AssetsCurator

AssetsCurator is an app built for automatically collecting assets like images,videos,audio voiceover and background music,based on video script and arranging them in downloadable zip folder by scene

Comment

Inspiration

Assembling a shot list, mood board, or pre-viz pack for a short video means juggling a dozen tabs — an AI image tool, two or three stock sites, a YouTube search, a text-to-speech page, a royalty-free music library — and then renaming every file into numbered folders by hand. Agents are great at the fetching, but curation is visual and iterative: a human has to see the board while it fills up and say "no, swap the hero of scene 2."

WebMCP made a different shape possible. Instead of bolting a chatbot onto a website, the website exposes tools and ChatGPT itself becomes the operator of a live canvas. That reframed the whole product: no in-app chat, no copy-paste — the human watches the board, the agent fills it.

What it does

AssetsCurator is a WebMCP media bin. Open the board in ChatGPT's in-app browser and ChatGPT gains ~23 registered tools to build a shot list scene by scene:

  • Generate AI stills, short clips, and audio (voiceover, score, SFX)
  • Search & attach royalty-free photos, footage, and music
  • Pin YouTube references (links and embeds only, never downloaded)
  • Surgically edit — replace a single asset in place, reorder scenes, set roles
  • Export a shot-ordered folder (01-open/, 02-pour/ …) with audio files, attribution, and .url shortcuts

Every write tool updates the visible UI before it returns, and the whole board works in 16:9, 9:16, or 1:1. With no agent present, the same actions are plain buttons — progressive enhancement.

How we built it

  • Next.js 15 (App Router), TypeScript, Tailwind v4.
  • Tools register on the top-level document via document.modelContext.registerTool().
  • A Zustand store is the single shared source of truth (persisted to IndexedDB). The agent's tools and the human's buttons call the same action layer, so their views can never diverge.
  • API routes proxy keys and run a free-first fallback router (firstOk() + a circuit breaker): each modality has a provider chain that cools a failed source and tries the next, so the board is never empty.
  • A cinematic prompt layer + higher generation resolution lift AI output from "AI default" to production-looking frames.
  • JSZip assembles the shot-ordered export.

Challenges we ran into

  • Top-level registration only. WebMCP hosts discover tools registered on the top-level page; iframe and declarative-form approaches didn't surface, which shaped the whole client architecture.
  • Free providers are fragile. During the build, HuggingFace retired FLUX from its free serverless inference (HTTP 410), Together's free tier began requiring a paid deposit, and a keyed Pollinations endpoint ran out of credit (402). Rather than chase one "good" provider, we leaned into the circuit-breaker + fallback design — the fragility became resilience.
  • Vertical video. Most free image generators emit a fixed square frame; only Pollinations honors arbitrary width/height, so it anchors the 9:16 / 1:1 feature end to end (generation, stock orientation, and display).
  • A silent zero-result bug. A Jamendo include parameter returned an empty (but "successful") result set and masqueraded as a provider failure until we traced it.

Accomplishments that we're proud of

  • A genuinely agent-native UX: no chatbot, the page is the canvas, and single-asset replace_asset proves the model is doing typed mutation on shared state, not DOM scraping.
  • Runs end to end on free providers with graceful degradation — a judge with zero keys still gets a full board.
  • Conflict-free human + agent co-editing through one optimistic store.
  • One-call export that drops straight into CapCut or Premiere.

What we learned

  • WebMCP rewards designing tools as verbs over shared state, where every write returns the new project so the agent can verify ids and self-correct.
  • Optimistic UI + a single store is what makes human-and-agent co-editing feel live.
  • Fallback architecture matters more than any single provider. Assume every free API will fail, quota out, or get deprecated mid-hackathon — and design for it.

What's next for AssetsCurator

  • Sequence/timeline export (EDL / Premiere XML) and per-shot duration trims.
  • More aspect-native video and a Ken Burns motion pass for stills.
  • Collaborative, multi-viewer boards.
  • Optional paid-tier providers behind the same router for higher-fidelity generation, with the free chain as the always-on floor.

Built With

Share this project:

Updates