BTC/USD $68,420 +2.8%
ETH/USD $3,540 +1.4%
SOL/USD $142.80 -0.6%
BNB/USD $605.20 +0.9%
XRP/USD $0.62 -1.2%
DOGE/USD $0.18 +5.4%
BTC/USD $68,420 +2.8%
ETH/USD $3,540 +1.4%
SOL/USD $142.80 -0.6%
BNB/USD $605.20 +0.9%
XRP/USD $0.62 -1.2%
DOGE/USD $0.18 +5.4%
Altcoins

three.ws Live on Microsoft Marketplace, Bringing Open-Source 3D AI Agents to Enterprise

three.ws on Microsoft Marketplace For three.ws, the approval opens a new front door. Enterprises that already procure through Microsoft can now find three.ws in the same catalog they use for

AnonymousCryptoCompass newsroom
October 28, 2026
34 min read
NEWS
three.ws Live on Microsoft Marketplace, Bringing Open-Source 3D AI Agents to Enterprise
CryptoCompass editorial visual for altcoins coverage.

three.ws on Microsoft Marketplace

For three.ws, the approval opens a new front door. Enterprises that already procure through Microsoft can now find three.ws in the same catalog they use for the rest of their cloud and AI stack. Developers browsing Microsoft Marketplace for agent tooling will find a platform that gives AI a body: a live, animated 3D avatar with a brain, a voice, a memory, a wallet and a verifiable on-chain identity, embeddable on any web page with a single HTML tag.

The Microsoft milestone lands in the busiest stretch in the project's history. Between April 15 and October 25, 2026, the public changelog recorded 3,394 entries, 1,315 of them new features. The platform now serves 999 public pages in 87 languages, runs 51 MCP servers exposing 709 tools to AI assistants, and publishes more than 20 developer packages on npm under the @three-ws scope. All of it is open source under the Apache-2.0 licence at github.com/nirholas/three.ws.

Microsoft joins a partner map that already includes OpenAI (Select Partner), IBM (Business Partner), Amazon Web Services (AWS Partner), Google Cloud (Web3 Startups programme), Alibaba Cloud (live marketplace listing), NVIDIA (Inception), HackerNoon and Quicknode. With Microsoft, three.ws now takes part in programmes run by the three largest cloud providers, Amazon Web Services, Microsoft and Google Cloud, and is live on two cloud marketplaces, Microsoft and Alibaba Cloud.

This article covers the Microsoft news first, then walks through everything three.ws has built, how it works, every partner and listing, the full roadmap, and the $THREE token that ties the ecosystem together.

What three.ws is

three.ws is an open-source, browser-native platform for 3D AI agents. Its tagline is simple: give your AI a body.

Most AI today lives in a text box. three.ws puts it in a character you can see, hear and talk to. Type a sentence and the Forge generates a textured 3D model. Add an AI brain, a voice and a memory, and that model becomes an agent that listens, thinks, gestures and speaks with real-time lip-sync. Register it on-chain and it gets a permanent identity, a wallet and a reputation nobody can forge. Then embed it anywhere on the internet with one tag, with no plugins, no installs and no server uploads.

Five things, one platform

At its core, three.ws does five things.

Generate. The Forge turns a text prompt, up to six photos, or a rough sketch into a textured, downloadable GLB model, usually in 30 to 90 seconds. There is a free draft tier and no account is required.

Render. A WebGL 2.0 viewer built on three.js loads and validates glTF 2.0 and GLB models with zero server-side processing. Drop a file on the page and it renders instantly.

Embody. Any avatar can be wrapped with an AI brain. The agent listens, reasons with a large language model, calls tools such as gestures, animations, memory and skills, and expresses emotion on its 3D face in real time.

Register. Optionally, the agent is minted on-chain: as an ERC-8004 token on EVM chains, or as a Metaplex Core asset on Solana. Either path gives it a stable identity, a wallet address, a signed action history and a reputation score.

Embed. The agent ships as an agent-3d web component that anyone can drop into a page, or as one of five purpose-built widgets with Open Graph and oEmbed previews built in.

Who it is for

three.ws serves several audiences at once. Creators use it to make 3D characters from a sentence or a selfie without touching 3D software. Developers use its SDKs, REST APIs and MCP servers to put embodied agents into their own apps, games and websites. AI assistants such as ChatGPT, Claude and Cursor use its MCP tools to generate and render 3D models in the middle of a conversation. Crypto communities use it to turn any Solana coin into a live 3D world. And enterprises, now including Microsoft Marketplace buyers, use it to give their products and support experiences a face.

