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.
JavaScript 3d-generationai-agentsanimationavatardigital-humangenerative-aiglbgltfimage-to-3dllmmcpmcp-servermodel-context-protocolsolanatext-to-3dthreejsweb-componentswebglwebxrx402
# Cloud Build upload allowlist (gitignore syntax; last match wins).
#
# The agent-sniper image (deploy/sniper/cloudbuild.yaml) builds from the REPO
# ROOT and its Dockerfile copies the root package manifests, api/, and the
# agent-payments-sdk/ workspace — NOT just workers/. The previous allowlist
# uploaded only workers/, so the sniper Cloud Build context was missing those
# paths and `COPY package.json` / `COPY api` failed. Re-include exactly what the
# worker images need, and keep multi-GB node_modules + .git out of the upload.
/*
!/workers/
!/api/
!/src/
!/agent-payments-sdk/
!/package.json
!/package-lock.json
!/.npmrc
!/.gcloudignore
!/.dockerignore
# --- three-ws-api (Cloud Run) additions -------------------------------------
# The root Dockerfile runs `npm ci` (workspaces) and serves the full site via
# server/index.mjs, so the context needs every workspace dir, the data/ and
# public/ dirs handlers read at runtime, dist/ (the built frontend, served
# from the image), and vercel.json (the route table the server interprets).
# character-studio ships manifest-only (216 MB of frontend source npm ci
# doesn't need); animation-sources (2.6 GB) and dist-lib/ stay out entirely.
# agents/ (fact-checker, tutor, endpoint-shopper, unstoppable, x402-buildout —
# 492K total, no node_modules) was missing from this allowlist entirely, so
# api/x402/fact-check.js and api/x402/tutor.js's `../../agents/...` imports
# 500'd ERR_MODULE_NOT_FOUND on every request in every deployed revision —
# found live in prod during the 2026-07-08 storefront cleanup (prompt 18)
# when the endpoint was re-listed as a real product. Root-caused via Cloud
# Run stderr logs (gcloud logging read), not guessed.
!/server/
!/Dockerfile
!/vercel.json
!/dist/
!/scripts/
!/data/
!/examples/
!/public/
!/pages/
!/sdk/
!/solana-agent-sdk/
!/mcp-server/
!/mcp-bridge/
!/packages/
!/agent-ui-sdk/
!/avatar-sdk/
!/page-agent-sdk/
!/multiplayer/
!/walk-sdk/
!/tour-sdk/
!/agents/
# services/ (644K, no node_modules) holds the long-running side services, and two
# api/** handlers import straight out of it: api/_lib/home/relay.js pulls
# services/home-relay/src/token.js and api/home/satellite.js pulls
# services/home-satellite/src/token.js. It was never re-included, so revision
# 00412 answered /api/healthz (and every home relay route) with
# ERR_MODULE_NOT_FOUND on 2026-09-04, found in live Cloud Run stderr. Exact same
# failure shape as the agents/ entry above, which is why check-gcloudignore.mjs
# now asserts both token modules land in the upload.
!/services/
# .agents/skills (2.2 MB, 210 files) is the skill catalog workers/okx-chat-bot's
# Dockerfile copies with `COPY .agents/skills ./.agents/skills`, so the AI
# subsession answering marketplace buyers boots knowing the task/chat lifecycle
# instead of improvising it. This allowlist is a root exclude plus re-includes,
# and .agents was never re-included, so gcloud stripped it from every upload and
# that COPY failed the build outright: verified 2026-08-11 on build 47b6b6ae
# ("COPY failed: file not found in build context ... stat .agents/skills"). The
# worker's only documented deploy path could not have succeeded. Same failure
# shape as the sniper and agents/ entries above.
!/.agents/
# community-skills/ is the community skills registry. api/_lib/community-skills.js
# reads registry.json and each skills/<slug>/SKILL.md at runtime for
# /api/skills/community, custom-skill imports and the MCP skill tools, and
# api/_lib/agent-custom-skills.js imports its tools/registry.mjs.
!/community-skills/
# tests/ stays out of the image, but tests/fixtures/ holds one runtime artifact:
# the 40-claim fact-check benchmark suite. api/fact-check-benchmark.js reads it
# to render the claim-count summary on /fact-check, and
# api/cron/fact-check-benchmark.js runs it. Without this re-include the endpoint
# answered `fixture: null` in every deployed revision and the scheduled re-run
# could not load its own suite. Re-include the parent, drop everything else under
# it, then re-include just the fixtures dir (236K).
!/tests/
/tests/*
!/tests/fixtures/
!/character-studio/
/character-studio/*
!/character-studio/package.json
# Re-exclude heavy/derived paths even inside the re-included directories.
**/node_modules/
.git/
# Python virtualenvs and caches inside the GPU workers. These are gitignored, so
# a clean `git worktree` deploy (the runbook path) never sees them, but a submit
# straight from a working directory would upload thousands of files: one
# workers/avatar-reconstruction/.venv-eval alone accounts for most of the ~10.5k
# files workers/ contributes. Verify with `npm run check:gcloudignore`.
**/.venv*/
**/__pycache__/
**/*.pyc
# Wrangler's local dev scratch (workers/pump-fun-mcp/.wrangler). Gitignored, but
# .gcloudignore replaces .gitignore rather than extending it, so a submit from a
# working directory uploaded it. Worse than wasted bytes: `wrangler dev` creates
# and deletes files under .wrangler/tmp/ while it runs, so the upload walk hit a
# path that vanished under it and gcloud CRASHED with
# "FileNotFoundError: .wrangler/tmp/dev-XXXXXX" - every Cloud Build submit from
# this repo failing for as long as a dev server happened to be up.
**/.wrangler/
# workers/*/.cache/ holds per-agent keypairs as raw secret-key byte arrays
# (workers/agora-citizens/.cache/*.json). Each worker gitignores its own .cache,
# but .gcloudignore REPLACES .gitignore rather than extending it, so a submit
# from a working directory uploaded live private keys into the Cloud Build
# source bucket. Nothing in any image needs them: the runtime reads its keys
# from Secret Manager. Caught by `npm run check:gcloudignore` on 2026-08-11.
**/.cache/