Skip to main content
Glama

Clayform

A 3D workshop built for AI agents. Agents model, sculpt, rig, animate and bake effects by editing a semantic scene document. Clayform meshes it, renders it for them to look at, measures what a picture can't show, and exports game-ready glTF. No Blender, no GPU, no browser: it is a Node library with an MCP server.

Docs: clayform.sadelabs.site · Made with it: the showcase

The 47 templates, rendered by Clayform's own software renderer

Every image in this README is a real Clayform render, regenerated by scripts/.

Why

Connecting a language model to a traditional 3D package works, but the model is working blind with a low-level API. It writes vertex and coordinate code, has trouble seeing what it made, and can't tell a floating ear from an attached one. Clayform changes the medium instead:

Problem for agents

Clayform's answer

Coordinates are where models make mistakes

Parts attach to each other's surface by side ("on head, top"), sink in by embed, and mirror themselves. Nothing can float by accident.

No sense of what is what

Every part has an id and a role (head, leg, wheel …). The parts view paints each one a color and prints a legend.

Can't see the result

A built-in software renderer returns labelled multi-view sheets (front, side, top, three-quarter) with ground and shadows as MCP images.

Can't judge the result from a picture

Critics measure: floating or split pieces, carves that cut a part in two, buried parts, detail too thin for the resolution, broken symmetry, models that would tip over, triangle budgets, and limbs that will stretch when animated. Every issue names the parts and says what to change.

Sculpting and keyframing are hand gestures

Anchored sculpts (inflate the cheek, radius 6 cm) and intent clips (walk, fly, drive, wave …) on an automatic rig.

Output isn't game-ready

glTF 2.0 with materials, vertex colors, skin and animations, reduced to a triangle budget. Validated with the Khronos glTF validator in the test suite.

Related MCP server: modelgen

Quick start

Requires Node 20+. Install from npm:

claude mcp add clayform -- npx -y @s1444/clayform mcp

For any other MCP client, the command is npx -y @s1444/clayform mcp. You can also run directly from GitHub: npx -y github:sezginkipel/clayform mcp. To host it for several agents, clayform serve speaks MCP over HTTP with a sandboxed workspace per session (hosting). Scenes are saved in ./.clayform/ (change it with --workspace <dir> or the CLAYFORM_WORKSPACE environment variable).

To install a fixed version, use the prebuilt tarball from Releases: npm install https://github.com/sezginkipel/clayform/releases/download/v0.1.0/clayform-0.1.0.tgz. To work on the code: git clone, then npm install and npm test. Then ask your agent for something: "Make a goblin from the biped template, green skin, pointy ears, and export it with a walk cycle."

What an agent does

// new_scene { "name": "Goblin", "template": "biped" }
// edit — one atomic batch:
[
  { "op": "set_palette", "set": { "skin": "#7fb04a", "shirt": "#6b4a2b", "pants": "#3d3322" } },
  { "op": "update_part", "id": "hair", "set": { "hidden": true } },
  { "op": "add_part", "part": {
      "id": "ear", "role": "ear", "shape": { "type": "cone", "height": 0.16, "radius": 0.045 },
      "attach": { "to": "head", "side": "left", "offset": [0, 0.2], "align": true, "embed": 0.2 },
      "rotation": [0, 0, -25], "mirror": true, "blend": 0.02, "material": { "color": "skin" } } },
  { "op": "update_part", "id": "nose", "set": {
      "shape": { "type": "cone", "height": 0.09, "radius": 0.03 },
      "attach": { "to": "head", "side": "front", "offset": [0, -0.1], "align": true, "embed": 0.2 } } }
]

render

render with mode: "parts"

Goblin, shaded

Goblin, one color per part with a legend

preview_motion { "clip": "walk" }:

Six frames of the goblin's walk cycle

preview_effect on the torch template's fire preset:

Eight frames of a baked fire flipbook

Critic output looks like this (from a deliberately broken edit):

size 1.00 × 1.32 × 0.46 m · 31,412 triangles · 2 body islands · built in 226 ms
ERROR [floating] orb is not connected to the rest of the body — use attach, raise embed, or mark it separate if it should be its own mesh
WARN [asymmetric] declared X symmetry is off by 1.3 cm on average; worst: orb (43.1 cm) — use mirror: true instead of hand-placing both sides

MCP tools

Tool

What it does

guide

The manual an agent reads once (topic schema returns the full JSON Schema)

list_templates

19 tuned starting points: biped, quadruped, bird, fish, slime, robot, snowman, car, spaceship, cottage, tree, rock, mushroom, crystal, sword, chest, barrel, potion, torch

list_parts

A library of reusable parts (eyes, ears, horns, tails, wings, limbs, wheels, windows, doors, hats), added with one edit op

new_scene · list_scenes · get_scene · import_scene

Scenes in the workspace; get_scene summarizes parts with resolved world positions

edit

Atomic batch of ops (add/update/remove/rename/duplicate parts, sculpts, clips, effects, settings, palette). Returns what changed and the critics

render

Labelled multi-view PNG; modes shaded, parts, clay, normals, depth; compare: "previous" shows before and after the last edit

inspect

Part summary, critics, and a check of every clip (ground contact, parts passing through each other)

measure

Exact distances, overlaps, sizes, proportions, and which part is at a pixel

compare_reference

Fit the model to a sketch or photo: a score, an overlay, and which parts are too wide or narrow

preview_motion

A clip as a labelled film strip

beauty

A presentation render: lighting presets, soft shadows, a transparent background, or a turntable as an animated PNG

preview_effect

An effect's frames

export

glb · obj · json · flipbook (sprite sheet PNG + JSON), with an optional triangle budget

set_layout · render_layout · export_layout

Place many scenes or templates together (rows, grids, circles, scatters), check that they do not pass through each other, export one GLB

history

Undo, redo, named snapshots

The scene document

Meters, +Y up, the model faces +Z, the model's left is +X. The full reference is in src/guide.ts (the text the guide tool returns).

  • Shapes: sphere, ellipsoid, box, capsule, cylinder, cone (both optionally faceted with sides), torus, prism, tube (a swept tube with per-point radius, for tails, limbs, horns), and mesh (an imported GLB or OBJ).

  • Combining: add with a smooth blend radius, carve, intersect; pattern (spots, stripes, noise, gradient), detail (surface displacement).

  • Sculpts: inflate, dent, flatten, crease, noise, anchored to a part and side or a point.

  • Clips: idle, walk, run, hop, fly, swim, drive, spin, hover, wave, nod, or keyframes, with keyframe tracks layered on top.

  • Effects: fire, smoke, sparks, magic, explosion, dust, snow, rain, bubbles, heal. Every parameter can be overridden, and the particle paths are closed-form, so loops are exactly periodic.

CLI

clayform templates
clayform render biped --mode parts -o biped.png
clayform inspect my-scene.clay.json      # exit code 2 when critics report errors
clayform export my-scene.clay.json -o hero.glb --triangles 3000
clayform motion quadruped walk -o walk.png
clayform effect torch -o fire.png        # sprite sheet + .json + .preview.png

As a library

import { getTemplate, buildScene, critique, renderSheet, simplifyBuild, exportGlb } from 'clayform';

const build = buildScene(getTemplate('robot')!.scene);
console.log(critique(build).issues);
const png = renderSheet(build, { mode: 'parts' }).png;
const glb = exportGlb(await simplifyBuild(build, { triangles: 4000 })).glb;

How it works

  1. Resolve parts in dependency order. Attached parts are placed by sphere-tracing the target's own distance field from the requested world side (without an offset, the extreme point that way), then sunk in by embed. Mirror twins are exact reflections.

  2. Mesh the signed distance field (smooth-min unions, carves, intersections, sculpt modifiers) with surface nets on a block-culled grid. Only the shell near the surface is sampled.