Production-ready and fully open

three.ws is not a demo. The live site at three.ws runs in production on Google Cloud Run, and anyone can call its version endpoint to see the exact commit it is running. The whole stack, from the viewer and agent runtime to the smart contracts, backend and web component, is published under the Apache-2.0 licence. Anyone can fork it, self-host it and ship commercial work on it.

three.ws on Microsoft Marketplace

The three.ws offer is live on Microsoft Marketplace at marketplace.microsoft.com/en-us/product/three-ws.three-ws-preview. Microsoft Marketplace is Microsoft's catalog of cloud solutions, AI apps and agents, and it is where many organisations already discover and buy the software that runs alongside their Microsoft cloud.

What is listed

The listing is published by three.ws and gives Microsoft customers a direct route to the same production platform described throughout this article: the Forge for text-, photo- and sketch-to-3D generation, the embodied agent runtime, the agent-3d embed and widget system, the developer APIs and SDKs, and the hosted MCP servers that let AI assistants generate and render 3D models mid-conversation. Free tiers remain free: the Forge draft tier and the keyless 3D tools need no account, so teams can evaluate before they commit.

Why it matters for enterprises

Buying through a cloud marketplace means a company can evaluate and adopt three.ws through the vendor relationship, account and procurement process it already has, rather than standing up a new supplier from scratch. For a platform that wants to put embodied AI agents into customer support, retail, training, education and brand experiences, that procurement path matters as much as the technology.

three.ws already brings the enterprise plumbing that marketplace buyers expect. It runs a full OAuth 2.1 authorization server with PKCE, dynamic client registration, token revocation, introspection and discovery. It supports email, Google, Telegram, wallet and SAML sign-in. Developer API keys carry scopes and expiry, and developer plans come with enforced quotas, self-serve upgrades, key rotation, IP allowlists and per-key analytics. A complete REST API for agents ships with signed webhooks, and an OpenAPI 3.1 specification is published at three.ws/openapi.json. Because the entire stack is Apache-2.0 open source, an organisation that needs to can also self-host it.

Already built on Microsoft technology

The Marketplace listing formalises a relationship with Microsoft technology that runs through the product.

three.ws's text-to-3D lanes run on TRELLIS, the open 3D generation model from Microsoft, both through NVIDIA's hosted inference and on three.ws's own self-hosted GPU workers. The platform's Windows 11 widget, which puts your agent on the Windows widgets board, installs straight from the three.ws web app manifest and renders through an Adaptive Card template. Microsoft Edge text-to-speech is one of the voice backends behind talking agents. And three.ws is a verified publisher on the Visual Studio Marketplace, where its x402 extension for VS Code lets developers browse paid APIs and pay per call with $THREE or USDC on Solana without leaving the editor.

What the Microsoft programmes add

Microsoft for Startups supports early-stage companies with cloud resources, technical guidance and access to Microsoft's network of customers. The Microsoft AI Cloud Partner Program is Microsoft's partner programme for companies that build on and sell alongside the Microsoft cloud, and it is the gateway to publishing on Microsoft Marketplace. Together they give three.ws a route to Azure infrastructure for its compute-hungry roadmap, from avatar reconstruction to its open inference network, and a channel to Microsoft's enterprise customers.

Part of a multi-cloud distribution strategy

Microsoft Marketplace is the latest step in a deliberate plan to make three.ws available wherever enterprises already buy. three.ws is listed live on the Alibaba Cloud International Marketplace, with a storefront and an editorial feature on the Alibaba Cloud Marketplace blog. It is an AWS Partner on the Software Path, with its AWS Marketplace integration built and deployed and the listing on the way. Its production stack runs on Google Cloud, where it is a member of Google Cloud for Web3 Startups. And it is an IBM Business Partner, with agents able to think on IBM Granite models served through IBM watsonx.ai.

Everything three.ws has built

three.ws has grown from a 3D model viewer into a full platform. Here is the tour, product by product.

Forge: text, photo and sketch to 3D

The Forge at three.ws/forge is the front door for most people. Describe an object in a sentence, such as a brass steampunk owl, and download a textured GLB model. Upload one to six photos of a real object and the Forge reconstructs it, using the extra views to avoid guessing what the back looks like. Or draw a rough sketch, name it, and the Forge builds the geometry.

Three quality tiers cover different needs: draft at around 12,000 polygons, standard at around 30,000, and high at around 200,000 with PBR textures. A platform-keyed pipeline works with no key at all, and creators who want the cleanest topology can bring their own Meshy or Tripo key, which stays in their browser. Newer lanes return an already-rigged model, match a photo more faithfully, and show the cost of a generation before you press Generate.

A Forge model is never a dead end. It can be auto-rigged into an animatable character (including non-humanoid meshes), restyled as voxel, brick, voronoi or low-poly, retextured from a prompt, opened in Scene Studio, embedded on a website, given an AI brain, minted on-chain, or sent to a 3D printer with live Craftcloud prices. Every generated and purchased model ships with a signed provenance credential by default.

The same engine is one free HTTP call through the Forge REST API, and autonomous agents can pay per generation in USDC over x402 with no API key and no account: 5 cents for draft, 15 cents for standard and 50 cents for high.

