Open-source platform for 3D AI agents. Turn text or a photo into a rigged, animated GLB avatar, give it an LLM brain, memory and a wallet, and embed it anywhere with one web component. Remote MCP server, x402 payments, Three.js, browser-native.
The prompt library
three.ws/prompts is a library of prompts written to
be copied out and pasted into your own assistant. Each one is a complete
instruction: what to build, which endpoint or tool to use, and what to hand back.
You replace the parts in <angle brackets> and send it.
It exists because the gap between "three.ws has a free text-to-3D API" and "my Claude just made me a rigged avatar" is a paragraph of prose that nobody should have to write twice. The library writes it once, correctly, for the paths that actually work.
Everything in it is real. The endpoints are live, the tool names are the ones the servers advertise, and nothing in a prompt asks an assistant to pretend.
The three ways in
Most prompts need nothing installed. The rest ask for one connection you make once and keep.
| Setup | What it means | How |
|---|---|---|
| No setup | The prompt tells your assistant which public endpoint to call. It works anywhere Claude can make an HTTP request or run a command. | Nothing to do. |
| Free MCP server | The three.ws 3D Studio as native tools: generate, rig, refine, look at a model, give it a persona. No account, no key, no payment. | claude mcp add --transport http three-ws-studio https://three.ws/api/mcp-studio, or add that URL as a custom connector in the Claude apps with authentication set to None. See MCP Studio. |
| Skills pack | The Agent Skills pack, so your assistant already knows the platform's endpoints and rules and your prompts can be one sentence long. | Install the three-ws-core plugin, or copy a folder from .agents/skills/ into your project's .claude/skills/. |
| Agent wallet | A funded agent wallet, for the prompts that buy or sell calls in the x402 economy. | Start at three.ws/start. |
Using one well
- Replace every placeholder.
<a weathered brass diving helmet>is a slot, not a suggestion. A prompt sent with the brackets still in it produces a diving helmet. - Keep the verification lines. Several prompts end by telling the assistant to look at the result before declaring success, or to poll a queued job instead of giving up on the first pending response. Those lines are there because removing them is how you get a confident answer about a model that never finished generating.
- Keep the stop lines. The prompts under "the agent economy" that can spend money are written so the assistant has to show you the amount and the recipient and wait for your yes. That sentence is load-bearing. Do not trim it.
- Edit freely otherwise. These are starting points. The library's job is to get the endpoint, the polling and the failure handling right so you can spend your attention on the part that is specific to you.
Machine-readable
The library is published as data, so an agent can read the whole thing without scraping the page:
curl -s https://three.ws/prompts.json # structured: categories, setups, tags, prompts
curl -s https://three.ws/prompts.txt # the same prompts as plain text
prompts.json carries the counts and indexes the page renders from
(categories, setups, tags, byCategory, bySetup), and every prompt
object holds id, title, summary, category, setup, prompt, returns,
tags and links. prompts.txt is the version to pipe into another assistant
whole:
curl -s https://three.ws/prompts.txt | pbcopy
Deep-link to a single prompt with its id: https://three.ws/prompts#rig-my-glb.
Filters live in the URL too, so a link can carry a whole view:
https://three.ws/prompts?category=create&setup=none.
Adding a prompt
The library is one file: data/prompt-library.json.
A prompt is an object in prompts, with the body written as an array of lines
so a diff shows what changed:
{
"id": "rig-my-glb",
"title": "Rig a GLB you already have",
"summary": "Add a humanoid skeleton and skin weights to a static model.",
"category": "create", // must exist in `categories`
"setup": "mcp-studio", // must exist in `setups`
"tags": ["rigging", "glb", "mcp"],
"prompt": ["Rig this GLB so it can be animated: <https://example.com/x.glb>", "", "..."],
"returns": "A rigged GLB plus a pose-studio link to check the joints.",
"links": [{ "label": "Rig Doctor", "href": "/rig-doctor" }],
"added": "2026-09-20"
}
Then rebuild the published files:
npm run build:prompt-library
scripts/build-prompt-library.mjs writes public/prompts.json and
public/prompts.txt, and validates the entry on the way through: ids are unique
and kebab-case, the category and setup exist, placeholder brackets are balanced,
every internal link is a route declared in data/pages.json, and the house ban
on dash glyphs holds. npm run check:prompt-library runs the same validation
without writing, and it is wired into npm run gate, so an edit that never got
built cannot ship a stale library.
Write a prompt the way you would write an instruction to a capable colleague who has never used the platform: name the endpoint or the tool, say what to do when it returns a pending job, and say what you want handed back. If a prompt can spend money, make it stop for a confirmation.
Related
- Agent Skills: the same capabilities as skills your assistant loads automatically, rather than prompts you paste.
- MCP Studio: the free 3D MCP server most of these prompts use.
- 3D API: the keyless HTTP endpoints behind the no-setup prompts.
- Embedding: what the web-component prompts are driving.