Articraft is an agentic system for generating articulated 3D assets — objects with working joints, like a desk lamp whose arms hinge or a fan whose head tilts — by having a language model write the geometry as code. Released under Apache-2.0 in March 2026 by a team spanning Oxford and Cambridge, it has reached roughly 1,430 stars and 177 forks, is backed by an arXiv paper, and unusually for a project this size carries only five open issues. Its approach inverts the assumption behind most 3D generation work. ## Code, Not Meshes The dominant path to AI-generated 3D runs through learned representations: diffusion over triangle soups, Gaussian splats, signed distance fields, or multi-view images lifted into geometry. These produce visually convincing static objects and struggle badly with articulation, because a mesh generator has no concept of a hinge. Joints have to be inferred afterward by a separate rigging step, and inferred joints are frequently wrong. Articraft skips the learned 3D representation entirely. It asks a language model to emit a Python `model.py` that constructs the object programmatically, with joints declared explicitly as part of the code. The output is not a mesh that happens to look articulated — it is a parametric program whose articulation is stated by construction. Kinematic structure comes for free because the model wrote it down, and the asset is editable by editing code. The practical consequence: `uv run articraft generate "Create a realistic articulated desk lamp with a weighted base, two hinged arms, and an adjustable lamp head."` produces a record with working joint controls, viewable in the bundled React viewer. Editing works through forking — `articraft fork <record_id> "make the handle longer"` creates a child record and leaves the parent untouched, giving generation a version history rather than a regenerate-and-hope loop. ## Bring Your Own Model Articraft is model-agnostic by design and reads provider keys for OpenAI, Gemini, Anthropic, and DashScope from a local `.env`, defaulting to `gpt-5.6-sol` at high thinking level when nothing is configured. Cost is bounded per generation with `--max-cost-usd`, an acknowledgment that agentic 3D generation burns tokens in a way users need to control. The more interesting option is the no-API-key path. The documentation instructs users to point an external coding agent — Claude Code, Codex, or Cursor — at the repository and prompt it to create an articulated object following `EXTERNAL_AGENT_DATA.md`, with a dedicated Codex plugin available. Because the output format is Python source, any agent that can write code to a repository can drive the system. That is a natural fallout of the code-generation approach, and it is the kind of interoperability that purpose-built 3D pipelines cannot offer. Image-conditioned generation is documented separately for producing assets from a reference photo rather than a text description. ## Local-First and the Dataset Split The project restructured into a local-first harness: this repository holds generation and viewer logic, while the released dataset lives in the separate `mattzh72/articraft-data` repo under CC-BY 4.0. Records are stored in a gitignored data root that defaults to `<repo-root>/data`, redirectable via `ARTICRAFT_DATA_DIR` or `--data-dir`. That separation keeps the code repo light and makes it clear that contributed data is licensed for model training and public redistribution — the contribution terms say so explicitly, which is more transparency about data usage than most projects offer. A compact library CLI handles the housekeeping: listing records, rebuilding manifests, checking integrity, and assigning categories. Setup is standardized on `uv` for Python packaging and `just` as the command runner, with `just setup` and `just viewer` as the two commands most users will need. ## Pros and Cons The strengths follow from the core decision. Representing assets as code makes articulation explicit rather than inferred, and makes the output human-readable, diffable, and editable — three properties no mesh-generation pipeline provides. Provider-agnostic model support plus the external-agent path means the system is not tied to any vendor's continued availability or pricing. Fork-based editing gives a real provenance chain. Apache-2.0 on the code and CC-BY 4.0 on the dataset is a clean, permissive pairing. Peer-reviewable research backing with a published paper and project page separates it from the many demo-grade 3D repos. And five open issues against 1,400 stars suggests the maintainers are actually closing things. The limitations are significant and one is a genuine hazard. Articraft compiles and inspects generated records by executing their `model.py` files as Python, which the README flags in a security note — running records or model scripts from untrusted sources means running arbitrary code, and this constrains any workflow involving shared or community-contributed assets. Python 3.13 and later are unsupported, pinning users to 3.11 or 3.12. Meaningful use requires either paid API keys or a coding-agent subscription, so per-asset cost is real and variable, which is why the cost cap exists. The programmatic approach also implies a ceiling: code-generated geometry suits mechanical, structured objects — lamps, fans, furniture, tools — far better than organic or highly detailed forms, and the showcased examples reflect that. No tagged releases exist yet, so version pinning means pinning a commit. The toolchain requires `uv`, `just`, and optionally `npm`, which is more setup than a single pip install. And the separated dataset repo adds a step for anyone wanting to browse existing work. ## Outlook The broader pattern here — using an LLM to generate a program that produces an artifact, rather than generating the artifact directly — keeps proving useful across domains where the artifact has structure worth preserving. Articulated 3D is a strong fit because joints are semantic facts that survive far better in code than in geometry. Whether it scales past mechanical objects is the open question; organic articulation like a hand or a quadruped is a much harder programmatic target. The external-agent integration is a quiet bet worth noting: as coding agents become the default interface for technical work, systems that expose themselves as code repositories rather than APIs get compatibility with tools that do not exist yet. ## Conclusion Articraft is worth trying for robotics simulation, game asset pipelines, and 3D research needing articulated objects with correct kinematics — particularly for mechanical and household objects where programmatic construction plays to the method's strengths. The fork-based editing and code-level output make it genuinely useful for iterating on assets rather than rolling the dice repeatedly. Teams should account for the arbitrary-code-execution behavior before running any record they did not generate themselves, budget for per-asset LLM cost, and look elsewhere for organic or highly detailed geometry.