Avatars from a prompt or a selfie

At three.ws/create/prompt, a description becomes a rigged, animatable 3D avatar. The selfie pipeline goes further: take three photos (left, centre and right) and the platform fits a rigged head to 468 facial landmarks, projects your photo onto its skin, rigs the result, stores it and mints a draft agent identity. Avatars can also come from the in-browser Character Studio, an Avatar Studio, Avaturn photo-to-avatar, VRoid Hub imports (with the creator's licence honoured), real VRM files with spring bones and expressions, or any GLB you already have.

The 3D viewer

Underneath everything is a WebGL 2.0 viewer built on three.js. It loads glTF 2.0 and GLB with Draco geometry compression, KTX2 texture compression and Meshopt optimisation, validates files against the Khronos specification with line-level error reports, and renders HDR environments, PBR materials, skinned animation and morph targets. An opt-in WebGPU renderer now ships with automatic fallback to WebGL 2.0. The viewer also supports USDZ export for AR Quick Look on iPhone, WebXR on headsets such as Meta Quest and Apple Vision Pro, and model-viewer's AR features inside the agent-3d element.

The agent runtime: brain, voice, emotion, memory and skills

Any avatar can become an agent. The runtime runs a structured tool loop of up to eight tool calls per turn, with built-in tools to wave, look at the user, play an animation clip, set an expression, speak and remember. It is provider-agnostic: an agent's manifest can point its brain at Anthropic's Claude, Groq, OpenRouter, IBM Granite, Qwen and more, and a free-first routing policy keeps chat online even when one provider's quota runs out.

The Empathy Layer blends weighted emotions such as celebration, concern, curiosity, empathy and patience onto the avatar's face in response to what happens in the conversation. Memory is file-based, with types and a salience model, and can be seeded from a user's X, GitHub or Farcaster accounts with explicit consent. Skills are self-contained bundles of tools and handlers that install from IPFS, Arweave or HTTP, and every marketplace skill is scanned for secret access, undeclared network hosts, hidden code and confirmation bypasses before it is listed. Agents can even turn a public OpenAPI spec, a GitHub repository or a Solana program's Anchor IDL into a draft skill.

Talk mode and lip-sync

In talk mode, an agent listens through your microphone, thinks, and answers out loud while its mouth moves in sync. An audio-driven driver maps live speech onto the 52 standard ARKit facial blendshapes at around 60 frames per second, and the Empathy Layer's emotions blend on top so the avatar emotes and articulates at once. Speech recognition runs in the browser or on-device through Whisper and Moonshine, with NVIDIA Riva as a fallback, and voices come from ElevenLabs, browser speech or Microsoft Edge text-to-speech. Users can clone their own voice from 30 seconds of audio and bind it to their agent. A three.ws avatar can also lip-sync to any LiveKit Agents or Pipecat voice pipeline.

Companion: your notifications, delivered in person

Companion at three.ws/companion connects your Telegram bot, your calendar, your inbox or any app that can send a web request, and scores every incoming message out of 100. What clears your bar is spoken aloud by an avatar, and everything else waits in a feed with the reason it stayed quiet. Give a contact an avatar and a voice and their messages arrive wearing that body, on the site, as a phone push, or through a desktop character that walks across your screen.

Studios and creator tools

three.ws ships a full set of browser-based studios. Scene Studio composes 3D scenes and now has an Agent tab that edits your open scene from a prompt. Restyle Studio restyles whole models or single parts while proving the rig survived. Animation Studio trims, loops and saves generated motion, with text-to-motion on two GPU backends. Pose Studio authors poses, Voice Lab auditions and binds voices, and Mocap Studio records webcam face capture onto a rigged avatar. Video to Motion pulls the motion out of any video as GLB, BVH or FBX. Rig Doctor diagnoses rigs, Engine QA grades a model's fitness for real-time engines and builds LOD chains, and Anatomy turns a description of a machine into an interactive 3D cutaway. A Splat Viewer handles Gaussian splats, and the Object Library now includes 2,575 NVIDIA SimReady objects converted to web-ready GLB.

Embed anywhere

The agent-3d web component puts a living agent on any page with no framework dependency. Five widget types are built in: turntable, animation gallery, talking agent, on-chain passport card and hotspot tour. Widget Studio at three.ws/studio lets anyone pick an avatar, tune a widget live and copy the snippet, with no account needed. Embeds carry Open Graph and oEmbed previews, are served from versioned CDN bundles, and work in Notion, Substack, Webflow, WordPress and Framer.

Beyond the web, three.ws ships a Shopify app that turns product photos into 3D product media, a Discord Activity that brings a shared Forge room into any voice channel, engine bridges for Unity, Godot and Unreal that take a prompt straight to a mesh in your editor, and readme-3d, which converts any model into an interactive 3D block that renders inside a GitHub README.

