@three-ws/see

Give an AI agent eyes for 3D.

Every text-to-3D API on earth answers with a URL to a binary file. A human clicks it. An agent cannot: a .glb is opaque to a language model, so the agent that just generated an asset has no way to check its own work. That is why agentic 3D has been stuck at one shot and a shrug.

This package closes the loop in one call. Hand it a GLB URL and get back frames of the model rendered from several angles, ready to put in front of any multimodal model, plus the geometry facts and a plain reading of what they mean.

npm install @three-ws/see

No key. No account. The service is free.

Look at a model

import { see } from '@three-ws/see';

const look = await see('https://example.com/robot.glb');

look.views;
// [ { view: 'three-quarter', theta: 35, phi: 78, imageUrl: 'https://…/a.png' },
//   { view: 'front', … }, { view: 'side', … }, { view: 'back', … } ]

look.notes;
// [ '12,400 triangles, a normal real-time budget for a hero prop or character.' ]

look.stats;      // { vertices, triangles, materials, textures, animations, … }
look.viewerUrl;  // spin it in a browser
look.arUrl;      // place it in a real room from a phone

Close the loop: generate, look, judge, fix

The reason this exists. An agent that can see its own output stops guessing.

import Anthropic from '@anthropic-ai/sdk';
import { see, toMessageContent } from '@three-ws/see';

const anthropic = new Anthropic();

async function critique(glbUrl, intent) {
  const look = await see(glbUrl);
  const message = await anthropic.messages.create({
    model: 'claude-sonnet-5',
    max_tokens: 512,
    messages: [{
      role: 'user',
      content: [
        { type: 'text', text: `This model was meant to be: ${intent}. Judge it. Is the subject complete and recognisable, is the far side finished, is anything melted or fused? If it needs work, say exactly what to change.` },
        ...(await toMessageContent(look)),
      ],
    }],
  });
  return message.content[0].text;
}

toMessageContent shapes a look into multimodal chat content: one labelled text block per angle followed by the image itself. It passes image URLs by default; { fetchImages: true } inlines base64 for APIs that will not fetch a URL themselves.

Pick your angles

await see(glbUrl, { views: ['back', 'top'], size: 768 });

View	What it catches
three-quarter	Form and depth, the way a product shot reads (default first)
front	Facing, symmetry, the subject's identity
side	Profile, depth collapse, flattened geometry
back	The half a single-view generator most often leaves unfinished
top	Footprint, layout, whether it is hollow
bottom	Missing caps, inverted normals, floating geometry

Up to six views per call, size from 128 to 1024 pixels (default 512). Unknown angle names are ignored rather than failing the call, so a guess still returns something useful.

Errors are typed

import { see, SeeError } from '@three-ws/see';

try {
  await see(url);
} catch (err) {
  if (err instanceof SeeError && err.code === 'rate_limited') {
    await new Promise((r) => setTimeout(r, err.retryAfter * 1000));
  }
}

invalid_url, render_failed, rate_limited, timeout, unreachable. A partially-rendered turntable is not an error: you get the frames that worked plus look.missingViews naming the angles that did not, because three good views beat a failure.

Options

Option	Default	Notes
views	['three-quarter','front','side','back']	Up to 6, from the table above
size	512	Square pixels, 128 to 1024
stats	true	Set false to skip the geometry pass
baseUrl	https://three.ws	Point at another deployment
timeoutMs	120000	Rendering holds a browser for seconds
fetch	global fetch	Inject for tests, proxies, tracing

Using MCP instead?

The same capability is an MCP tool, look_at_model, on https://three.ws/api/mcp-studio. There the frames come back as MCP image content blocks, so a multimodal client renders them straight into the conversation and the model sees the model with no glue code at all.

Related

• three.ws/docs/3d-vision: the full surface, REST and MCP
• POST /api/3d/look: the endpoint underneath (GET it for a discovery doc)
• @three-ws/retarget: retarget animation onto any humanoid GLB

Apache-2.0.