01
Natural language + trusted templates
Every prompt grounds on a maintained template family (boxes, brackets, jigs, panels, plates, fasteners…) before parametric extension. Safer than open-ended generation, faster than a GUI.
Meshwish is an AI-native parametric CAD builder for 3D printing. From a sentence to a 3MF — geometry, parameters, validation and AI repair, all closed-loop. No more prompt-to-mesh black box.
prompt
A wall-mountable router bracket, 250mm wide, 60mm deep, with keyhole hangers on the back and ventilation slots on the front.
Every model ships with editable parameters
Designed for
01
Every prompt grounds on a maintained template family (boxes, brackets, jigs, panels, plates, fasteners…) before parametric extension. Safer than open-ended generation, faster than a GUI.
02
A structured AST is the single source of truth (model.json + model.cad). AI patches it locally; the system enforces types, units, ranges; parameters become UI controls automatically.
03
Phase 1 prioritizes watertight, robust booleans and fast preview. BREP / STEP via OpenCascade and build123d come in Phase 2 when you actually need them.
04
Built-in slicer preflight: bed fit, volume, wall thickness, overhangs, self-intersections. Models that fail to print do not reach your download folder.
05
Compile errors, non-manifold meshes and parameter sweep crashes become AI-readable signals that feed back into the generate → validate → repair loop.
06
Source is a text AST. Diff it, share it, round-trip to OpenSCAD / KCL / build123d-style code. No platform lock-in.
Not just text-to-CAD: a full generate → compile → render → sweep → preflight → repair → export loop. Failed models never reach your downloads.
Pull intent, dimensions, materials and constraints out of natural language.
Match a template family or synthesize a structured AST. Not arbitrary scripts.
Manifold / CSG kernel produces a watertight mesh.
Front / top / side / iso / section / print-bed previews.
Sweep allowed ranges to catch geometry collapse before the user does.
Bed fit, wall thickness, overhangs, volume and filament estimate.
Bundle with parameter presets, print hints and rendered cards.
Users see a clean DSL, the AI patches a structured AST, and the system derives the parameter panel, type/unit checks and sweeps from it. Three views of one source.
router-mount.cad
v0.1 · AST// auto-synced with model.json
model "router-mount" {
params {
width: 250 mm [120..320]
depth: 60 mm [40..100]
wall: 2.4 mm [1.6..4.0]
hanger_pitch: 180 mm [60..240]
vent_slots: 6 [2..12]
}
body = rounded_box(width, depth, 90 mm, r=4 mm)
.shell(wall)
.vent_slots(count=vent_slots, axis="x")
back = body.face("-y")
.keyhole_pair(pitch=hanger_pitch)
export { "3mf", "stl" }
} Auto-generated parameter panel
MVP ships 20–50 high-quality parametric templates, each with print validation, presets and tests.
Size, wall thickness, grids, hinges, magnet pockets, common stacking systems.
Keyholes, magnets, VESA, 3M-pad pockets, cable channels.
Parametric workpieces, clamping hints, stackable plates.
Library of common cutouts (USB / DB9 / RJ45 / SMA…).
CJK-aware fonts, embossed / engraved, mounting variants.
Involute gears, knurled knobs, snap-fits, axle joiners.
Five templates end-to-end: DSL → Manifold/CSG → STL/3MF → multi-view render → parameter sweep. Each compiles in under 1–3 seconds.
20–50 high-quality templates + prompt-to-template matching + AI repair loop + browser builder (preview / code / params) + print preflight.
OpenCascade / build123d backend: STEP export, fillets / chamfers, sketch constraints. Continue editing in FreeCAD / Fusion / Onshape.
Template marketplace, enterprise libraries, versioning / audit, cloud slicing, API / Agent SDK, organizational permissions.
Those are "write code, get geometry." Meshwish is "prompt + AST + closed-loop validation": code, parameter panel and natural language are three views of one source. AI patches a controlled DSL while mesh checks, parameter sweeps and slicer preflight close the loop.
OCCT is powerful but heavy: topology naming, boolean robustness and WASM size all slow product validation. 3D printing does not need STEP. We get mesh CSG / 3MF / preflight right first; build123d / OpenCascade ship in Phase 2 for true engineering interop.
We do not rely on raw text-to-mesh models. The AI outputs a structured AST that must pass non-empty, watertight, parameter-sweep and slicer-preflight checks — otherwise it goes back into the repair loop. Failed models never reach your download folder.
You do. The source is a text AST, diffable and round-trippable to OpenSCAD / KCL / build123d-style code. No hard platform lock-in.
We are in the Phase 0 technical spike: validating 5 templates end-to-end through DSL → mesh → 3MF → printer. Join the waitlist for first-batch invites.
Waitlist members get priority invites at Phase 1 MVP launch, plus a free print-preflight quota across 20+ templates.