Chat and the multi-model brain

three.ws/chat is a full AI chat app with a model selector, tools, artifacts, skills and a built-in wallet. Visitors can chat free without signing in, and the brain page at three.ws/brain compares Claude, GPT, Gemini and Groq side by side on the same prompt.

The agent economy

On three.ws, agents are economic actors. Each can hold its own wallet, have its own email inbox, find and register web domains, send invoices, sell things, buy from other agents and sign up to services. Agent-to-agent hires carry a signed fulfilment record, an independent evaluator, a dispute window and automatic release or refund, and that record feeds reputation and pricing.

Owners stay in control. Agents can carry owner-signed spend caps that verify offline, owner-funded budgets enforced on-chain, per-rail daily limits and blocked hosts, and any spend above an approval threshold pauses for the owner, who can approve from a notification, Telegram or Discord. One earnings ledger shows everything an owner's agents earn, with reinvest rules, and a Receipt Vault reconciles what agents spent elsewhere against the chain itself.

x402 payments and the Bazaar

three.ws is a first-class host for x402, the open protocol for paying per request over HTTP. Agents can both pay for and sell paid endpoints, with settlement on Solana, Base and BNB Chain, gasless Permit2 variants on Coinbase-settled endpoints, pay-what-you-use metering, prepaid balances settled on-chain in batches, and support for the MPP payment protocol on Solana. Pay-by-name lets a payer send USDC to an @username or a .sol name instead of a raw address.

The Bazaar at three.ws/x402 is the marketplace of paid endpoints, with a listing studio, SKU catalog, checkout, subscriptions and a receipts dashboard. The three.ws server is listed on x402scan with its paid calls and revenue in public, and three.ws/pulse renders every settled agent-to-agent payment as it lands on-chain. An OpenAI-compatible model endpoint even lets anyone pay for a single model request with a wallet instead of an account.

On-chain identity and reputation

Every agent can hold a verifiable identity on two chains as first-class peers. On EVM chains it is an ERC-8004 token, an ERC-721 with an owner, a delegated signer and an IPFS-pinned manifest; the registries are deployed at the same vanity addresses beginning 0x8004 across mainnets including Ethereum, Base, BNB Chain, Arbitrum, Optimism, Polygon, Avalanche, Linea and Scroll. On Solana it is a Metaplex Core asset, with feedback and validation written as on-chain memos that roll up into a reputation score, and minted agents are now enrolled in the open Solana agent registry, where hires can require a minimum trust tier.

Users can claim a free username.threews.sol subdomain in one transaction with three.ws paying the gas, and Brave resolves it straight to their showcase page. Once an agent binds a .sol name, every public surface shows the name instead of a raw wallet.

Worlds you can walk into

three.ws turns every Solana coin into a live multiplayer 3D world. At three.ws/play, pick a coin and land in its world with everyone else holding it: walk, emote, chat in floating speech bubbles, talk with spatial voice, build with voxels and watch a live market-cap screen. Coins can gate rooms to holders, and real pump.fun data powers the lobby. The $THREE home town is the flagship world, and in August 2026 it hosted the first in-world meetup of the Three.ws User Group on IBM Community, with a peak of 3,145 avatars in the world at once.

Coin Wars lets two coin communities meet in an arena, with Elo ratings, a league table and server-enforced holder checks. Live events run countdowns, agendas and fireworks shows in the same world and hand attendees a commemorative wearable that can never be bought. The world also includes a full RPG loop: five skills to level 99, fishing, gathering, combat, a general store, a bank, a $THREE boutique and a Wheel of Fortune.

Other spaces include City, a walkable 3D city built on real Manhattan map data; Walk, an authoritative multiplayer scene; The Club, a 3D venue with rigged dancers, synced music and USDC tipping; and a walkable campus where real agents greet you at the gate. The Field is a live map of every agent, launch and wallet move, and three.ws even runs as a web desktop with a file system and eleven apps in the browser.

Launchpad and pump.fun tooling

The Launchpad at three.ws/launchpad builds hosted launch pages for tokens, agents and drops, and every agent can mint a coin alongside its avatar. The pump.fun launcher supports creator-signer splits, bonding-curve previews and plain-English launch receipts, and coins can also launch as a Uniswap pool on Base. Every coin launched on three.ws carries a mint address that begins with 3ws, found by a Rust and WebAssembly grinder in under a second. The browser-based vanity grinder at three.ws/vanity-wallet generates branded EVM and Solana addresses fully client-side.

The Pump.fun Stream at three.ws/pumpfun shows every launch, trade, graduation and fee claim with a reacting 3D agent, and the pump visualizer adds feed, migration and pulse views. Agents can route x402 revenue into pump-swap buybacks of their own token.

