Printable

MCP server for AI-driven 3D modeling, optimized for FDM-printable geometry.

Connect any MCP-capable agent (Claude Code, Goose, Cursor, Codex, Cline, etc.) to a 3D modeling backend and get tools tuned for the design loop that actually produces parts you can print.

Why this exists

LLMs are good at writing geometry code but bad at the things that make a part actually printable: clearances on moving joints, overhangs, bridging, units, watertight meshes. Printable encodes those constraints into both the tool surface (typed booleans, printability checks, clearance sweeps, cross-sections) and the prose rules the agent reads (cardinal print-path rule, FDM clearance values, mechanism patterns).

The result: you can ask any MCP-aware agent "build me a hinge with 5mm barrel and 20×15mm flanges" and get a part that comes off the bed working.

Backends

• Blender (blender_* tools, 25 of them). Full design loop with rendering, cross-sections, printability validation. Requires Blender 3.6+ installed and the included addon enabled.
• OpenSCAD (scad_* tools, 5 of them, see docs/openscad/README.md). Code-first parametric backend. No app, no addon — shells out to the openscad CLI and uses trimesh for mesh validation. Requires OpenSCAD installed.

Cross-backend handoff happens via STL — both backends import and export it.

Architecture

Agent  <--stdio/MCP-->  server.py  <--TCP :9876-->  Blender addon
                              \
                               +----shell-out------>  openscad CLI + trimesh

• server.py — FastMCP server. Exposes the tool surface and embeds the always-on rules in the MCP instructions field, with pointers into docs/ for deeper guidance.
• addon/ — Blender addon. TCP server on 127.0.0.1:9876. Commands run on Blender's main thread via bpy.app.timers.
• docs/ — agent-agnostic prose guidance: print-in-place rules, design loop, image displacement. Also exposed as MCP resources under printable://… URIs (see Documents below) so any resource-aware MCP client can pull them via the protocol — no filesystem access required.
• .claude/skills/ — thin Claude shims (description-triggered loading) that point into docs/. Other agents use resources or filesystem.
• evals/ — policy-based regression tests that verify agents actually follow the rules. See evals/README.md.

Status

v0.1.x — alpha. Blender backend is feature-complete and dogfooded against real prints; OpenSCAD backend covers the parametric workflow but has fewer validation tools. API is stable enough to use but may shift before 1.0.

Requirements

• Python 3.10+ for the MCP server.
• Blender 3.6+ if you're using the Blender backend (the bundled addon needs to be installed once and enabled in Blender's Preferences).
• OpenSCAD CLI if you're using the SCAD backend (auto-discovered on PATH and in the standard install locations on Windows / macOS / Linux).
• An image-capable agent model. The visual-feedback tools (blender_get_screenshot, blender_render_tiled, blender_render_turntable, blender_cross_section*, blender_render_printability_heatmap, blender_render_with_dimensions, blender_render_before_after, scad_render_views, scad_cross_section) return base64-encoded PNGs that the agent has to actually see to use them. Text-only models will still get tool results but can't interpret the rendered geometry — the design loop relies on the agent looking at renders and cross-sections to verify what it's built. Examples that work well: Claude Sonnet 4.x+, GPT-4o/5, Gemini 2.x Pro/Flash. Text-only models will technically run but won't catch shape-level mistakes.

Setup

Install from PyPI

pip install mcp-printable       # or: uv pip install mcp-printable
printable-install-addon         # copies the bundled addon into Blender's addon dir

Then in Blender: Preferences → Add-ons → enable "Printable Blender Bridge".

Install from source

git clone https://github.com/AaronGoldsmith/mcp-printable
cd mcp-printable
uv sync                         # or: pip install .
python install.py               # equivalent to printable-install-addon

Wire into your agent

For Claude Code (~/.claude.json or project .mcp.json), pick one of two options.

Option A — uvx (recommended; no prior install needed):

{
  "mcpServers": {
    "printable": {
      "type": "stdio",
      "command": "uvx",
      "args": ["--from", "mcp-printable", "printable"]
    }
  }
}

uvx will pull mcp-printable from PyPI on first run and cache it. To pick up new releases later, run uvx --refresh --from mcp-printable printable once or pin a version with mcp-printable@0.1.2.

Option B — global pip install (simpler if you already have mcp-printable on PATH):

{
  "mcpServers": {
    "printable": {
      "command": "printable"
    }
  }
}

Requires pip install mcp-printable to have put printable on your PATH first.

For other agents (Goose, Cursor, etc.) — same command, wrapped in your agent's MCP server configuration.

