three.ws 3D for VS Code

Preview, animate, refine, optimize, and review 3D models where you already are. This extension makes .glb and .gltf files first-class in VS Code: double-click one and it renders, animates, and reports on itself in an editor tab, next to the code that ships it. Try any of 2,800 library animations on its rig and bake the one you like into the file. Describe a change and get a new version. Optimize a copy for the web without uploading it. Diff it against the committed version before you open the pull request. And when you embed it, the editor checks the <agent-3d> tag the way it checks your TypeScript.

Nothing here is a preview stub. The viewer is three.js with the same lighting, decoders and retargeter the three.ws site uses, the report and the diff are the platform's own inspectors, and generation calls the live studio API.

• A real viewer, as the default editor for model files. Orbit, pan, and zoom with the mouse. Animation clips get a picker, a play/pause button, and a scrub bar. Toggle a wireframe, the skeleton, the ground grid, or a slow auto-orbit. Draco, KTX2/Basis, and meshopt assets decode locally from decoders shipped inside the extension, so a compressed three.ws avatar renders offline.
• Model report. Triangles, vertices, meshes, materials, texture count and weight, animations, rig size, and glTF extensions, plus optimization notes. It is the same inspector that powers the three.ws /validation page, running in the extension host: your model is never uploaded anywhere to be measured. The status bar shows triangles, size, bones and clips while a model is open.
• Snapshots and turntables. Snapshot writes the current view as a PNG beside the model. Turntable writes a looping GIF of a full orbit, with the playing animation stepped in sync, for a README or a pull request.

• Try a library animation. Animate in the toolbar (or 3D: Try a Library Animation) searches the three.ws motion library, 2,800 clips baked on the canonical humanoid skeleton, and plays your pick on the open model. The clip is retargeted onto the model's own bones in the viewer by the platform's retargeter, which knows Mixamo, Avaturn, VRM/VRoid, Unreal, Daz, MakeHuman, Blender .L/.R and simple shoulderL conventions. If a rig cannot be driven, the message says how many bones matched and offers to rig it.
• Bake it in. Bake clip writes the retargeted clip into a copy of the file as real glTF animation samplers and channels, so it plays in Unity, Unreal, Blender, <model-viewer>, or <agent-3d>. Meshopt-compressed models are re-encoded, not inflated.
• Animate from words. 3D: Animate this Model from a Text Prompt sends "waving confidently with the right hand" to the text-to-motion worker on the three.ws GPU fleet and retargets the sampled clip the same way.

• Text to 3D. 3D: Generate a Model from Text for an object or prop, 3D: Generate an Avatar from Text for a character. Both run on the free lane (no key, no account), save the GLB into your workspace, and open it.
• Refine by describing a change. Refine (or 3D: Refine this Model) takes "make it metallic" or "bigger helmet" and generates a new version anchored to the current one, carrying its form and materials forward. Versions save as name-v2.glb, name-v3.glb, and the lineage is remembered, so refining a refinement extends the same history.
• Auto-rigging. 3D: Rig a Model for Animation sends a static model to the three.ws rigger and brings back one with a humanoid skeleton and skin weights.
• AI quality check. Check quality renders the current view and asks a vision model to score realism and completeness against the subject, list the defects it can see, and suggest a fix. One click feeds that fix into a refinement. Because the render is what travels, it works on a model that has never left your disk.

• Optimize for the web, locally. 3D: Optimize a Copy for the Web runs the asset pipeline's passes in the extension host (dedup, prune, weld, resample, meshopt; the Compact preset adds vertex quantization) and writes name.web.glb with a before/after summary. Textures are left as they are.
• Compare with the committed version. 3D: Compare with the Committed Version reads the file at HEAD straight out of git, opens it in a second viewer beside yours, and writes a structural diff as Markdown: what changed in geometry, materials, textures, skeleton and animations, with a severity that says whether a consumer of the model breaks. The same comparison is offered after a refinement or an optimization.
• <agent-3d> embeds, checked as you type. In HTML, JSX/TSX, Vue, Svelte, Astro, Markdown and PHP files the extension reports an embed with no source (error), no size (warning: it collapses to 0×0), a library <script> on the latest channel (warning), a pinned script with no integrity (hint) or a stale one (error: the browser refuses the file), unknown or misspelt attributes, and invalid enum values. Every finding has a quick fix, including "pin to the current release with its hash". Hover any attribute for its docs, get completions for attribute names and values, and use the CodeLens to preview the referenced model in the viewer.
• Embed snippet, correctly pinned. 3D: Insert <agent-3d> Embed Snippet resolves the current library version and its subresource-integrity hash from the live release manifest and writes a ready-to-paste snippet at your cursor.
• Forge gallery. Browse recent public three.ws creations in the sidebar, preview any of them in the viewer, and import the ones you want.

Search for three.ws 3D in the Extensions view, or install it by ID:

code --install-extension threews.vscode-3d

It is listed on the Visual Studio Marketplace and, for VSCodium, Cursor, Windsurf and other Open VSX editors, on Open VSX.