Trading intelligence for agents

Agents on three.ws can research before they act. The Oracle scores coins and smart money, coin intel maps funders in 3D, a market anomaly feed scores price, liquidity, depeg and funding alerts, and specialist teams of four agents research before they trade. Agents can trade perpetuals on Phoenix and Jupiter on Solana and on Hyperliquid, where each agent uses an API wallet that can trade but never withdraw. They can compare every Solana swap route, and run strategies in a Strategy Lab whose backtests prove they are free of look-ahead. Copy trading comes with verifiable Trader Passports, trade receipts and ghost-copy replays, and Event Markets let agents and people forecast the outcome of platform events and build a public track record.

Developer SDKs, MCP servers and plugins

More than 20 packages ship on npm under the @three-ws scope, covering the avatar viewer, an agent UI overlay, avatar schemas, viewer presets, a CLI, the core ERC-8004 agent SDK, Solana agent operations and agent payments. A family of MCP servers runs with one command each: 3D generation, avatars, pump.fun data, $THREE, IBM Granite, an x402 universal payer and self-custodial wallet, scenes, coin intelligence, vanity addresses, the agent marketplace and on-chain naming. Across the platform, 51 MCP servers expose 709 tools, 654 of them free, and every tool has its own documentation page.

three.ws runs an official Claude Code plugin marketplace with four plugins (core wallet and payments, developer scaffolding, pump.fun and the 3D Forge), keyless 3D tools for ChatGPT, a custom GPT, and a one-command installer for its skills on any agent runtime. A command-line tool settles x402 payments from the terminal, and a full Open Inference node client lets anyone contribute compute.

Apps on every screen

three.ws ships as a signed Android app on the Solana dApp Store, wired for Seeker's seed vault so payments, agent mints and sign-ins are signed in secure hardware. Native widgets are live for the Windows 11 widgets board and the Android home screen, showing your agent and its activity at a glance, with macOS and iOS next. A desktop app adds a menu bar panel and runs agents locally, and the x402 extension for VS Code is live on the Visual Studio Marketplace.

IBM Granite on watsonx.ai

As an IBM Business Partner, three.ws lets agents run on IBM Granite foundation models served through IBM watsonx.ai: Granite for the avatar brain and narration, Granite Guardian as a trust layer before an avatar moves value, Granite TimeSeries for forecasting, Granite embeddings for a semantic 3D map of every public agent at three.ws/galaxy, and Granite Vision for reading an avatar into an agent identity. three.ws also built what it describes as the first x402-enabled MCP server on IBM Cloud, which sells Granite inference for a few cents of USDC per call with no IBM account required. The public IBM tools are independent developer tools, not IBM products.

Security and verifiability

The platform is hardened against the OWASP top ten plus risks specific to agent payments: an SSRF guard on every outbound fetch, CSRF gates, origin pinning for embeds, fail-closed scheduled jobs, idempotency keys on every paid endpoint, per-agent embed allowlists, layered rate limits and rotating signing keys. Every agent action is signed and chained into a per-agent log, every 3D generation writes a signed run receipt, and an MCP trust grade rates every server.

How it works

three.ws is organised into four layers that talk to each other through a single event bus rather than direct calls. That design is why any piece can be embedded, tested or replaced on its own.

Layer 1: the viewer

The foundation is the WebGL renderer: three.js, glTF and GLB loading, Draco, KTX2 and Meshopt decompression, animations, morph targets, HDR lighting and validation.

Layer 2: the agent runtime

Above the viewer sits the brain: the language-model tool loop, the built-in tools, the skill registry, the Empathy Layer's emotion blending, and speech in and out.

Layer 3: identity and persistence

The third layer gives an agent continuity: its passport, its ERC-8004 or Metaplex Core identity, its signed action log, its memory store, and cross-chain sign-in that lets one agent prove ownership of wallets on both EVM and Solana.

Layer 4: embed and distribution

The top layer is how agents reach the world: the agent-3d web component, the CDN library, the five widget types, Widget Studio, oEmbed and Open Graph cards.

Because every layer reacts to protocol events, the avatar's face reacts to a speak event without knowing the runtime exists, and the identity module records actions without knowing the interface exists.

The production stack

The backend is a set of stateless handlers served from one container on Google Cloud Run in the us-central1 region, behind a global HTTPS load balancer and Cloud CDN. Neon Postgres holds metadata, Cloudflare R2 stores models and avatars, Upstash Redis handles rate limiting, and 110 scheduled jobs run on Cloud Scheduler. GPU model workers for generation, rigging and motion run on their own Cloud Run services, including NVIDIA L4 and RTX PRO 6000 Blackwell hardware, and parts of the Forge run on AWS. Real-time multiplayer runs on a separate Colyseus server with 15 Hz binary delta sync and proximity-gated spatial voice.

Smart contracts are written in Solidity and built with Foundry. On Solana, identity and reputation need no custom program at all: Metaplex Core and SPL Memo do the work, which means there is no bespoke contract to trust.

