three.ws Now Speaks 100+ Languages: How We Localized a 3D AI Agent Platform for the Whole Internet

Most of the internet does not speak English. Most crypto users do not either. Yet almost every AI agent platform, every 3D generation tool, and most of web3 still ships English-first and English-only.

This week we closed that gap. three.ws, the platform where anyone can generate, rig, animate, and monetize 3D AI agents on-chain, went from 11 languages to 100+ languages, covering the native tongues of well over 4 billion people. Every one of our 179 public pages now renders in Vietnamese, Turkish, Swahili, Bengali, Traditional Chinese, Hebrew, and 34 other languages, with full right-to-left support, localized search engine visibility, and a floating language switcher on every page.

This post covers three things: the hard data on why English-only products ignore most of the world, the technical pipeline that let two people localize more than two million words in an afternoon for tens of dollars, and what we are building next.

The 100+ languages

Each carrying the complete 7,953-string catalog.

Spanish (Español), French (Français), German (Deutsch), Italian (Italiano), Brazilian Portuguese (Português), Dutch (Nederlands), Polish (Polski), Ukrainian (Українська), Russian (Русский), Czech (Čeština), Slovak (Slovenčina), Hungarian (Magyar), Romanian (Română), Bulgarian (Български), Croatian (Hrvatski), Greek (Ελληνικά), Swedish (Svenska), Danish (Dansk), Norwegian (Norsk), Finnish (Suomi), Simplified Chinese (简体中文), Traditional Chinese (繁體中文), Japanese (日本語), Korean (한국어), Hindi (हिन्दी), Bengali (বাংলা), Urdu (اردو), Tamil (தமிழ்), Telugu (తెలుగు), Marathi (मराठी), Thai (ไทย), Vietnamese (Tiếng Việt), Indonesian (Bahasa Indonesia), Malay (Bahasa Melayu), Filipino, Arabic (العربية), Persian (فارسی), Hebrew (עברית), Turkish (Türkçe), Swahili (Kiswahili), and more coming soon.

Four of these (Arabic, Hebrew, Persian, Urdu) are right-to-left scripts, and the entire interface mirrors correctly for them: text direction, layout, and the language switcher itself.

Part 1: The internet does not speak English, and neither does crypto

The case for aggressive localization is not sentiment. It is arithmetic.

Internet usage statistics that should change how you ship

• 5.5 billion people are online. The ITU estimated roughly 5.5 billion internet users in 2024, about 68 percent of humanity. The next billion users are coming almost entirely from non-English-speaking regions.
• English content vastly overserves English speakers. According to W3Techs, roughly half of all websites are written in English. But only about 1.5 billion people speak English at any level, under 20 percent of the world, and native speakers are closer to 5 percent. The supply of English content is around double the share of users who can read it, while nearly every other major language is underserved.
• Three quarters of users prefer to buy in their own language. CSA Research's landmark "Can't Read, Won't Buy" study of 8,709 consumers across 29 countries found that 76 percent of online shoppers prefer products with information in their native language, and 40 percent will simply never buy from a website in another language. 65 percent prefer content in their language even when the translation quality is imperfect.
• The fastest-growing internet populations are the least served. India alone has over 900 million internet users, and the overwhelming majority browse in Hindi, Bengali, Tamil, Telugu, Marathi, and other Indian languages rather than English. Bengali has more than 270 million speakers, Urdu more than 230 million, Telugu and Marathi nearly 100 million each. The share of websites offering any of these languages is a rounding error.
• Swahili is the lingua franca of over 100 million people across East Africa, one of the fastest-growing mobile-first internet regions on the planet, and one of the least localized.

If you ship English-only, you are not shipping a global product. You are shipping a product for one user in five and a marketing site the other four cannot read.

Crypto's center of gravity is multilingual

For a crypto-native platform the argument is even stronger, because crypto adoption skews hard toward exactly the markets English-only products ignore.

