Skip to content

3D engine that AI agents drive

4.4 KB
Raw Blame

Questions people ask

Short answers to the questions that came up most after the launch. Each answer links to the page with the detail.

Does Crucible replace three.js?

No. You can use them side by side. three.js is a general 3D library that you program in JavaScript. Crucible is a smaller renderer with a scene format that an AI agent can edit through MCP tools, without writing engine code. Crucible uses the three.js Draco, meshopt and KTX2 decoders when you pass them in (import and export).

Is it faster or smaller than three.js?

It is smaller. A lit box is 49,858 bytes gzip in Crucible and 133,652 bytes in three.js 0.186.1. A box loaded from a GLB is 77,122 against 155,814 bytes. Both were measured on 2026-11-03 (size bench).

It is not faster in general. On 2026-11-04, three.js drew the faster full frame in most runtime bench scenes. Crucible was faster on two of them: many copies of one object, and moving shadows. The performance page has the numbers.

Why JavaScript, and why not Vulkan or DirectX?

Crucible draws in the browser, so it uses WebGL2, which every current desktop and phone browser has. A page or an agent can show a scene with no install. A native Vulkan or DirectX backend is not planned. A native path tracer package is an open idea in issue #626.

What is the benefit over asking an agent to "make it in three.js"?

With three.js, the agent writes a program and runs it to find out if it works. With Crucible, the scene is data: parts, geometry, matrices and materials. The agent changes it with tool calls, and each call checks its input and says what to fix.

On six modeling tasks with the same model (agent bench, 2026-11-07, one run each):

System Passed Tool calls Tokens (in and out) Wall time
Crucible MCP 6 of 6 3.2 375k 24 s
three.js by code 5 of 6 8.8 269k 359 s
Blender MCP (official server) 6 of 6 3.8 177k 517 s

The tasks were made for the Crucible bench, so they favour it. The Blender times include a slow render inside an emulated container, so do not compare them directly.

Does it use a lot of tokens?

Yes, more input tokens than the other two systems in the table above. The agent client reads every tool schema on every call. The Crucible tool schemas are about 19,000 tokens (node scripts/measure-tool-cost.mjs, measured 2026-11-10). Crucible uses fewer calls and much less time, but each call carries more context. A size check on every pull request stops the schemas from growing without a reason (optimize for agents).

What does the agent see after each call?

Each edit returns the new revision, the result of each operation and a diff of what changed. The agent can then:

  • Read the scene state with scene_inspect.
  • Get a picture with scene_look.
  • Compare several camera shots or branches in one labelled image with scene_review.

The agent asks for a picture only when it needs one, so most calls cost no image. See agent workflows.

Can an agent undo an edit?

Yes. scene_restore with revision goes back to any of the last 4 revisions in one call, with no checkpoint made first. For an experiment that may take longer, call scene_checkpoint first, and scene_restore goes back to it exactly, geometry and textures included. scene_apply with a branch name tries an edit on a copy and leaves the scene as it is. The MCP guide has the limits.

Can I check that an edit did not change the rest of the scene?

Yes. Every scene_apply result has a diff with the IDs of the nodes, materials and geometries that the edit added, removed or changed. Nodes keep the IDs that the recipe gives them, so you can also run scene_inspect before and after the edit and compare the two records.

Which tools do agents use most?

This is not measured yet. Issue #746 adds an opt-in local log of tool calls, and we will post the table on that issue.

How do I help?

Read the contributing guide and pick an issue. Issues marked good first issue are small and self-contained, such as OBJ and STL export (issue #747).