anyCreature README

The project's README from github.com/Ariescar/anyCreature. As one markdown file: README.md.

anyCreature

Text → a game-ready 3D creature, in one session.

An AI session takes an order like "make me a menacing mountain giant", asks at most two questions, and delivers a skinned, animated, vertex-coloured, AO-baked GLB plus an offline showroom viewer.

License: MIT Version Engine Output Tooling

anyCreature — a winged creature compiled from a one-line order

Every creature is compiled from one JSON spec. No mesh files, no downloaded art packs, no photogrammetry. The worked example that ships with this repo, example/wolf.json, is 2,211 vertices and 31 joints written out by the engine from plain text.

Quick start

bash setup.sh                              # deps + red/green ruler calibration
node engine/cli.js example/wolf.json out/wolf.glb

setup.sh must print calibrate OK before you trust anything else: it builds one spec that has to pass and two that have to be blocked, so you know the rulers separate good from bad on your machine.

To run the full text-to-creature session, give your agent (Claude Code, Cursor, a VM agent — anything that runs shell commands) EXACTLY this:

Read MANUAL.md in this folder and follow it exactly — the cards in cards/ are the pipeline, do not improvise around them. First run bash setup.sh and confirm it prints "calibrate OK". Then make me: <describe your creature>. Real-world height: <e.g. 1.8 m>. Ask me exactly one question — card 01's — and nothing else. Not cost, not token budget, not how long, not whether to publish. Decide the rest and tell me what you chose when you deliver.

Three things that make or break the result:

Driving it harder

Knob Where What it changes
smooth_angle spec root, or per volume/part Crease threshold in degrees, default 50. Higher = smoother body, lower = more facets.
shading spec root Whole-body vertical colour ramp and grain — {gradient:{top,bottom}, noise:{size,amount}}.
build: "rigid" spec root The only way to get a fully faceted body past the checker. For robots and crystals, not for animals.
ao spec root Per-vertex ambient occlusion; false to skip the bake.
embed_spec spec root Every GLB embeds its own authored spec (asset.extras.source_spec) + a parts manifest — any agent can extract, edit, recompile, or graft parts between creatures (harness/graft.py). false opts out.
harness/claims.json QC The one claims sheet — signature presence, 6:3:1 hierarchy, focal contrast, saturated area, rig/animations, triangle budget. Everything is judged at boss standard; adjust tri_budget for cheap creatures.

What is actually in here

Folder What it holds
engine/ The ACS engine. One JSON spec in, one skinned GLB out. Zero runtime dependencies — node engine/cli.js spec.json out.glb is the whole interface.
cards/ The five stage cards the executing session reads, plus the spec syntax on one page.
harness/ Measuring tools (silhouettes, mask metrics, claims judge), the delivery packer and the publisher.
example/ A bred, approved light quadruped to read and copy from.
calibration/ One sample that must build and two that must be blocked — proof the rulers separate good from bad.

How it works

flowchart LR
    A[Order<br/>one sentence] --> B[Interview<br/>≤ 2 questions] --> C[Silhouette<br/>brief] --> D[LOW<br/>design free]
    D --> E{Gate 1<br/>RECOGNISED}
    E -->|fail| D
    E -->|pass| F{Gate 2<br/>PUNCHIER}
    F -->|reverted| D
    F -->|pass| G[MID<br/>all parts] --> H[HIGH<br/>colour + anims] --> I[SHIP<br/>stamp · deliver]

The two gates

Neither gate is self-graded. The session that designed the creature is the worst possible judge of whether it reads, because it already knows what it drew. So the silhouettes go to a context-free reader agent that has never seen the order, and the only question is "what is this?"

Gate 1 is RECOGNISED: all four views have to land. Gate 2 is PUNCHIER: a new round may only make the silhouette bolder than the last one — a round that tames the shape is reverted, even if it is "more correct". LOW's deliverable is an exaggerated silhouette, not an accurate one.

The engine floors

The compiler talks in three registers, and they mean different things:

Prefix Meaning
BLOCK: The build stops. A hard floor was crossed — faceted body, part floating off its host, a mirrored twin distorted past 30%, an attack that never reaches.
warn: It built, but something is probably wrong. Read it.
info: A number you asked for, or an assumption the compiler made on your behalf — like which way an anchored fin ended up facing.

Floors are not style opinions. They are the failures that survive review and ship broken: a body that renders in hard facets, a horn that hovers a centimetre off the skull, a right thigh collapsed to 40% depth by mirrored skinning.

What gets measured

four silhouette views the blind reader is shown

Every round renders these four views, reduces them to masks, and computes the numbers the design card declared a target for — width over height, mass thirds, torso depth contrast, leg fraction, silhouette turn count, zigzag alignment, and IoU against the previous round as a regression guard. The 24px thumbnail is what the blind reader actually sees; if it does not read at 24px, it does not read.

The doctrine in one paragraph

Head-to-head experiments showed the model designs BOLDLY when left free, and every attempt to teach it design upfront made the output tamer — so this harness ships a clean painter and a strict inspector. The creation side gets only the order, the engine syntax, and a short pit-map of engine-local traps; ALL quality control lives in gates read by context-free reader agents, never self-graded. Form beats obedience, everywhere.