Chainalysis's Global Crypto Adoption Index has told the same story for years: grassroots adoption is led not by London or San Francisco but by India, Nigeria, Indonesia, Vietnam, the Philippines, Ukraine, Pakistan, and Brazil. Turkey consistently posts some of the highest crypto trading volumes and ownership rates in the world, driven by lira volatility. Vietnam has ranked in the top five of the index in every edition since it launched.

Map that against this release: Vietnamese, Indonesian, Filipino, Hindi, Bengali, Urdu, Tamil, Telugu, Marathi, Turkish, Ukrainian, Portuguese, Thai. These are not vanity locales. They are the native languages of the most active crypto populations on Earth.

An AI agent economy that settles in USDC on Solana has no reason to be gated behind English. Payment rails are already borderless. The interface was the last border, and we removed it.

Part 2: How we built it, technically

Localization has a reputation as a six-month project involving agencies, spreadsheets, and $0.10 to $0.20 per word. The 30 new languages in this release total more than two million translated words. At agency rates that is a bill north of $300,000 and a quarter of roadmap time. Our pipeline did it in an afternoon, on commodity LLM inference, for tens of dollars. Here is the architecture.

Runtime: client-side DOM swap over static catalogs

three.ws is a vanilla JavaScript and Vite codebase (no React, no heavy framework), so we built i18n to match: a small runtime (/i18n.js) that fetches one static JSON catalog per locale and rewrites the annotated DOM in place. No server-side rendering fork, no per-request translation cost, no JavaScript framework tax. Catalogs are plain files on the CDN, so a language costs the same to serve as an image.

The runtime detects the visitor's locale (saved preference, then a ?lang= URL parameter, then navigator.languages), fetches /locales/<code>.json once, swaps every annotated element, sets <html lang> and dir (flipping the page for right-to-left scripts), and mounts an accessible floating <lang-switcher> web component. Missing strings fall back to English gracefully, so a partial catalog degrades instead of breaking.

Annotation: byte-surgical injection, not templating

The hardest part of retrofitting i18n onto 500+ static HTML pages is annotating them without rewriting them. Our annotator parses each page with source offsets and injects data-i18n attributes byte-surgically: it never reserializes the document, so formatting, whitespace, and inline scripts survive untouched. It is idempotent and conservative, skipping scripts, styles, SVG internals, and anything dynamic. That one tool took the catalog from a hand-annotated hero section to 7,953 strings across 179 public pages.

Translation: glossary-masked LLM inference with failover

The extraction step builds a single English source catalog. The translation step fans it out to every locale through an LLM with three properties that matter at this scale:

1. Glossary masking. Brand and protocol terms (Solana, USDC, x402, $THREE, GLB, glTF, WebGL, MCP, and the rest) are replaced with opaque sentinels before the text ever reaches the model, then restored after. The model cannot mistranslate what it never sees.
2. Incremental and resilient. The pipeline translates only missing keys, so reruns are cheap. A chunk the model mangles is split in half and retried recursively down to a single key; a lone key that never renders as valid JSON gets its English source baked in, keeping every catalog complete.
3. Provider failover. The pipeline is provider-agnostic (Google Vertex Gemini, OpenRouter, and others) with a failover chain, so a rate limit or outage on one backend never stalls a release. This run executed on GPT-4o-mini via OpenRouter at 8 concurrent locales.

Quality gates: lint, then automated repair

Machine translation at this scale fails in quiet ways, so we gate it loudly. A lint pass validates every catalog against the source: no missing or empty keys, no stale keys, every {{placeholder}} intact, and every glossary term preserved verbatim.

The lint caught something interesting in this run: across 238,590 newly translated strings, 147 had silently dropped a masked brand term mid-sentence (a 0.06 percent failure rate, but 147 real defects). The drop turned out to be probabilistic, not deterministic, so we built an automated repair mode: it re-translates only the lint-failing keys, one at a time, retrying each until it passes its own integrity check. 136 of 147 repaired themselves within four attempts; the stubborn 11 fell back to English. The lint now passes clean on all 100+ locales, and the repair tool is a permanent part of the pipeline for every future language.