To build it from source instead:

git clone https://github.com/nirholas/three.ws.git
cd three.ws/packages/vscode-3d
npm install
npm run package
code --install-extension vscode-3d-0.2.0.vsix

For development, open packages/vscode-3d in VS Code and press F5 to launch an Extension Development Host. Maintainers: releasing a new version is covered in PUBLISHING.md.

Open any .glb or .gltf in the explorer. That is the whole setup: the viewer is the default editor for those files, and the three.ws 3D icon in the activity bar lists every model in the workspace alongside the public gallery.

A typical loop: 3D: Generate an Avatar from Text, describe a character, and the GLB lands in models/ and opens. Press Animate, type wave, and watch it wave. Press Bake clip to save astronaut-waving.glb. Press Refine, type add a glowing visor, and get astronaut-v2.glb. Press Optimize for astronaut-v2.web.glb, then Insert embed to put it on a page:

<script
  type="module"
  src="https://three.ws/agent-3d/1.5.2/agent-3d.js"
  integrity="sha384-…"
  crossorigin="anonymous"
></script>

<agent-3d
  body="https://three.ws/cdn/creations/…/mesh.glb"
  style="width: 400px; height: 500px; display: block;"
></agent-3d>

If that snippet later drifts (someone swaps the URL to latest, or the hash goes stale after a release), the editor underlines it and the quick fix repairs it.

Generation, refinement, rigging, the library, and the gallery go through public, unauthenticated three.ws endpoints:

• POST /api/mcp-studio is the JSON-RPC studio. The extension calls the forge_free, text_to_avatar, rig_mesh, refine_model, and check_job tools. No key, no account, no payment: the platform's server-side keys cover provider cost.
• GET /api/animations/library lists the motion library; each clip is fetched from the CDN as a three.js AnimationClip JSON when you pick it.
• POST /api/forge-motion starts a text-to-motion job; GET ?job= polls it.
• POST /api/forge-quality-check scores a render you send as a PNG.
• GET /api/forge-gallery is the public feed of recent creations.
• GET /agent-3d/versions.json is the library release manifest the embed snippet and the embed diagnostics pin against.

Everything else is local. The viewer runs three.js inside a webview whose content security policy allows exactly one nonce-gated script and no remote origin, so the only things that cross the network while you look at a model are the model file and, when you pick one, a clip. Retargeting runs in the webview with the platform's animation-retarget.js; baking, optimization, inspection and the structural diff run in the extension host with glTF-Transform. The Draco and Basis decoders are copied out of three.js at build time and served from the extension, never from a CDN.

Those three need a URL the browser (or the studio) can fetch. Models this extension generated or imported carry their three.ws URL, so the commands just work. For a local file that has never been uploaded, the extension asks for the https URL where it is hosted rather than handing you a snippet that renders nothing. Animating, baking, optimizing, comparing, and the quality check all work on a purely local file.

npm run watch   # rebuild the host bundle and the webview on change
npm test        # builds, then runs the suite

npm test activates the bundled extension against a stand-in extension host and asserts that every command, view, language provider, and custom editor the manifest promises really registers, that the viewer webview is nonce-gated and CDN-free, that an HTML file with a broken embed yields the right diagnostics and quick fixes, and that a canonical clip retargets onto the committed reference avatar and bakes into it. The studio, library, motion, refine, quality, embed, optimize, compare and naming modules are unit tested against local HTTP servers and real committed GLBs.

• docs/vscode.md walks through the whole workflow.
• @three-ws/vscode-x402 is the sibling extension for paying x402 endpoints from the editor.
• docs/embedding.md documents every <agent-3d> attribute the diagnostics and completions know.
• @three-ws/glb-diff is the structural differ behind the compare command, also available as a CLI for CI.
• @three-ws/glb-tools does inspection and baking from the shell.

Apache-2.0. Built by three.ws.

If vscode-3d saves you time, star it on GitHub. Stars are how other developers and AI agents find the repositories worth trusting, and they cost you one click.

Know someone who would use it? Post on X · Share on Bluesky · Share on LinkedIn · Submit to Hacker News · Share on Reddit

Coding agents and LLM tooling can read this repo directly: AGENTS.md, llms.txt, llms-full.txt. Point an agent at https://github.com/nirholas/vscode-3d and it has the context it needs.

• All repositories by nirholas: the full catalog, grouped by topic
• three.ws: the platform for 3D AI agents with Solana wallets, a skill marketplace and x402 payments
• Questions or ideas: open an issue or start a discussion

three.ws is a platform for 3D AI agents with Solana wallets: avatars, a skill marketplace, x402 payments and more than seventy MCP servers. vscode-3d is one package from it.

• Documentation site: https://nirholas.github.io/vscode-3d/
• npm: https://www.npmjs.com/package/vscode-3d
• Canonical source: https://github.com/nirholas/three.ws (this repository is a generated mirror of packages/vscode-3d)
• Agent-readable summary: llms.txt and AGENTS.md
• Issues and ideas: https://github.com/nirholas/vscode-3d/issues