Agent skills (optional)

This repo bundles four Claude Code skills under .claude/skills/ — short shims that point at the same docs/ content the MCP exposes as resources. Claude Code auto-discovers them when you launch it in the repo:

git clone https://github.com/AaronGoldsmith/mcp-printable
cd mcp-printable && claude

The 4 skills:

• print-in-place — design rules for moving-parts mechanisms (hinges, ball-sockets, snap fits)
• blender-design-loop — plan→build→verify→validate→export workflow
• image-displacement — turn a 2D image into 3D printable relief
• blender-app — launch / restart / multi-instance Blender setup

For Claude Code, copy a skill into ~/.claude/skills/ to make it available across all projects. For other agents (Codex, Cursor, Goose, ...) — see AGENTS.md, which links each skill into the equivalent location for your agent and explains the MCP-resource fallback for agents that don't load project skills.

Tool families

Blender (23 tools)

Scene — blender_get_scene_info, blender_get_object_info, blender_clear_scene, blender_restore_checkpoint (roll back to the auto-saved checkpoint after a destructive mistake), blender_rename_object

Code — blender_execute_code (arbitrary bpy/bmesh; auto-checkpoints), blender_boolean (typed UNION/DIFFERENCE/INTERSECT — prefer this over execute_code)

Visual feedback — blender_get_screenshot, blender_render_tiled, blender_render_turntable, blender_cross_section, blender_cross_section_gallery, blender_render_printability_heatmap, blender_render_with_dimensions, blender_render_before_after

Print validation — blender_validate (one tool for HEALTH / OVERHANGS / THIN_WALLS / CLEARANCE checks; checks=['ALL'] runs the full suite — replaces the former mesh_health / check_overhangs / check_thin_walls / full_printability_check tools), blender_check_clearance, blender_check_clearance_sweep, blender_check_intersection, blender_check_retention

Export — blender_export_stl, blender_import_stl, blender_save_blend

OpenSCAD (5 tools)

scad_compile, scad_render_views, scad_cross_section, scad_validate_printability, scad_import_stl. Shells out to the openscad CLI; uses trimesh for mesh validation. Full docs: docs/openscad/README.md.

The Design Loop

Always-on rules embedded in the MCP server's instructions field — every agent that connects sees them automatically. Summary:

1. Plan. Compute coordinates and dimensions in one execute_code call that PRINTS them. Verify the math BEFORE creating geometry.
2. Build. 1–3 operations per execute_code, then blender_validate(checks=['HEALTH']).
3. Verify. Renders for shape, cross-sections for internal truth.
4. Validate. blender_check_clearance_sweep for any joint. blender_validate(checks=['ALL']) before export.
5. Export. blender_export_stl (no args = bundle all parts).

Full doc: docs/blender/design-loop.md.

For mechanism design (hinges, ball-sockets, snap fits, articulated chains): docs/design/print-in-place.md. This is backend-agnostic — same rules apply if you're using OpenSCAD.

Documents

Every doc below is served two ways:

1. As an MCP resource under printable://… — the preferred path. Resource-aware clients fetch via resources/read, get the same content the maintainer ships, and don't need filesystem access to the project. Resources travel with the MCP server itself, so a pip install mcp-printable user has the docs even without cloning the repo.
2. As a file in docs/ — fallback for filesystem-based clients, and for humans browsing the repo.

If you're writing an MCP client, prefer the URI. The filesystem path is documented mainly so a human can click through from this README.

docs/
├── design/                          # backend-agnostic design rules
│   └── print-in-place.md
├── blender/                         # Blender-specific
│   ├── design-loop.md
│   ├── image-displacement.md
│   └── blender-app.md
└── openscad/                        # OpenSCAD-specific
    └── README.md

Testing

python -m pytest tests/ -v          # unit tests, no Blender required
python evals/runner.py              # policy-based regression evals (see evals/README.md)

Unit tests cover TCP protocol framing, image compositing (PIL), MCP tool registration. The eval suite checks that agents using the MCP actually follow the always-on rules (e.g. clear scene first, prefer blender_boolean, no monolithic execute_code blocks) — procedural checks from the tool trace, plus LLM-judged outcome policies for things like "moving parts have a continuous print path to the bed."

Roadmap

Things on the list, not yet shipped:

• Validated parameter recipes — turnkey parameter sets for common print-in-place mechanisms (wheel-on-axle, flip-tile, ball-and-socket, snap fit) so an agent can ask for "a toy car wheel" and get a known-good geometry without re-solving the clearance + retention math each time. Earlier drafts existed but weren't dialed in enough to ship as authoritative; new ones will land once they're validated against real prints.
• Blender app-lifecycle tools (blender_launch / blender_status / blender_kill) — let the agent spin Blender up itself instead of needing the user to start it first.
• Second OpenSCAD parity pass — match Blender's clearance-sweep / retention / thin-wall checks on the SCAD side.

Contributing

The interesting design surface is in docs/. New always-on rules belong in docs/design/ (backend-agnostic) or docs/blender/ / docs/openscad/ (backend-specific). If a rule should be enforced, also add a policy file under evals/policies/ and a scenario under evals/scenarios/ so the eval runner picks it up.

License

MIT — see LICENSE.