  3. Attribute every vertex: normal from the field gradient, ambient occlusion from the field, blended part color and pattern, dominant part, and skin weights restricted to the dominant part's parent and children.

  4. See: a deferred software rasterizer (triangle ids + barycentrics, then one shading pass) draws the mesh, an analytic ground plane, a projected key shadow and a soft contact shadow. Labels use a built-in 5×7 bitmap font.

  5. Measure: connected components, the volume centroid against the ground-contact hull, mirrored field sampling, and per-part visibility and thickness.

  6. Export: quadric simplification (meshoptimizer) that keeps color, normal and material borders, then GLB with one primitive per material, a joint per part, and sampled animations.

Documentation

The same pages, with every image, are at clayform.sadelabs.site.

Showcase

Projects made with Clayform are listed at clayform.sadelabs.site/showcase. The first is Cargo Century, a browser transport tycoon whose 118 buildings, industries, stations, trees and vehicles are all Clayform scenes.

Cargo Century's Clayform models

Made something with it? Send it in, or open a pull request with a small JSON file (the rules).

Limits (today)

  • The look is stylized: smooth and clay-like, colored per vertex or in a baked texture. Material presets add normal and roughness maps, but nothing is photoreal. Hard mechanical edges are rounded at the cell size unless edges is sharp, so raise resolution for crisp props.

  • It does not generate organic detail from a text or image model yet. You can import a mesh made elsewhere (mesh shape, GLB or OBJ) and keep working on it. Built-in hand-off to an image-to-3D model is on the roadmap.

  • Animation is procedural plus keyframes on rigid parts: feet and arms use IK, characters can export humanoid bone names for retargeting, faces are morph targets and cloth ripples as a baked wave. Limbs bend only where parts meet. Layouts settle items straight down, but there is no rigid-body or cloth simulation: nothing tips, slides or collides while it moves. Motion critics check ground contact, sliding feet and parts passing through each other.

  • The renderer is software rasterization with a shadow map. beauty makes clean pictures for a README or a store page, but there is no global illumination, glass or subsurface light.

  • Nobody has run the bench yet (bench/), so this README makes no quality comparison.

Roadmap

Next up: image-to-3D hand-off (v0.4) and a published bench (v0.9). See ROADMAP.md, milestones and the changelog.

Development

npm test                              # includes the Khronos glTF validator and runs every docs example
npm run check                         # types
npx tsx scripts/gallery.ts            # regenerate docs/gallery.png
npx tsx scripts/docs-images.ts        # regenerate the README images
npx tsx scripts/gen-docs.ts           # regenerate docs/reference.md, docs/templates.md, the JSON Schema
npx tsx scripts/bench.ts <dir>        # score a bench run

See CONTRIBUTING.md.

License

Apache-2.0 © Sezgin Kipel

Related MCP Connectors

Related MCP Servers

  • A
    license
    Not graded
    quality
    A
    maintenance
    Enables AI agents to create, edit, and export 3D models (GLB, USDZ) from natural-language descriptions and YAML, with rendering and problem detection.
    6 npm
    MIT
  • A
    license
    A
    quality
    A
    maintenance
    Enables an AI agent to drive a live Blender session, building, sculpting, measuring, checking, rigging, lighting, rendering and exporting 3D models through safe operations in world units. Every geometry-changing call reports achieved dimensions, open edges and failures as numbers and review images, so modelling mistakes become visible instead of silent.
    113
    MIT
  • A
    license
    Not graded
    quality
    C
    maintenance
    Enables agents and Blender users to generate game-ready art through a set of Blender tools covering concept image generation, image-to-3D mesh creation, mesh cleanup, PBR material assignment, greybox level layout, camera pathing and still or sequence rendering. Every action runs headless Blender, returns machine-readable JSON with pass/fail self-checks such as triangle budget, scale, pivot and clearance, so agents can verify and fix each step unattended.
    MIT