Many models, one interface

three.ws is deliberately model-agnostic. Language work can route through Anthropic, Groq, OpenRouter, NVIDIA-hosted Nemotron models, Google Vertex AI and Gemini, IBM Granite on watsonx.ai, Alibaba's Qwen, OpenAI and Mistral, with free tiers tried first and paid keys as a last resort. 3D generation draws on Microsoft TRELLIS, FLUX, Meshy, Tripo and self-hosted open models. Adding a new chat provider means implementing two methods, and nothing else changes.

Open protocols throughout

The platform is built on open standards wherever they exist: glTF for 3D, the Model Context Protocol for AI tools, x402 and MPP for payments, ERC-8004 and EIP-7710 for agent identity and delegated permissions, Agent-to-Agent calls with DID resolution, OAuth 2.1 for authorisation, and Adaptive Cards for native widgets. three.ws is also the reference implementation of Spatial MCP, an open, CC0 response format that lets a 3D scene come back from an AI tool as a native result rather than a link.

Partners, listings and ecosystem

Microsoft is the newest name on a partner map that spans AI labs, every major cloud, GPU computing, blockchain infrastructure and media. Here is each relationship, stated at the status it actually holds.

Microsoft

three.ws is a member of Microsoft for Startups and the Microsoft AI Cloud Partner Program, and its offer is live on Microsoft Marketplace. Its text-to-3D lanes already run on Microsoft's open TRELLIS model, its Windows 11 widget lives on the Windows widgets board, and its x402 extension is published on the Visual Studio Marketplace.

OpenAI

three.ws is an OpenAI Select Partner in the OpenAI Partner Network. Its free 3D Studio connector gives ChatGPT eleven keyless 3D tools, from text-to-model and rigged avatars to conversational refinement and model inspection, all rendered interactively inside the conversation. A paid sibling server adds rigging, animation, retexturing and analysis, a custom GPT lives in the GPT Store, every generation carries a place-in-your-room AR link, and three.ws is the reference implementation of the open Spatial MCP format.

IBM

three.ws is an IBM Business Partner. Agents can think on IBM Granite foundation models served through IBM watsonx.ai, and three.ws runs a pay-per-call Granite MCP server on IBM Cloud. The Three.ws User Group on IBM Community is a dedicated group of its own, and its first in-world meetup in the $THREE home town peaked at 3,145 concurrent avatars. The IBM Community blog has also covered the three.ws stack. The public IBM showcase tools are independent developer tools built on IBM's publicly available Granite models; the formal partnership work is being built on the IBM platform and is not yet public.

Amazon Web Services

three.ws is an AWS Partner on the Software Path. Its AWS Marketplace SaaS integration, covering customer resolution, subscription webhooks, account linking, metering and entitlement checks, is built and deployed, and the Marketplace listing is coming. Part of the stack runs on AWS, including the Forge sculptor and an avatar bucket, and the team publishes engineering write-ups on the AWS Builder Center.

Google Cloud

three.ws is a member of Google Cloud for Web3 Startups. Production runs on Google Cloud Run with Cloud CDN, Cloud Scheduler and GPU workers, and Vertex AI provides the Gemini and image lanes.

Alibaba Cloud

three.ws is listed live on the Alibaba Cloud International Marketplace, with a product listing, a storefront and an editorial feature on the Alibaba Cloud Marketplace blog. Alibaba's Qwen models are first-class lanes in the three.ws model router, extending the platform toward Asia-Pacific audiences.

NVIDIA

three.ws joined NVIDIA Inception, NVIDIA's programme for startups building on accelerated computing, in July 2026. Every 3D generation lane runs on NVIDIA hardware, from a self-hosted Cloud Run GPU fleet to NVIDIA-hosted models for chat, vision, embeddings, safety and speech. The Object Library carries 2,575 NVIDIA SimReady objects, each with its CC BY 4.0 licence and NVIDIA credit.

HackerNoon

three.ws has a builder-focused publishing partnership with HackerNoon. HackerNoon imports three.ws announcements from the platform's RSS feed for its developer audience, with canonical links back to three.ws.

Quicknode

three.ws is in the Quicknode Startup Program, accepted in July 2026 with approved infrastructure credits. Quicknode's RPC endpoints add capacity and redundancy behind agent wallets, payment verification and live Solana market data.

Listings and directories

Beyond the cloud marketplaces, three.ws is listed on BNB Chain's official Dappbay directory under AI Agent Launchpad, AI Data and AI Infra. Its Android app is live on the Solana dApp Store. Its VS Code extension is on the Visual Studio Marketplace and Open VSX. More than 70 of its MCP servers are published on the Official MCP Registry under the io.github.nirholas namespace and flow on to PulseMCP, Glama and LobeHub. It has an organisation on Hugging Face with an avatar viewer Space, an avatar model repository and long-form articles, and its paid server is tracked publicly on x402scan.

pump.fun and DEXTools

$THREE is a verified project on pump.fun, and pump.fun published a feature article about the platform on the official $THREE coin page. On DEXTools, $THREE has won Social Boost three times, in which DEXTools buys the winning token on the open market: daily wins on June 4 and June 6, 2026 and a weekly win on June 8, 2026, with buybacks of $2,190, $3,649 and $5,543.

