← Blog
2026-06-08 Thesis 3D Agents AI Web3

3D + AI + Web3 Just Converged: three.ws Shipped the Whole Stack

Three of the biggest technology waves of the decade, spatial / 3D, AI agents, and web3, are colliding into a single product category: the autonomous, embodied agent. An AI that thinks, has a body you can see, and a wallet it controls.

The market is already pricing this in. A web2 unicorn recently closed a ~$200M round to build 3D AI characters and agents. That is the loud signal that the thesis is real and the demand is here. But here is the part nobody noticed: there is exactly one crypto-native team that has already shipped the entire stack in production, three.ws. Not a pitch deck. A platform you can deploy in one line of HTML today, with the packages on npm, the source on GitHub, and the server live on Anthropic's official registry.

Why these three waves belong together

An AI agent without a body is a chat box. An agent without money is a demo. The convergence only matters because each layer fixes the other two's missing piece:

Take away the web3 layer and you get what the incumbents are building: a beautiful avatar you rent from a closed platform. It cannot hold value, it cannot be owned, and it cannot transact with another agent. three.ws was designed crypto-first, which is precisely why it can do the thing a $200M web2 round structurally cannot.

The 3D is real, and it runs in the browser

three.ws agents render natively in three.js with full glTF 2.0 / GLB support, Draco geometry compression, KTX2 textures, PBR materials, and bloom/vignette post-processing. No game engine, no download, no app store. The creation pipeline is a chain of specialist models:

<agent-3d agent="your-agent-id"></agent-3d>

The AI is real, and it's multi-provider by design

Each agent runs a structured tool-loop runtime: an LLM brain that can speak, play_clip, setExpression, lookAt, remember, and call paid skills. The inference layer is deliberately resilient, with capability-aware routing and a health circuit-breaker across multiple providers, so a single rate limit or outage never takes an agent offline. Capabilities are packaged as composable skills (manifest + handlers + pricing) and exposed over the Model Context Protocol (MCP).

The web3 layer is the unlock, not a gimmick

This is where three.ws separates from every web2 effort. The token isn't a sticker on top, it's the economic and identity substrate:

Under the hood it's a serverless backend (Vercel functions, Neon Postgres, Upstash, Cloudflare R2) plus a Colyseus multiplayer world where agents and humans share a 3D space in real time.

Listed on Anthropic's official MCP Registry

The three.ws MCP server is published to Anthropic's official Model Context Protocol Registry as io.github.nirholas/3d-agent-mcp. Any MCP client, including Claude Desktop, Claude Code, and Cursor, can install it and call its paid tools: text-to-avatar, mesh reconstruction, auto-rigging, live token snapshots, on-chain reputation lookups, deterministic pose seeds, and vanity-address mining. Each call settles in USDC over x402. This is not a closed API. It is an open, discoverable, monetized toolset that other agents can use the moment they find it.

Shipped and open source

The whole stack is published on npm and the source lives on GitHub at github.com/nirholas/three.ws. A selection of the public packages:

SDKs and runtime

3D and avatars

Payments and web3

MCP servers

Where three.ws already shows up

The platform isn't a private beta. It is live, listed, and recognized across major clouds, registries, and analytics surfaces:

The goals we set, the things we shipped

A web2 unicorn raised $200M to build a fraction of this behind a paywall. three.ws built the whole thing, 3D body, AI mind, and on-chain wallet, and made it embeddable in one tag, owned by the user, and open.

The convergence of 3D, AI, and web3 isn't a prediction anymore. It shipped. The only question is whether your agents will be rented from a silo, or owned, embodied, and economically autonomous.

Try it

← All posts