Creating 3D Models with Claude AI and Blender — From Conversation to .blend File

Download Blender from blender.org — it is free, open-source (GPL license), and runs on Windows, macOS, and Linux. Install version 4.0 or newer. Blender includes a full Python interpreter and a comprehensive Python API that allows every operation in the program to be scripted. This is what Claude uses to create 3D models — every button, menu, and tool in Blender has a Python equivalent.

After installation, open Blender once to confirm it launches. You will see the default startup scene with a cube, camera, and light. Close it — for the AI workflow, Blender will be controlled remotely.

There are two ways to connect Claude to Blender. Choose the one that fits your setup:

Method A — Blender MCP Server (Claude Desktop): Install a Model Context Protocol (MCP) addon in Blender that turns it into a local server. Claude Desktop connects to it directly and can execute Python code, take screenshots of the viewport, and read scene information — all while you watch the model build in real time in the Blender window.

Method B — Headless Blender API (Claude Code): Run Blender in headless mode (no GUI) wrapped by a FastAPI server. Claude Code sends Python scripts via HTTP, Blender executes them and returns file outputs. This is faster for batch work and runs on a remote server, but you do not see the viewport live — you review the finished .blend file afterward.

Method A is better for interactive design where you want to see each change. Method B is better for automated pipelines, server-side generation, and working without a display.

Install the Blender MCP addon from GitHub (search for 'blender-mcp' — the addon provides an MCP-compatible server that runs inside Blender). In Blender: Edit → Preferences → Add-ons → Install from Disk → select the downloaded .zip. Enable the addon. A 'Blender MCP' panel appears in the sidebar (press N to toggle the sidebar, then select the MCP tab).

Click 'Start Server' in the MCP panel. Blender is now listening for connections from Claude. In Claude Desktop settings, add the Blender MCP server as a tool source (the addon documentation provides the configuration snippet). Once connected, Claude gains access to tools like execute_blender_code, get_scene_info, and get_viewport_screenshot — everything needed to design 3D models through conversation.

Create a FastAPI server that wraps headless Blender. The server accepts Python scripts via a POST /execute endpoint, runs them in Blender's background mode (blender --background --python), and returns the output files. A minimal server needs just one endpoint:

Start Blender with: blender --background --python-expr "exec(script)" where script is the Python code received from Claude. Save the output files (.blend, .stl) to a job directory and serve them via a GET /files endpoint. Claude Code calls these HTTP endpoints to send scripts and retrieve the finished model files.

This is the same architecture used for AI image generation pipelines (ComfyUI) — a headless creative tool wrapped by an API, running on a server with GPU access.

Tell Claude what you want to build in plain language. Be specific about dimensions, materials, and physical properties. For example:

'Design a 1 kg cross-peen blacksmith hammer head. The striking face should be a rounded square, the peen should be horizontal (thin in Z, wide in Y) for pushing metal. Add Elder Futhark runes ᚦᛟᚱ (TOR) engraved 3 mm deep on both flat sides. Scale the geometry so the volume equals exactly 127.4 cm³ (1 kg of hardened steel at 7850 kg/m³ density). Include an oval eye hole for the handle.'

Claude will translate this description into a sequence of Blender Python operations. The more precise your description, the closer the first iteration will be to what you want. Use real measurements (millimeters, grams, degrees) rather than relative terms ('make it bigger').

Claude creates a Python script that uses the Blender Python API (bpy module) to build the model. A typical script sequence for the hammer head:

1. Create a cube primitive and scale it to the rough hammer dimensions
2. Subdivide the mesh to add geometry for shaping
3. Loop through vertices to sculpt the cross-peen taper, striking face profile, and center bulge
4. Apply a Subdivision Surface modifier for smooth curves
5. Cut the eye hole using a boolean DIFFERENCE operation with a cylinder
6. Calculate the mesh volume and apply a uniform scale factor to hit 1 kg of steel
7. Create small cube cutters for each rune stroke and boolean-subtract them from the surface

Claude sends this script to Blender (via MCP or API). Blender executes it and produces the 3D geometry. With MCP (Method A), you watch the model appear in the viewport in real time.

