3D Studio — natural language → print-ready 3D
A Claude Code plugin + agent harness that turns a natural-language
description (and optional reference images) into print-ready 3D files
(STL · 3MF · GLB) you can drop straight into Bambu Studio, OrcaSlicer, PrusaSlicer,
or MakerWorld — validated against the 2026 AI 3D Print-Readiness Benchmark, with a
React + three.js preview UI.
The agent does the modeling. The primary engine is a local, manifold-by-
construction CSG kernel that the agent writes geometry against — deterministic,
free, needs no GPU or API key, and produces watertight solids by construction
(~100% slicer pass). An optional generative backend (Meshy / local models) handles
genuinely organic shapes, always followed by validate → repair.
New in v0.4 — the print-readiness moat (beat Meshy where it's structurally weak)
3D Studio competes with generative-mesh tools (Meshy/Tripo) not on organic mesh
aesthetics — their saturated home turf — but on the properties a diffusion-mesh cloud
tool is architecturally barred from matching. See docs/V3-COMPETE-WITH-MESHY.md.
- Real slice-to-G-code D2 (verified).
studio3d slice / validate --slice drives a
detected OrcaSlicer/Bambu/PrusaSlicer/Cura headless to produce real G-code with real
print time + filament grams — verified end-to-end against Bambu Studio 2.7.1 (e.g.
"25m 43s, 3.5 g"). With no slicer it falls back to an explicitly labeled proxy
(d2_method), so "print-ready" is never silently self-certified.
- Packaged cross-OS slicer setup.
studio3d slicer status detects the OS/arch and any
installed slicer (native / AppImage / .app / flatpak); studio3d slicer install gives the
exact per-OS recipe (OrcaSlicer AppImage on Linux, brew --cask on macOS, winget on
Windows) and performs it with --yes.
- Forever-editable parametric source. Every bundle ships
model.py + params.json.
Change one knob (studio3d tweak --set wall=3) and regenerate deterministically. A
Meshy mesh has no knobs to turn.
- Print-Readiness Certificate.
certificate.json — prompt → script/file hashes →
D1–D4 → slice result → human-approval slot — an auditable provenance chain.
- Kernel-metrics-fused critique. The
design-critic runs on a stronger model and
reads genus / component-count / wall-p05 / overhang alongside the renders (catches
defects invisible in a 4-view image), with an anti-stagnation escalate verdict.
- Heal-the-generative-path. Any generative mesh is forced through repair + a
manifold round-trip toward a watertight solid before validation.
- Per-part AMS multicolor from one CSG union.
paint() each part and combine with
multicolor_union(...). The exporter writes the Bambu/Orca native format — separate
objects with per-object extruder (filament) assignment in model_settings.config plus
an N-filament project_settings.config built from Bambu's own template — not core-3MF
colorgroups (which Bambu ignored: "no filament colors found"). The per-object filament
routing is verified to survive a Bambu Studio round-trip (objects → deduped slots
1,2,3…). Open in Bambu/Orca with a printer selected to map parts to AMS slots; identical
part colors dedupe to one slot. (Headless GUI color display isn't scriptable here, so
that last visual step is confirmed by opening the file, not in CI.)
- Faceted STEP export (
studio3d step, or --formats …,step) — a dependency-free
AP214 MANIFOLD_SOLID_BREP (no OCP kernel needed on CPython 3.14). The output is
structurally verified (sewn closed shell, consistent orientation, coplanar faces);
acceptance by a specific CAD kernel is best-effort and not yet round-trip-confirmed in
Fusion/SolidWorks here.
- Assemblies. Multi-part design plans with
mates, a single clearance_mm knob,
and an automated interference() collision check.
- DFAM/CSG domain knowledge base (
studio3d kb) grounds authoring in documented rules.
- Disconnected-part gate.
print_ready fails when a model that should be one solid
comes out in detached/levitating pieces (n_components > expected) — a watertight-per-
piece model is not print-ready if it's in fragments.
- Internal fabrication-reliability self-check (
studio3d muse). This is NOT the MUSE
benchmark and its number is not comparable to it: it runs bundled reference scripts
(the answer key) through fabricate→validate and reports how reliably the deterministic
kernel + validator produce print-passing solids. It measures pipeline reliability, not
text→CAD generation quality (the hard part — turning a prompt into geometry — is done by
the agent, and a blind held-out eval is the honest measure of that).
- Browser Customizer. The viewer surfaces the parametric knobs, the certificate, the
