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.
Sketchfab showcase
The best community-made forge models are published to the official three.ws account on Sketchfab, the largest public 3D model site. Each model page there links back to the creation on three.ws, so the showcase doubles as a discovery channel: someone browsing Sketchfab finds a model, follows the link, and can remix it or forge their own.
This is a curated feed, not a mirror. Only models the community has already validated get pushed, a few per week.
What gets picked
Selection runs three times a week (Mon/Wed/Fri) and takes up to
SKETCHFAB_UPLOADS_PER_RUN models per run (default 2), in this order:
- Weekly Forge-Off winners. The creation crowned by community vote each week on the forge board is the strongest curation signal on the platform and always goes first.
- Top-voted board models. Anything with at least one community upvote, highest votes first.
- Creator-validated models. Creations whose maker explicitly accepted the result or downloaded the GLB, newest first. This tier keeps the showcase alive while board voting ramps up. Raw unreviewed output is never pushed.
A model is skipped when its GLB exceeds the Sketchfab upload cap (45 MB guard), when it was already uploaded, or when it failed three times.
Brand safety. The official account never publishes firearms or explicit
content, regardless of what the forge itself allows. A conservative
word-boundary denylist filters the selection query and re-checks before
upload, and the NemoGuard content classifier in api/_lib/publish-safety.js
runs as a second, fail-open layer. Both are limits on what this account
publishes, not on what users may forge. Blocked creations are parked in the ledger (blocked) and never
re-picked. Refinement children are also excluded from the creator-validated
tier, since their prompts are instructions, not titles.
What an upload looks like
Every published model carries:
- The generation prompt at the top of the description, in quotes. The
model's name is the prompt with a leading article stripped and clamped to
Sketchfab's display limit; a creation whose stored prompt is a placeholder
rather than a description (the image-to-3D path stores its route name,
image-to-3d, as the prompt) is named from its forge category instead ("Forged Sci-Fi"), falling back to "3D Model" for the catch-all buckets, so a winning image-derived model never reaches the account titled after a route. - AI disclosure: Sketchfab's "Created With AI" tag (slug
createdwithai), which its AI-generated content policy requires on AI-generated models, plus a plain statement that the model was AI-generated on the three.ws Forge. Without the tag Sketchfab can file the model as "Human Created" in its search filters. - Backlinks with UTM parameters (
utm_source=sketchfab,utm_medium=referral,utm_campaign=showcase): one to the creation's share page (/forge/share/<id>), one to /forge. The UTM tags make Sketchfab referrals measurable in analytics, which decides whether the cadence goes up or down. - Tags:
createdwithai,ai-generated,generative-ai,text-to-3d,threews, plus the model's category.
Models are published viewable and inspectable but not downloadable: creations belong to their creators, and the showcase does not relicense them.
How it runs
GET /api/cron/sketchfab-showcase (Cloud Scheduler, Bearer $CRON_SECRET):
- Refreshes the async processing status of recent uploads
(
uploadedbecomeslivewhen Sketchfab finishes processing). - Selects candidates and claims each in the
sketchfab_uploadsledger before any network call, so a retried or concurrent run can never double-upload. - Downloads the stored GLB and posts it to the
Sketchfab Data API v3
(
POST /v3/models, multipart).
?dry_run=1 returns the current selection without uploading anything.
Configuration
| Env var | Meaning |
|---|---|
SKETCHFAB_API_TOKEN |
Data API token of the official account (Sketchfab settings, Password & API). Unset: the cron is dormant and skips cleanly. |
SKETCHFAB_UPLOADS_PER_RUN |
Models per run, default 2, clamped 1-5. At the Mon/Wed/Fri schedule the default publishes up to 6 models a week. |
State lives in the sketchfab_uploads table (one row per creation:
pending, uploaded, live, failed with the error recorded, or blocked
when the brand-safety gate parked it). Code:
api/cron/sketchfab-showcase.js
and api/_lib/sketchfab.js.
Backfilling the AI tag on older uploads
New uploads carry createdwithai from 2026-10-09 (commit 919e54e82). Models the account
published before that date do not, and Sketchfab can file them as "Human Created". The backfill
script adds the tag to every model that lacks it and changes nothing else:
eval "$(node scripts/read-service-env.mjs '^SKETCHFAB_API_TOKEN$')"
node scripts/sketchfab-retag.mjs # dry run: prints exactly what would change
node scripts/sketchfab-retag.mjs --apply # PATCH /v3/models/{uid} for each model missing the tag
node scripts/sketchfab-retag.mjs # second dry run: must report 0 to change
Without a token, --user <username> runs a read-only dry run against the public listing. Sketchfab
replaces the whole tag list on PATCH, so each request carries the model's existing tags in order
plus createdwithai, and nothing else. A 429 is retried after Sketchfab's Retry-After.
Tests: tests/sketchfab-retag.test.js.