three.ws says Nicholas Resendez's six-person team combines 3D generation, animation, model routing, browser embeds and USDC payments in one stack.
By RuntimeWire Staff · Published
Primary source: Hugging Face Newsroom
Why it matters #
Three.ws shows how falling 3D-generation costs can turn embodied agents into a developer product, while its wallet features raise the security bar far above a typical avatar tool.
Nicholas Resendez (@nichxbt), the three.ws founder responsible for product, engineering and the agent platform, is trying to turn the AI agent from a chat window into a 3D character that can remember users, call tools, move through augmented reality and pay for services.
The architecture behind the product was assembled as the platform has developed, and documented in detail in an August 29 Hugging Face Community Article. Published by the three.ws account, the article reads like an engineering inventory. The platform routes prompts across multiple 3D-generation models, rigs and animates the resulting files, connects them to language models, gives agents optional on-chain identities and distributes them through websites, ChatGPT, Android and MCP servers. Because the Community Article is a three.ws-published source, its product specifications and operating counts are self-reported unless otherwise noted.
Resendez describes the work as one integrated product built around four requirements: body, brain, wallet and distribution. That breadth is three.ws' defining bet. Convai and Inworld AI have built runtimes for interactive characters, while Soul Machines sells enterprise-oriented digital people. Meshy, Tripo and Rodin concentrate on generating 3D assets. Resendez is packaging generation, animation, agent behavior, publishing and payments inside an Apache-2.0 repository that developers can inspect and run themselves.
A founder assembling the whole pipeline
Three.ws identifies Resendez as the founder responsible for product, engineering and the agent platform. An earlier Entrepreneur author profile describes him as a developer and security researcher who previously founded Grindez and uALREADY and studied information technology and marketing at the University of Wisconsin-Milwaukee. He also wrote a 2020 book about automating Twitter accounts without writing much code.
That background fits the way three.ws has been built: a large collection of services, packages and integrations organized around making complicated infrastructure accessible from a browser. Three.ws says in its published company materials that its engineering-heavy team has six people, although the reviewed materials publicly name only Resendez and developer Catherine Macoviak.
The first step is Forge, which accepts a text prompt, one to six photos or a sketch and returns a downloadable GLB model. Three.ws routes jobs according to cost and availability instead of committing to a single generation model. Its registry includes Microsoft's TRELLIS through NVIDIA NIM, Tencent's Hunyuan3D models, TripoSR and TripoSG-scribble, with paid or bring-your-own-key routes to other providers.
Three.ws offers a 12,000-polygon draft tier, a 30,000-polygon standard tier with 2K textures and a 200,000-polygon high tier with PBR materials. The platform says its selfie reconstruction pipeline can produce a fixed-topology avatar in about five seconds on a CPU, using 52 ARKit expression shapes and 15 visemes for speech. Those performance figures come from three.ws and have not been independently benchmarked.
The next engineering problem is movement. A rigging worker predicts a 52-bone skeleton, while a retargeting layer translates animations across naming conventions used by tools including Blender, VRoid, Unreal, Roblox and Mixamo. Three.ws says a model that cannot be driven falls back to a default body instead of freezing in the familiar T-pose. Its motion library contains roughly 3,000 clips, alongside text-to-motion, video-to-motion and webcam facial-capture tools.
The model is a router, too
Resendez has taken the same multi-provider approach to the agent's brain. Three.ws routes requests among commercial and open models and gives agents tools for speech, memory, gaze, animation and screen interaction.
The runtime includes memory that can be used to preserve information about users and insert it into the model's context. Agents can also control facial expressions, head movement and gestures. Voice support includes browser speech and integrations with external voice services.
Developers can place the resulting character on a page through a web component. Three.ws also publishes SDKs, REST endpoints and MCP servers, giving Resendez several routes into products that already have their own interfaces. Three.ws reported 101 npm packages, 72 MCP servers in the official registry, 106 rigged characters and more than 500 public-domain props as of August 25. These company-reported counts describe the amount shipped, rather than the number of people using it.
That distinction matters for a project with such a wide surface area. Three.ws has demonstrated a substantial body of working code and publicly accessible product pages. Active users, paying customers, live third-party embeds and recurring revenue remain the figures that would establish whether developers want the entire stack instead of selecting narrower tools from established character-runtime and 3D-generation providers.
Wallets make the product harder to dismiss, and harder to trust
The wallet layer is where Resendez's technical ambition meets the largest operational risk. Three.ws uses the x402 payment protocol to let an agent encounter a paid HTTP endpoint, settle the charge in USDC and retry the request with payment proof. Resendez's team supports settlement on Solana, Base and BNB Chain, while the broader platform includes optional blockchain identities and autonomous trading tools.
Three.ws reported 110,416 on-chain USDC settlements and 803,483 payment verifications through its facilitator as of August 25. Those company-reported counts do not establish transaction value, unique paying users or revenue. Three.ws separately says creators can keep 100% of pay-per-call earnings during its open beta.
Resendez has built transaction and spending controls into the runtime before an agent can move funds. The controls are intended to let a transaction proceed, require approval or block it. That is the right engineering target, though any product that gives software custody and autonomous trading powers will ultimately be judged through security audits, incident handling and real losses prevented.
The three.ws investor deck says Resendez is raising capital to make the platform a default embodiment layer for AI. It does not identify a round size, valuation or investors. The deck presents premium generation, developer subscriptions, enterprise contracts and transaction fees as potential revenue lines, while free generation and embeds supply distribution.
For now, Resendez's strongest evidence is the stack itself. Three.ws has compressed a collection of jobs that usually require separate 3D, AI, payments and web-infrastructure vendors into one open repository and a browser flow. The next test is narrower and less theatrical: whether developers trust one small team to operate all of those layers in production.