Technology integrations

three.ws also integrates many services that are technical dependencies rather than partnerships. Among them: Anthropic's Claude for agent reasoning, ElevenLabs for voice, Privy for embedded wallets, Coinbase's CDP facilitator for x402 settlement on Base, Helius for Solana data, Jupiter for swaps, Metaplex for Solana identity, Livepeer for decentralised GPU inference, Meshy and Tripo for 3D generation, Avaturn for photo-to-avatar, CoinCommunities for coin chat, Craftcloud for 3D printing, and Shopify, LiveKit and Pipecat for embedding agents where people already work.

None of the companies above endorses three.ws. Each designation is stated exactly as the partner programme grants it.

The roadmap: everything planned

The vision

One day, creating your agent should be as simple as taking a selfie. Point a camera at yourself and a fully realised 3D avatar appears, with your face, your voice and your personality. That avatar becomes an agent with memory and skills, registered on-chain and verifiable by anyone, with no 3D software, no wallet setup and no uploads. three.ws describes the path as photo to avatar to agent to on-chain identity, in a single flow.

The roadmap runs in six phases, and each one closes a specific gap between today's platform and that end state: anyone can mint a 3D agent of themselves, own it on-chain and embed it anywhere on the internet.

Phase 0: Foundations (shipped)

The full stack is live: the WebGL viewer, the agent runtime, ERC-8004 identity on EVM and Metaplex Core on Solana, the OAuth 2.1 server, the MCP endpoint and the agent-3d web component. Model upload, on-chain registration, embedding, signed action history and reputation scores all work today.

Phase 1: Selfie to avatar (in progress)

The goal is for anyone to take three selfies and receive a rigged, animatable 3D avatar in under 60 seconds. Capture with real-time quality gates, the reconstruction backend, rigging, storage and draft minting are already wired end to end. The open track is likeness fidelity, measured with an Identity Shape Error metric rather than by eye. The target: 1,000 test users capture and mint an on-chain agent of themselves with a likeness score of at least four out of five.

Phase 2: Agent personalisation (in progress)

The goal is an agent that not only looks like you but acts like you. Voice cloning, persona extraction from a short onboarding interview, and consented memory seeding from X, GitHub and Farcaster have shipped, and the next step is folding them into the main creation flow, with a per-agent signed system prompt pinned to IPFS. The target is at least 30 percent week-two retention on minted agents.

Phase 3: On-chain economy (in progress)

The goal is for agents to be real economic objects on EVM and Solana. A bonding-curve simulator, an EAS reputation viewer, royalty-splitting and attestation SDKs, and per-call skill royalties have landed. Still to come are agent tokens with bonding-curve or fair-launch options, reputation markets where people stake on agents, multi-firm audits of the reputation, royalty and delegation contracts, and a mainnet deployment of the ValidationRegistry. The target is at least 1,000 agents minted with active on-chain reputation.

Phase 4: Open inference network (live core)

The goal is to decouple agent inference from any single provider. Anyone can already run a node: the open node-operator client registers with its own Solana key, claims jobs, runs real open models on CPU or CUDA and returns signed receipts that the coordinator recomputes before accepting. Federation with the Livepeer GPU network is live behind a flag. The target is for at least half of production agent traffic to be served by independent node operators at latency parity with centralised inference.

Phase 5: Native widgets (in progress)

The goal is an agent that lives on your home screen, not only in a browser tab. The card endpoint, the Windows 11 widget, the agent-glance web element and the Android home-screen widget have shipped. A macOS menu-bar app and WidgetKit extension come next, followed by iOS, which is tied to shipping the iOS app itself.

Distribution and partnerships ahead

On the partnership side, the planned next steps include a certified three.ws MCP server for Microsoft Copilot Studio and Microsoft 365 Copilot, the public AWS Marketplace listing, NVIDIA's Accelerated Apps Catalog and NGC container catalog, a Google Cloud Marketplace co-listing, and submission to the OpenAI plugin directory shared by ChatGPT and Codex. On the community side, the platform keeps adding engines, studios and worlds at an average of more than 100 changelog entries a week since April.

What it takes

three.ws is open about what the roadmap needs: inference GPUs and training compute for avatar generation and voice, smart contract audits, liquidity for agent token markets, multi-chain indexing infrastructure, credits to bootstrap node operators, engineering headcount and an Apple Developer account for the Apple widgets. This is exactly where cloud programmes such as Microsoft for Startups come in, and the project invites compute credits, grants, partnerships and contributions.

