Brand exploration at agent speed

A 12-slot Canada Day moodboard for a premium bottled-water brand — explored end-to-end by an agent through the Storyboard MCP in under three minutes.

A campaign creative director has 90 minutes between meetings and wants to see twelve directions for a Canada Day campaign before pitching the launch team tomorrow. The brief: premium bottled spring water, multigenerational Canadian families, soft golden moments — never industrial. Every direction has to be on-brand, but the board has to be diverse enough to actually drive a decision.

The agent flow

The director's agent, holding a single Daydream API key, talked to the Storyboard MCP server. Three tool calls. No prompt engineering. No model selection. No image-asset wrangling.

1. brand_kit_create — name, hex palette, prompt keywords, forbidden terms. One call. Returns ckit_ccd658ae45.
2. moodboard_create — title + intent + the kit id. One call. Returns mood_488b3b9ac3.
3. moodboard_spread — twelve slots, max_cost_usd: $4.50, six style hints rotated across the slots. One call, fans out.

The server auto-rotated four image models across the twelve slots so the board carried real stylistic diversity — not twelve flux-dev renders that all look the same. The palette and the keywords were injected into every prompt automatically because the moodboard was locked to the brand kit. Forbidden terms ("plastic", "industrial", "synthetic") were stripped from every shot.

The board

The agent received a public viewer URL the marketer can share with stakeholders without an account, exporting credentials, or paying for another seat. Live: storyboard.daydream.monster/m/mood_488b3b9ac3

From still board to hero video

Boards drive launch decisions — but campaigns need motion. One additional create_media action=animate call on the strongest slot animates the lead frame into a 5-second commercial-ready hero shot. The agent picked slot 4 (the gemini-image soft daylight portrait), routed to seedance-i2v, and 3.5 minutes later returned a polished pour-and-condensation hero. Cost: ~$0.30.

Pure North hero spot — seedance-i2v, 5s, seed 121909774, animated from slot 4.

What the agent did not have to handle

• Pick which image model to use per slot. The server's capability router did it.
• Inject the brand palette into every prompt. The kit handled it.
• Filter brand-policy words. The forbidden-terms list ran server-side.
• Track cost manually. cost_usd_estimated arrived on every slot's response payload.
• Manage retries on transient fal-side flakes. The 4-cap fallback chain absorbed them.
• Build a viewer page. The MCP auto-published with a 14-day share URL.

How this compares

This is a published-capability comparison — vendor-cited specs and pricing as documented in their respective developer docs.

Capability	Storyboard MCP / Livepeer	Runway	Midjourney	Flora
Multi-model diversity in one call	✓ 4 caps round-robined	Single Gen-4 family	Single Niji/v7 family	Node-by-node, one model per node
Brand-kit palette + keyword auto-inject	✓ brand_kit_create + lock	Manual per-prompt	Manual per-prompt	Manual per-node
Cost transparency before submit	✓ max_cost_usd pre-flight	Credits, no $ pre-check	GPU-min after-the-fact	Credits, no pre-check
Agent-callable (MCP)	✓ Native	HTTP API, no MCP yet	Discord-bot or unofficial API	UI-first, no MCP
Pricing for a 12-slot exploratory board	~$0.50	$1.20–$2.40 (Gen-4 credits)	~$1.00 (Standard tier)	~$0.85 (mixed nodes)
Shareable viewer URL (no login)	✓ Auto-published	In-app only	In-app or download	Public canvas requires sub

The competitive numbers above are from each vendor's public pricing page and quickstart docs as of mid-2026. We did not run the same brief on the other platforms for this case study — those numbers are estimates from their published rates.

Where this would have broken without recent work

The 12 slots all hit on the first pass. Before the Z+AX sprint that landed across May 14–15, several of the same agent flows would have been more painful:

• AX-1 (typed error envelope) — moderation rejections on recraft-v4 for stylization name-brands now return {code: "moderation_blocked", retryable: false, hint: "try gemini-image or describe the aesthetic without name-brands"} instead of an opaque fal error string. Agents branch on code; before, every error looked the same.
• AX-4 (idempotency) — A flaky agent retry of the same moodboard_spread with an idempotency_key now hits the 24-hour cache and returns the same twelve URLs with idempotency_replay: true. No double-spend.
• AX-5 (latency SLA) — list_capabilities now publishes p50/p95 per capability. Agents picking under a wall-clock budget compare against p95 directly.
• AX-6 (URL normalization) — Every URL returned through the canonical helper. Mixed-case host, http→https on known providers, signed tokens preserved verbatim. Agents downstream see uniform URLs.