Honesty about limits

Scripts

Script Role
engine/cli.js Spec → GLB. The whole engine interface.
harness/silmetrics.mjs Renders the four silhouettes and the 24px thumbnail; emits the layout metrics.
harness/maskmetrics.py Per-view mask measures and dullness flags (sq_fill, mirror_sym, straight_max).
harness/judge.mjs Claims judge — part shares, focal contrast, saturated area, rig/anim/triangle budgets.
harness/hero.mjs hero.png (1024² transparent, 45°) and hero.jpg over studio grey.
harness/deliver.py Stamps identity into the GLB, writes the offline showroom viewer, builds the upload pack.
harness/publish.mjs Gobkit publisher. Only runs after an explicit yes, and always tries the upload before offering the manual page.
harness/calibrate.py The red/green self-check setup.sh runs.

Requirements

Node 18+ and Python 3.9+. Windows: run setup.ps1 in PowerShell (or setup.sh in Git Bash/WSL). setup.sh installs three, playwright, numpy, pillow and scipy, then runs the calibration self-check. The engine alone needs nothing but Node; the dependencies are for the measuring and render tools.

Output contract

Every GLB this harness writes satisfies docs/OUTPUT_CONTRACT.md — two chunks, no trailing bytes, no external URIs, no textures (colour is baked into vertex colours), semantic material names, convention bone names. Verify any file with node harness/glbcheck.mjs <file.glb>; the same module runs as a dependency-free import inside a server or Worker.

Security

Short-lived local servers used by the render tools bind to 127.0.0.1 only, and Chromium runs with its OS sandbox enabled. The Gobkit submission key in harness/gobkit.json is public by design — it authorises posting to the community wall and nothing else. Details and reporting: SECURITY.md.

Licence

MIT — see LICENSE. One third-party component is bundled (harness/assets/three-bundle.js, three.js, MIT); attributions for it and for everything setup.sh installs are in THIRD-PARTY-NOTICES.md.

About

Gobkit: AI-native 3D infrastructure for agents and vibe coders. Stream game-ready 3D assets straight into your workflows via API. Agent-friendly. Studio-quality. Built by a team of artists and engineers.

Author: Ariescar | Alsomindtech

👹 gobkit.com · 📖 Devlog

anyCreature (Ariescar/anyCreature)

procedural-generation
Searching Database…

467 ·GitHub
AnyCreature
What is this?anyCreature turns a plain-language description like 'a menacing mountain giant' into a game-ready 3D creature — a skinned, animated, vertex-coloured GLB file plus an offline viewer — produced in one AI agent session.
Is this a standalone app, or something my coding agent runs for me?It's a CLI pipeline (an engine plus a set of staged 'cards' and a harness) meant to be driven by an AI coding agent such as Claude Code or Cursor — there's no GUI, just `node engine/cli.js spec.json out.glb`.
How do I start generating a creature?Run `bash setup.sh` (or `setup.ps1` on Windows) until it prints 'calibrate OK', then point your agent at MANUAL.md and have it follow the card pipeline for the creature you describe.
What do I need installed before running it?Node 18+ and Python 3.9+. setup.sh additionally installs three.js tooling, playwright, numpy, pillow and scipy for rendering and measuring — the engine itself needs only Node.
Any known failure modes to watch for?Isolated-part blind reads can misread certain heads shown alone (elephants/birds), specs written before engine version 1.2.0 may now fail new checks, and running separate short-lived subagents for each pipeline stage is slower and costlier than one agent doing it start-to-finish.
Does it run on Windows, or just Mac/Linux?All three — Windows is supported via setup.ps1 in PowerShell (or setup.sh under Git Bash/WSL), alongside native macOS and Linux support.
Does any of this talk to a remote server?The render and measurement tools only spin up short-lived local servers bound to 127.0.0.1; the sole outward action is an optional, explicit step to publish a finished creature to the Gobkit community wall.
What is Gobkit?Gobkit is the author's AI-native 3D-asset infrastructure/API for streaming game-ready models to agents and 'vibe coders'. anyCreature is the open-source engine behind it, and finished creatures can optionally be published to its community wall.
What does the embedspec option actually give me?It makes the exported GLB carry its own authored JSON spec and a parts manifest inside the file, so any agent can later extract, edit, recompile, or graft parts between creatures.
Why two separate quality gates instead of one?The session that designed the creature is the worst judge of its own silhouette. Gate 1 (RECOGNISED) checks a context-free reader agent can identify the creature from four views; gate 2 (PUNCHIER) blocks any revision that makes the silhouette less bold than the previous round, even if it's more 'correct'.
What exactly does the bundled example wolf show?example/wolf.json is a 2,211-vertex, 31-joint quadruped used purely as a syntax reference — it ships only idle and move animations, not attack, and the engine actively blocks builds that are just this wolf with the numbers barely changed.
What are similar projects?img2threejs, procedural-film, genex-desktop, Procedura, mcp-for-blender, flutter_scene, blender-production-skills, image-to-3dlab
Which topics is it in?gltf, 3d-modeling, 3d-generation, procedural-generation, computer-graphics, game-development, 3d, cursor, animation, generative-ai, agent, claude-code
Creator
Ariescar
Size
1.1 MB
Stars
467
Updated
CommitSep 21Releasev1.3.2Sep 21
Age
2mo
Commits · 90 Days