- app/domains/adstudio/images.py: trimmed excerpt of the private production file showing phrase_from_title + the SAM cutout crop math actually wired into a feature, not just the bare adapter (storage/fetch layer removed, nothing else simplified). 5 new tests, 33 passed/2 skipped total. - README.md: links github.com/caelum0x/menivor, the real MIT-licensed OSS core Menivor is built on (distinct from this extraction and from the private commercial monorepo). - Fresh 2026-11-02 proof artifacts (real Muse Image call made today). - Paste-ready final-submission form answers. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Menivor × Meta Model API — hackathon submission
Menivor is a live, production AI video-ad/UGC generation platform (script → TTS → avatar → captions → compose → publish). Its production codebase is a private monorepo (70-table Postgres schema, real billing, four client surfaces) — too large and too commercially sensitive to publish wholesale.
This repository is the real, runnable extraction of the specific code built for the Meta Global AI Developer Hackathon (hosted with Cerebral Valley): two new provider adapters that wire Meta's Model API (Muse Image, SAM 3.1) into Menivor's existing provider-routing architecture, plus their real test suite.
What's in this repo
app/domains/routing/providers/muse_image.py— real adapter for Muse Image (POST /v1/images/generationsand/v1/images/edits, OpenAI-Images-API compatible, $0.01/image).app/domains/routing/providers/sam_segment.py— real adapter for SAM 3.1 (POST /v1/responses, decodes the documented<|box;...|>detection grammar into bounding boxes for a product-photo cutout).app/domains/routing/base.py— the genericProviderAdapterprotocol +ProviderRegistrythese two adapters implement (copied verbatim from the private repo; it is intentionally generic and carries no business data).app/domains/routing/providers/kling_fal.py— not the real Kling/fal.ai video adapter (that stays private and is unrelated to this submission); this file is trimmed down to just the sharedKlingErrortaxonomy class every adapter in the real package — including these two — imports and raises.tests/— the real pytest suite for both adapters, copied verbatim except for two tests that exercise the full private provider registry (skipped here with a documented reason; they pass unmodified in the real codebase).SUBMISSION.md— the full hackathon application/submission writeup (track, project idea, technical skill, links).DEMO-SCRIPT.md— the shot-by-shot script for the required 3-minute demo video.
Why this design
Both adapters fail closed with a clearly-labeled KlingError(kind="auth") when
MODEL_API_KEY is unset, and this repo intentionally never commits a real key
(see HACKATHON.md for the current status of real calls/proof artifacts).
Nothing here is mocked or faked: the HTTP protocol, request/response shapes,
pricing, and error taxonomy are all the real, documented Meta Model API wire
format. The moment a real key is set, muse_generate_image() and
sam_segment_image() make real calls.
See HACKATHON.md for the exact commit-range boundary
between pre-existing work and this hackathon's Oct 27 – Nov 4 build window.
Verify it yourself
python3 -m venv .venv && source .venv/bin/activate
pip install requests pytest
python3 -m pytest tests/ -q
# 27 passed, 2 skipped
Hackathon window separation
See HACKATHON.md — it pins the exact git tag boundary
(pre-hackathon-window) between what predates the Oct 27–Nov 4 build window
(this repo's two initial commits, both 2026-10-26) and what's genuinely built
during it, and tracks window-work as it lands.
What's in this repo (continued)
app/domains/adstudio/images.py— the real SAM 3.1 integration call site (phrase_from_title, the cutout crop math), not just the bare adapter — trimmed from the private production file (storage/fetch layer removed, nothing simplified). Shows the adapter actually wired into a feature.
Links
- Live product: https://menivor.com
- Open-source core (real, separately maintained): github.com/caelum0x/menivor — the FastAPI pipeline + Next.js/Expo/Rust clients Menivor is built on top of (MIT, built on MoneyPrinterTurbo). This is a genuinely different repo from this hackathon extraction: it's the general screen-recording → captioned UGC/avatar video core, MIT-licensed and runnable standalone, whereas this repo is specifically the two new Meta Model API adapters plus their real call site. Neither this repo nor the OSS core expose the private commercial monorepo's billing/multi-tenant/schema internals.
- Production repository (private, commercial): github.com/caelum0x/videogen