Inspect the model. With Method A, orbit around it in the Blender viewport. With Method B, download the .blend file and open it in Blender, or ask Claude to render a viewport screenshot. Check dimensions, proportions, and detail fidelity. Common issues on the first iteration and how to direct Claude to fix them:

• Text looks backwards on one side → specify which face (+Y or -Y) needs mirroring
• Model too small or too large → give the target dimension in mm
• Engravings disappeared on curved surface → Claude needs to query the actual surface position at each engraving location
• Boolean operations destroyed the mesh → tell Claude to use individual cutters per stroke, not joined geometry
• Features on wrong axis → always communicate using axis references (+X, -Y, +Z), never 'front' or 'back' which are ambiguous in 3D

Each fix is a conversation turn: describe the problem, Claude adjusts the script, re-executes, and you review again. Two to five iterations typically produces a finished model.

Rotate the model to check every face. Verify the rune engravings read correctly on both flat sides — text on the +Y face should mirror naturally when viewed from the +Y direction, and text on the -Y face should read correctly from the -Y direction. Check the eye hole is clean and centered. Confirm the cross-peen tapers smoothly without jagged edges.

In Blender, press Numpad 1 for front (-Y view), Ctrl+Numpad 1 for back (+Y view), Numpad 3 for right (+X view), and Numpad 7 for top (+Z view). These orthographic views reveal alignment issues that perspective view hides.

Save the model as a .blend file (File → Save As, or via script: bpy.ops.wm.save_as_mainfile(filepath='model.blend')). The .blend format is Blender's native format and the best choice for archival and sharing:

• Smallest file size — approximately 50% smaller than the equivalent STL (536 KB vs 1.2 MB for the hammer head)
• Preserves all modifier stacks, materials, scene settings, and edit history
• Fully editable — another maker can open it, change the rune text, resize for a different weight, or adjust any parameter
• Can be converted to any other format (STL, OBJ, glTF, 3MF) at any time via Blender or the Blender API

The .blend file is the single source of truth. All other formats are derived from it.

Export the model as STL for compatibility with 3D printers and slicer software (File → Export → STL, or via script: bpy.ops.wm.stl_export(filepath='model.stl')). STL is the universal format that every slicer (PrusaSlicer, Cura, BambuStudio) and every 3D printer accepts.

The STL file contains only the triangle mesh — no materials, no modifiers, no edit history. It is larger than the .blend (the mesh data is stored as plain text coordinates) and is not easily editable. Think of it as the 'compiled' output: good for printing, but you go back to the .blend source to make changes.

For best results, ensure all modifiers are applied before export (in the script, this is automatic). Check 'Selection Only' if your scene has multiple objects and you only want the hammer head.

When creating or editing a blueprint on Youblob, add a design step (content type: 'design') and upload the .blend file as the primary design file. The .stl can be uploaded as a second design step or as a supplementary download. Youblob's blueprint viewer will display an interactive 3D preview for STL files and a download button for .blend files.

Include the model dimensions, weight, and material density in the step description so other makers know the exact specifications. For parametric models (like this weight-scaled hammer), document which variables to change for customization — other makers should be able to modify the design for their needs without starting from scratch.

Lessons learned from designing the Norse hammer head and other models with Claude:

• Use axis references, not directions: Say '+Y face' and '-Z axis', never 'front' or 'top' — these are ambiguous in 3D and lead to mirrored results.
• Boolean operations need individual cutters: Joining multiple primitive shapes before a boolean creates non-manifold geometry. Keep each cutter as a separate object.
• Query the actual surface: After subdivision and modifiers, the mesh surface is not where the original dimensions suggest. Claude should measure vertex positions before placing engravings or cuts.
• Scale by volume, not dimensions: To hit a target weight, compute mesh volume using the signed tetrahedra method, then apply a uniform scale factor: (target_vol / current_vol) ^ (1/3).
• Iterate in 2-5 rounds: The first script gets 80% right. Each review-and-fix cycle catches remaining issues. Five iterations typically produces a production-ready model.