on models you already have in Blender, read asking AI to cut poly count. For repairs before printing, read can ChatGPT fix non-manifold geometry?
Frequently asked questions
mcp for 3d modeling
An MCP server gives an AI assistant tools to call. For 3D modelling there are two kinds: servers that drive a desktop app such as Blender (local, next to the app), and servers that run a cloud service such as Vi3W, which generates, rigs, retextures and retopologises models remotely.
ai 3d model topology
Raw AI 3D models usually come out as dense, uneven triangle meshes that are heavy and awkward to edit. Retopology rebuilds them with fewer, more even faces, and baking keeps the surface detail in the texture.
mcp server for 3d modeling
Pick by where the model lives. Blender MCP servers for scene editing in Blender; the Vi3W MCP server for generating models and cleaning, rigging and retexturing them without installing anything.
Last reviewed

