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.
data/
The data files that drive the site. Most of them are hand-edited JSON that build scripts turn into public pages, feeds, and manifests. A few are machine-written ledgers and results that scripts append to. Nothing in here is served raw to users except through the consumers listed below.
Two files matter most:
pages.jsonis the manifest of every public page. It feeds the sitemap,llms.txt,features.json, the human/sitemappage, and the changelog (a page'saddeddate auto-generates its launch entry).changelog.jsonis the curated changelog.npm run build:pagesmerges both files intoCHANGELOG.md,public/changelog.json, andpublic/changelog.xml, and fails the build on a malformed entry.
Top-level files
| File | Purpose | Read by | Written by | Regenerate / validate |
|---|---|---|---|---|
pages.json |
Manifest of every public page (site metadata plus 10 sections of page entries: path, title, description, priority, changefreq, added date). Source of truth for the sitemap, llms.txt, llms-full.txt, features.json, and page-launch changelog entries. |
scripts/build-page-index.mjs, api/sitemap/[type].js (dynamic sitemap), api/page-og.js, scripts/check-pages.mjs, scripts/audit-page-index.mjs, npm run smoke:prod |
Hand-edited | npm run build:pages regenerates all outputs; node scripts/audit-page-index.mjs --strict fails on public routes missing from the manifest (runs in prebuild) |
changelog.json |
Curated changelog entries, newest first: everything holders care about that is not a new page launch. The editorial layer merged with page added dates. |
scripts/build-page-index.mjs, scripts/changelog-telegram.mjs, scripts/changelog-x.mjs, api/_lib/changelog-push.js (via the baked public/changelog.json) |
Hand-edited (append) | npm run build:pages validates every entry (date format, title, summary, allowed tags) and rebuilds CHANGELOG.md, public/changelog.json, public/changelog.xml. npm run changelog:gaps finds shipped work that never got an entry |
changelog-telegram-state.json |
Posted-state for the manual Telegram push script (a posted array of date:title keys). The production cron keeps its own state in Postgres (app_settings); it seeds from this file on first run. |
scripts/changelog-telegram.mjs, api/_lib/changelog-push.js (seed only) |
scripts/changelog-telegram.mjs |
Do not run npm run changelog:push for routine releases: delivery is automatic via /api/cron/changelog-push (Cloud Scheduler) and a manual push can double-post. Use node scripts/changelog-telegram.mjs --dry-run for previews and owner-directed backfills only |
changelog-x-state.json |
Same as above for the X (@trythreews) release thread. | scripts/changelog-x.mjs, api/_lib/changelog-push.js (seed only) |
scripts/changelog-x.mjs |
Same rule: cron-delivered in production; node scripts/changelog-x.mjs --dry-run for previews only |
agent-identities.json |
Agent Identity Studio demo identities behind /agent-identities. Each entry is a real run of the production identity pipeline: inputs plus pipeline outputs (PFP renders, full-body renders, backend used). |
src/agent-identities.js |
scripts/okx-identity-demo.mjs writes results back; never edited by hand |
node scripts/okx-identity-demo.mjs |
timeline.json |
Events for the interactive 3D history timeline at /timeline: 63 chronological milestones with category, importance (1 to 5, drives visual scale), and source URLs. |
src/timeline.js |
Hand-edited | None (edit and reload /timeline) |
walk-social.json |
Quote cards for the Walk landing page social strip. | pages/walk-landing.html |
Hand-edited | None |
robinhood-stock-tokens.json |
Registry of the 95 Robinhood Chain Stock Tokens (chain 4663): address, symbol, name, decimals, price feed. Generated on-chain during the Wave-1 SDK build (shared beacon slot, per-token multicall, feed latestRoundData). |
api/_lib/robinhood.js (stockRegistry(), loaded once at module init) |
Generated from chain; the refresh logic lives in robinhood/robinhood-chain-sdk/scripts/refresh-registry.mjs (writes the SDK copy; sync this file from it) |
node robinhood/robinhood-chain-sdk/scripts/refresh-registry.mjs |
erc8004-bsc-mint-ledger.json |
Append-only ledger of ERC-8004 agent identity mints on BSC (chain 56): agentId, owner, agentURI, txHash, gas cost, timestamp. | Audit trail; read by hand when reconciling mints | scripts/erc8004-mint-bsc.mjs |
node scripts/erc8004-mint-bsc.mjs |
x402-directory-registrations.json |
Per-directory state of our x402 service listings (402index and friends): last response, status, timestamp per endpoint URL. Re-run until every listing reports ok. | scripts/x402-register-directories.mjs (as its resume state) |
scripts/x402-register-directories.mjs |
node scripts/x402-register-directories.mjs |
Subdirectories
| Directory | Purpose | Read by | Written by | Regenerate / validate |
|---|---|---|---|---|
rss/ |
Curated news items. items.json (107 items) controls the RSS feed at /rss/announcements.xml AND one permalink page per item at /news/<slug>. SCHEMA.md documents every field; read it before editing. |
api/_lib/rss-feed.js, api/rss/announcements.js, scripts/build-news.mjs |