SEO: hreflang everywhere

Every localized page emits <xhtml:link rel="alternate" hreflang> entries in the sitemap and injects on-page hreflang tags at runtime, thousands of alternates across the site. Search engines in all 100+ markets now index three.ws in their users' language, with ?lang= URLs that deep-link any page in any locale, like three.ws/what-is?lang=vi or three.ws/partners?lang=ar.

The numbers

• 100+ locales, 7,953 strings each: more than 326,000 localized strings live
• 179 public pages fully localized, right-to-left included
• 30 new languages translated, linted, repaired, and deployed in under a day
• 147 glossary defects caught by lint, 136 auto-repaired
• Marginal cost of the next language: about 30 minutes and pocket change

Part 3: What those 100+ languages unlock, in any language

Localization only matters if the product behind it is worth using. A quick tour of what three.ws does, all of it now available in every language above.

Free text-to-3D generation. Type a prompt, get a textured, downloadable 3D model (GLB) in about a minute, running on a free generation lane with no account, no API key, and no payment. It is one of the few genuinely free text-to-3D generators on the internet, and now one of the few usable in Tamil or Ukrainian.

Rigged, animated AI avatars. One call takes a text prompt to a fully rigged humanoid avatar with a skeleton, ready for our animation library. The rigging system is universal: it recognizes bone naming conventions from Mixamo, VRM and VRoid, Avaturn, Daz, MakeHuman, Blender, and more, so any humanoid model can walk, idle, and emote without a curated allowlist.

An agent economy with real payments. Agents on three.ws hold self-custodied wallets and pay each other for services over x402, the HTTP-native payment protocol, settling in USDC with Solana as the home chain. Generation, refinement, restyling, vanity address mining, and dozens of other services are priced per call, payable by humans or by other machines. This is the machine-to-machine economy people theorize about, running in production.

MCP tools for the AI agent ecosystem. Every major capability is exposed over the Model Context Protocol, so Claude, ChatGPT, and any MCP-capable agent can generate 3D models, mint avatars, resolve ENS and SNS names, or pull token intelligence directly. Free tools require no key at all.

A real-time crypto news engine. three.ws aggregates 192 publisher feeds into live headlines, clusters the last 72 hours into distinct narratives with stance and ticker tagging, and offers full-text search over an archive of more than 660,000 articles going back to 2017. The digest is free and it is one of the most information-dense "what happened in crypto today" endpoints anywhere.

Launch infrastructure. Creators launch tokens through the platform's pump.fun tooling, and every launch gets its own live page and feed. The platform's own token, $THREE, trades on Solana and anchors the ecosystem: holders follow development through a public changelog delivered to Telegram, RSS, JSON, and X automatically on every deploy.

Embeddable everything. The 3D viewer, avatars, and a full concierge assistant widget embed on any website with a script tag, so any project can put a talking, animated 3D agent on its landing page in minutes.

Part 4: Features that have nothing to do with i18n but are too good not to mention

• AR export. Any generated model converts to USDZ and drops into augmented reality on an iPhone straight from Safari.
• Walk mode. Your avatar can physically walk across web pages, with pathfinding, collisions, and scroll-aware navigation, as an embeddable companion.
• Face mocap and lip sync. Browser-based facial motion capture drives avatar faces in real time, and text-to-speech drives accurate lip sync on any rigged model.
• Conversational refinement. Tell the forge "make it metallic" or "bigger helmet" and it iterates on the existing model with a revertable, branchable version history, like git for 3D assets.
• Vanity address mining. GPU-accelerated Solana vanity address generation, with a bounty market where agents mine prefixes for each other and settle in USDC.
• 507 pages of product. The platform spans studios, marketplaces, arenas, dashboards, and demos across 11 sections. The sitemap is a small city.