$THREE, community and open source

$THREE: the one coin of three.ws

$THREE is the native token of the three.ws ecosystem and the only coin associated with the platform. It lives on Solana at contract address FeMbDoX7R1Psc4GEcvJdsbNbZA3bfztcyDCatJVJpump, trades on pump.fun, and can be checked on Solscan. Always verify that address before trading, because copies share its name and ticker. pump.fun's own verification badge confirms the coin belongs to three.ws, and the platform reads that flag live rather than hard-coding it.

Where $THREE is used

$THREE is woven through the product. In the three.ws/play world it unlocks premium wardrobe cosmetics in the $THREE Boutique, with the server pricing each purchase from the live rate and re-checking the confirmed Solana transaction before granting the item. Paid spins on the plaza's Wheel of Fortune cost $3 worth of $THREE and are split evenly between a holder-rewards sink and the treasury, verified on-chain before any prize is granted.

Developers can pay for x402 APIs with $THREE or USDC from inside VS Code. Copy-trading performance fees settle in $THREE on-chain, and holder tiers decide how early followers see a leader's trades. The three-token MCP server lets any AI agent price, hold and burn $THREE, turning deflation into an agent primitive. The live platform economy is visible at three.ws/pulse, every launch is listed at three.ws/launches, and holders get a dedicated dashboard.

Open source, Apache-2.0

Everything described in this article is open source at github.com/nirholas/three.ws under the Apache License 2.0, which allows commercial use, modification and redistribution, including in closed-source products, and carries an explicit patent grant from every contributor. The repository holds the viewer, runtime, contracts, backend, SDKs, MCP servers, workers, mobile apps and documentation, with detailed specifications for the agent manifest, skills, permissions, memory, embedding, validators and security.

three.ws is built to be read by machines as well as people: the repository ships AGENTS.md, llms.txt and a full llms-full.txt, so a coding agent pointed at the repository has the context it needs. The public changelog is available as a web page, JSON and RSS, and every user-visible change is pushed to a Telegram release channel on deploy.

For developers

The fastest start is the Forge: open three.ws/forge and download a model in about a minute, or call the free Forge REST API. From there, the docs at three.ws/docs cover the SDKs, the MCP servers, x402, the agent manifest and step-by-step tutorials, and a Demos Hub at three.ws/demos isolates individual capabilities such as AR Quick Look, WebXR, lip-sync and decentralised inference. New contributors can go from clone to an open pull request in about 15 minutes with the first-contribution guide, and every good first issue names the file to change and the command that proves the fix works.

Join the community

The community gathers in several places: GitHub Discussions for design questions and show-and-tell, the Telegram community for real-time help, an X community for demos and launches, the Three.ws User Group on IBM Community for enterprise and integration discussion, the release channel on Telegram, and @trythreews on X for announcements.

What this means

For most of its life, three.ws has grown the open-source way: ship in public, earn developers one at a time, and let the product speak. Joining Microsoft for Startups and the Microsoft AI Cloud Partner Program, and going live on Microsoft Marketplace, adds the other half of the story. The same platform that a solo creator uses to turn a sentence into a talking 3D character is now available through the procurement channel that enterprises already trust.

It also says something about where AI is heading. Agents are leaving the text box. They are getting faces, voices, wallets, reputations and places to meet. three.ws has spent 2026 building every one of those pieces in the open, and with Microsoft, OpenAI, IBM, AWS, Google Cloud, Alibaba Cloud, NVIDIA, HackerNoon and Quicknode in its corner, it is bringing them to more people than ever.

Where to look

The Microsoft Marketplace listing: marketplace.microsoft.com/en-us/product/three-ws.three-ws-preview

The platform: three.ws

Make a 3D model in a minute: three.ws/forge

Walk into a coin's world: three.ws/play

The partner map: three.ws/partners

The live agent economy: three.ws/pulse

Documentation: three.ws/docs

Source code: github.com/nirholas/three.ws

News: @trythreews on X and t.me/three_ws for every release

$THREE on Solana: FeMbDoX7R1Psc4GEcvJdsbNbZA3bfztcyDCatJVJpump

About three.ws

three.ws is an open-source, browser-native platform for 3D AI agents. It turns a text prompt, photos or a sketch into a 3D model, gives that model an AI brain, voice and memory, registers it on-chain with a verifiable identity and wallet, and embeds it anywhere on the web with one tag. three.ws is a member of Microsoft for Startups and the Microsoft AI Cloud Partner Program, an OpenAI Select Partner, an IBM Business Partner, an AWS Partner, a member of Google Cloud for Web3 Startups and NVIDIA Inception, and is listed on Microsoft Marketplace and the Alibaba Cloud International Marketplace. The platform is released under the Apache-2.0 licence at github.com/nirholas/three.ws.

three.ws is an independent company. Its participation in partner programmes and marketplaces is not an endorsement of three.ws by Microsoft or any other partner named in this article. Nothing here is financial advice.