Hand-edited (items.json) |
node scripts/build-news.mjs regenerates public/news/ and data/_generated/news-routes.json (also runs in prebuild) |
skills/ |
Skill content for the skills marketplace. seed.json (115 skills, each with category and full SKILL.md content) seeds the marketplace_skills table. The category directories (analysis/, community/, defi/, development/, general/, news/, portfolio/, protocol/, security/, trading/) hold the source SKILL.md packs, one per skill. metamask-agent-wallet/ and metamask-agent-workflows/ are vendored partner skill packs (real directories, not symlinks: scripts/audit-deploy-artifacts.mjs fails the build on committed symlinks here). |
scripts/seed-skills.js; runtime skill feeds read data/_generated/skill-metadata.json and data/_generated/local-skill-packs.json instead (see below) |
Hand-edited | npm run seed:skills ingests seed.json into the database |
archives/ |
Scraped X post archives (nichxbt_tweets_2026-05-10.json, trythreews_tweets_2026-05-16.json, trythreews_tweets_2026-06-03.json) plus tweet-metrics.json, an archived launch-week metrics snapshot. The RSS feed uses the archives as fallback and mirror source; rss/items.json was last merged from the 2026-06-03 archive. |
api/_lib/rss-feed.js (archive mode) |
scripts/refresh-tweet-metrics.mjs produced the metrics snapshot; tweet archives are one-time imports |
None (append new archive files; then merge into rss/items.json by hand) |
quality-bench/ |
The forge realism benchmark: prompts.json (fixed 23-prompt set; never edit an existing id, add new ones), refs/ (3 CC0 reference photos for image-to-3D cases, sources in refs/SOURCES.md), and runs/ (one committed JSON per bench run, append-only, created on first run). Full docs in quality-bench/README.md. |
api/quality-bench.js (dashboard), api/cron/quality-bench.js (weekly regression smoke test, diffs against the latest committed run) |
scripts/quality-bench.mjs writes run files |
node scripts/quality-bench.mjs (see --dry-run, --lane, --compare=latest,previous) |
_generated/ |
Prebuild artifacts, gitignored except one file. news-routes.json (from scripts/build-news.mjs), skill-metadata.json (from scripts/build-skill-metadata.mjs), and local-skill-packs.json (from scripts/build-local-skill-packs.mjs) keep heavy imports out of the deployed functions (api/chat-skills.js, api/skills-manifest.js, src/skills/local-packs.js read them). fact-check-benchmark.json is the exception: a committed, manually-run accuracy report that /fact-check renders via api/fact-check-benchmark.js. |
scripts/build-page-index.mjs, api/chat-skills.js, api/skills-manifest.js, src/skills/local-packs.js, api/fact-check-benchmark.js |
The three build scripts above (automatic in prebuild); scripts/fact-check-benchmark.mjs for the committed report (real chain, real cost per run) |
npm run build regenerates the prebuild trio; node scripts/fact-check-benchmark.mjs re-runs the accuracy report |
Adding a page
- Build the page and wire its route.
- Add an entry to the right section of
pages.json:
{
"path": "/my-page",
"title": "My Page",
"description": "One sentence a stranger could act on.",
"priority": 0.7,
"changefreq": "weekly",
"added": "2026-07-28"
}
Required: path, title, description, added (YYYY-MM-DD). scripts/audit-page-index.mjs --strict (part of prebuild) fails the build if a public route is missing from the manifest. Optional fields used elsewhere in the file: priority and changefreq (sitemap), showcase, auth: "required", indexable: false, tags.
- Run
npm run build:pages. Theaddeddate auto-generates the page's changelog launch entry; nochangelog.jsonentry is needed for a new page.
Adding a changelog entry
For everything users would notice that is not a new page (features, improvements, fixes, SDK releases, security work), append to the top of entries in changelog.json:
{
"date": "2026-07-28",
"title": "Holder-readable headline",
"summary": "Plain-language explanation of what changed and why it matters. No commit jargon.",
"tags": ["improvement"],
"link": "/some-live-path"
}
Required: date (must match YYYY-MM-DD), title, summary, and tags (non-empty array). Allowed tags, enforced by the validator in scripts/build-page-index.mjs: feature, improvement, fix, sdk, infra, docs, security. link is optional and must be a live three.ws path.
Then run npm run build:pages. It validates the entry (a malformed one fails the build) and regenerates CHANGELOG.md, public/changelog.json, and public/changelog.xml.
Delivery to holders is automatic. /api/cron/changelog-push (Cloud Scheduler, every 20 minutes) reads the feed baked into the running image, diffs it against database state in app_settings, and posts anything new to the holders' Telegram channel (@three_ws) and the @trythreews X release thread. An entry can only be announced after the deploy that ships it is live. Do not run npm run changelog:push or npm run changelog:push:x for routine releases: their file state (changelog-telegram-state.json, changelog-x-state.json) is separate from the cron's database state, so a manual push double-posts. Those scripts are for --dry-run previews and owner-directed backfills only.