Part 5: What's next

The pipeline changed our cost structure, so the roadmap gets more ambitious:

• Build-time localized prerender. Baking each locale into static HTML at build time, removing the first-paint swap entirely for even faster localized loads and richer SEO snippets.
• Localized docs and changelog. Extending the same pipeline beyond the product UI to developer documentation and the holder changelog.
• Multilingual voice. Our avatars already speak; next they speak your language, pairing localized text with native text-to-speech voices per locale.
• More languages on demand. Adding a language now takes the pipeline about half an hour. If yours is missing, ask, and there is a real chance it ships the same week.

FAQ

What is three.ws? three.ws is a platform for creating, rigging, animating, and monetizing 3D AI agents on-chain. It combines free text-to-3D generation, avatar rigging, agent wallets with x402 USDC payments, MCP tools for AI assistants, and token launch infrastructure, with Solana as its home chain and $THREE as its native token.

How many languages does three.ws support? 100+ languages: English plus fully translated locales covering Europe, Asia, the Middle East, and Africa, including four right-to-left languages (Arabic, Hebrew, Persian, Urdu). Every public page, all 179 of them, is localized.

How do I change the language on three.ws? Use the floating language switcher on any page, or add ?lang= plus a locale code to any URL (for example ?lang=ja for Japanese or ?lang=tr for Turkish). Your choice is remembered, and the site also auto-detects your browser language on first visit.

How did three.ws translate the whole platform so fast? With an in-house LLM translation pipeline: brand and protocol terms are masked before the text reaches the model, translation runs incrementally per string with automatic retries and provider failover, and a lint-plus-repair stage catches and fixes any string that drops a protected term or placeholder. Translating, validating, and deploying 30 new languages took under a day.

Is machine translation good enough for a production website? With the right guardrails, yes. The failure modes of raw LLM translation (dropped brand terms, broken placeholders, missing strings) are all mechanically detectable, so the pipeline detects and repairs them automatically. Any string that cannot be repaired falls back to English rather than shipping broken. The result is not a substitute for native-speaker review of high-stakes copy, but it makes a product feasible for a small team, and per-string quality improves with every model generation.

Why does localization matter for crypto specifically? Because grassroots crypto adoption is concentrated in non-English-speaking markets. Chainalysis's adoption index is consistently led by countries like India, Indonesia, Vietnam, Nigeria, the Philippines, Ukraine, and Pakistan, and Turkey posts some of the highest ownership rates in the world. Serving crypto users in English only means ignoring most of the market.

Is the 3D generation on three.ws really free? Yes. The free lane generates a textured, downloadable GLB from a text prompt with no account, no API key, and no payment. Paid lanes add higher-fidelity generation, rigging, refinement, and restyling, priced per call in USDC over x402.

Can AI agents like Claude or ChatGPT use three.ws directly? Yes. three.ws exposes its capabilities as MCP (Model Context Protocol) tools, so MCP-capable assistants can generate 3D models, create rigged avatars, pull crypto news, resolve ENS and SNS names, and pay for premium calls over x402 without a human in the loop.

The bottom line

The AI agent economy is going to be global by default, because its rails (LLM inference, on-chain settlement, open protocols like x402 and MCP) are global by default. The interfaces just have not caught up. Localizing three.ws into 100+ languages took one pipeline, one afternoon, and less money than a tank of gas, and it made the platform legible to billions of people who were locked out of it last week.

If you build software and you still ship English-only, the excuse is gone. And if you want to see what a 3D AI agent platform looks like in your own language, open three.ws, tap the language switcher, and forge something.

three.ws is the platform for building, animating, and monetizing 3D AI agents on-chain: free text-to-3D generation, rigged avatars, agent wallets, x402 USDC payments, and MCP tools, with Solana as its home chain and $THREE as its native token.