Open3D

New in v1.2 · Windows

Ask an AI to inspect and edit the file you have open.

Open3D Viewer now speaks the Model Context Protocol. Point Claude Desktop, ChatGPT, Cursor or Gemini CLI at the model in the viewer window and ask it, in plain English, to fill a hole, bridge a tube cavity, boolean-cut a cylinder through it, or restyle it in walnut. Local server, undoable edits, no upload.

Get Open3D for Windows → How the setup works

How the setup works

Open3D speaks the Model Context Protocol locally, on your own machine. AI clients like Claude Desktop, ChatGPT, Cursor and Gemini CLI connect to it, see what you have open, and can call tools to inspect or edit it.

  1. Install Open3D Viewer for Windows. The MCP bridge ships with the app — no separate download.
  2. Open Settings → Connect an AI tool in the viewer. Flip the switch on. Click Install next to your AI of choice.
  3. Restart the AI client (Claude Desktop, ChatGPT/Codex, Cursor, Gemini CLI). It picks up Open3D on next launch.
  4. Load a 3D file in Open3D. Ask your AI something like "Using the open3d MCP, list the meshes and take a screenshot."

Which AI clients are supported

Claude Desktop
Chat app · Anthropic

Free download. Sign in with Google or email. One-click install from the Open3D Settings screen.

ChatGPT / Codex Desktop
Chat app · OpenAI

Modern ChatGPT desktop app (not "ChatGPT Classic") from the Codex install. One-click config into ~/.codex/config.toml.

Cursor
IDE · code editor

Cursor's built-in chat can call MCP tools. One-click install from Open3D. Same for Windsurf and Zed.

Gemini CLI
Terminal · Google

Runs in your terminal. Open3D auto-installs it via npm on your first click if it isn't already there.

The chat app called simply Gemini (Google's desktop Companion, Alt+Space) and ChatGPT Classic from the Microsoft Store don't support MCP — there's nothing on their side to talk to. Web Gemini and web ChatGPT don't either. Modern desktop clients above are the realistic set today.

What the AI can actually do

See the model

Take screenshots from any angle, list every mesh with its triangle count and material, report geometric stats (edge length, curvature, bounding box), and probe specific points with rays — all without downloading the file bytes.

Edit geometry

Fill open holes, bridge tube-shaped cavities with quad strips, smooth bumpy regions, boolean union / difference / intersection two meshes, cut with a plane, split by connected component, merge multiple meshes back into one. Every op is on the app's undo stack — Ctrl+Z reverts.

Restyle

Set base color, metalness, roughness and opacity directly, or pick from 18 hand-tuned presets: walnut, oak, maple, pine, ebony, brass, gold, silver, copper, steel, chrome, matte black, gloss black, white PLA, grey PLA, glass, rubber, concrete.

Not yet: user-supplied PBR texture packs (albedo / normal / roughness maps). Planned.

Save the edits

Nothing writes to disk until you or the AI explicitly ask. Export as GLB or STL, reset every edit to the original state, or step forward and back through the history one op at a time.

Example prompts

Type these in your AI chat; Open3D does the rest.

Take a screenshot of the current model from the top and iso view.
This ring has an open cavity in the middle. Close it with a proper tube bridge, not a flat disc.
Make the wooden parts look like walnut and the metal parts like brass.
Punch a 5mm-diameter hole through this bracket, centered on the top face.
Fill any open holes on the model, then smooth the patched area so the fix isn't visible.
Split this scan into its connected parts, keep the biggest one, and save it as walnut_stool.glb.

Privacy: your file never leaves your machine

The MCP server runs locally on your device and isn't reachable from the outside world. Every call is gated by a bearer token generated the first time you turn the feature on. The AI vendor (Anthropic, OpenAI, Google) sees your prompt, the arguments the AI decides to pass, and the results the server returns — but the raw file bytes never get uploaded and no third-party service touches the geometry directly.

If you're testing on a confidential file (a client STEP, a vendor part) and want to be strict, rotate the token from Settings and re-install into whichever clients you still want to keep.

FAQ

Which AI assistants can talk to Open3D?

Claude Desktop, ChatGPT (the modern desktop app / Codex install), Cursor, Windsurf, Zed and Gemini CLI. Install is one click from the viewer's Settings screen — no config-file editing needed. Gemini's Companion desktop app (Alt+Space), ChatGPT Classic and the web versions of either don't support MCP.

Does the file leave my machine?

No. The MCP server binds to 127.0.0.1 and requires a bearer token. Tool call results (screenshots, mesh lists) are sent back to the AI vendor as part of the chat context, but the file bytes themselves never get uploaded.

Is the MCP server free?

Yes. The MCP server, the bundled open3d-mcp.exe bridge and every tool are part of the free Open3D Viewer app. You still pay for whatever AI subscription you use on the client side.

Can I use this on iPhone or Android?

Not yet. The MCP server ships on the Windows build first because that's where AI chat clients live. Mobile platforms don't run background HTTP servers reliably enough to make this a good experience.

What if I want to write my own MCP client?

The Advanced section of Settings gives you a local MCP URL and a bearer token. Any client that speaks the Streamable HTTP MCP transport can connect. There's also a plain REST surface for scripts that don't need the full MCP protocol.

Can the AI break my file?

Every edit is on the Open3D undo stack — Ctrl+Z reverts, or you can use the AI's own undo_edit. save_edits is the only tool that writes to disk, and the AI has to call it explicitly. If you're paranoid, save a copy of the source file before you start.