Crafft — AI Creative Studio with Claude MCP

Source: crafft.ai homepage, crafft.ai integrations/mcp, docs.crafft.ai welcome, docs.crafft.ai MCP integration

Crafft is a Public-Beta “AI Creative Studio” — a node-based canvas that wires multi-modal AI models (image, video, voice, music) into one production workflow, plus a Claude MCP connector that lets Claude plan workflows, write prompts, and trigger generations from chat. Self-positioned as “AI Ad Creative Factory” — production for ad creatives, product demos, UGC content, image variations, voiceover, music — and explicitly contrasted with “prompt box and a prayer” generic AI tools. Sits at the intersection of AI ad/content production (closest cousins: HeyGen Studio Automation, Higgsfield ad-agency workflow, Hyperframes) and the AI-web-design / spec-prompt thread because it formalizes a Claude-driven creative pipeline as the “Foreman” pattern.

Key Takeaways

• The Factory metaphor is the product. Three-pillar architecture per the docs: (1) Assembly Line / Canvas — node-based visual canvas where each node is an AI generation step, output of one feeding the next; (2) Blueprint / Templates — save a workflow once, run it with new inputs to spin up unlimited variations; (3) Foreman / Claude via MCP — connect Claude and let it build the canvas for you from natural-language requests.
• Claude MCP setup is genuinely 30 seconds and uses the claude.ai Connectors UI. Add a custom connector in claude.ai Settings → Connectors (or Customize Connectors bottom-left), name it “Crafft,” paste https://mcp.crafft.ai/mcp, sign in. Once added, Claude can create projects, build & wire workflows, run generations or re-run specific nodes, and execute saved templates with new inputs — all from chat.
• Multi-modal model coverage in one subscription. Per the homepage banner: Seedance 2.0 + GPT Image 2 highlighted as live; per FAQ: Kling and Veo for video, Flux and Seedance for images, Minimax for voice. New models added regularly. Universal credit currency across all models — different models cost different credit amounts based on compute. Unused credits carry over.
• **Three pricing tiers, all $0/monthduringPublicBeta.\ast \ast Starter(550credits/mo),Solo(1,200credits/mo,"Popular"),Agencies(6,100credits/mo,withteamworkspaces,prioritysupport,credit-packtop-ups,beta-featureearlyaccess).Alltiersshow$0 in the captured pricing block — pricing is not yet surfaced (Public Beta) ^[ambiguous]. Annual discount of 10% advertised but the math is irrelevant while the base is $0.
• Has a co-pilot with brand memory. “It plans the workflow, selects the right models, writes prompts, and generates everything. It has memory — it learns your brand style and preferences over time.” This is the in-app counterpart to the external Claude MCP Foreman.
• Agent Skills (beta) is on the docs map. The Foreman section of the docs has both mcp-integration and a separate skills (beta) page. Explicit alignment with Claude’s official Agent Skills format. Page not extracted in this pass — flagged in Open Questions.
• Production-not-generation framing. The docs-page slogan is “Other tools generate. Crafft produces.” — and the positioning matches: visual logic (“if you can draw a line, you can produce”), agency-quality output, scale without effort (“50 localized variations of an ad as easily as you produce one”).
• Workspace + Team primitives ship. Workspaces & Teams page in docs sidebar — multi-seat collaboration is in scope.

What Claude Can Do via Crafft MCP

Per docs.crafft.ai/the-foreman-automation/mcp-integration, the connector exposes these capabilities:

• Create a project
• Build and wire a workflow on the canvas
• Run a generation or re-run specific nodes
• Execute a saved template with new inputs

Three example prompts from the docs:

1. “Create a new project and generate 3 video variations from this product description.”
2. “Run my ‘Coffee Brand’ template with these 5 different background descriptions.”
3. “Check what’s in my latest project and generate a video from the first image.”

How It Fits Alongside the Other Claude-MCP Creative Stacks

Three other Claude-MCP-connected creative production surfaces are documented in the wiki. Side-by-side:

Surface	Output focus	Canvas?	Claude MCP role	Wiki entry
Crafft	Multi-modal ad creatives (image + video + voice + music chained)	Node-based visual canvas	Builds the canvas for you, runs templates	this article
HeyGen Studio Automation	AI avatar video at scale	No (programmatic via Claude Code orchestration)	Drives the API directly via Claude Code skills	HeyGen Studio Automation
Higgsfield (image-to-video, ad-agency workflow)	Cinematic image-to-video and ad creatives	No (model-by-model via SDK + webhooks)	Skills-based orchestration via Higgsfield MCP	Higgsfield Claude Code Ad-Agency Workflow
Hyperframes	Programmatic motion graphics from HTML	No (HTML/Remotion-style code)	Code-generation via Claude Code	Hyperframes

Crafft’s distinctive bet is the visual node-based canvas as the durable artifact. Workflows are saved, reused, templated, and edited visually — not code. Claude becomes the natural-language layer over a graph the human can also touch directly. Compare to HeyGen Studio Automation, which is code-only orchestration; Higgsfield, which is model-by-model SDK calls; and Hyperframes, which is HTML-first.

For wiki context, this also rhymes with Refero Styles’s same-week strategy: expose a curated stack to Claude via MCP so the agent can search-and-study before it builds. Refero exposes design-token catalogs; Crafft exposes a canvas + multi-modal model fleet.

Implementation

Tool/Service: Crafft (crafft.ai) + Crafft MCP

Pricing: Public Beta — Starter / Solo / Agencies tiers all show $0/Month at fetch time. Free starter credits to explore, no credit card required. Real pricing presumably arrives at GA exit ^[inferred].

Setup (Crafft canvas):

1. Sign up at https://www.crafft.ai/dashboard.
2. Open the canvas, drop in nodes (image / video / voice / music / utility).
3. Wire output → input across nodes; run a workflow.
4. Save winning workflows as Templates; re-run with new inputs to scale.

Setup (Crafft MCP for Claude):

1. In the Crafft Dashboard, go to Settings → Integrations and copy the MCP URL.
2. In claude.ai go to Settings → Connectors (or “Customize Connectors” bottom-left).
3. Add a Custom Connector named Crafft, paste https://mcp.crafft.ai/mcp.
4. Click Add → Connect, sign in with your Crafft account.

Integration notes:

• The MCP is a remote Model Context Protocol endpoint — added via claude.ai’s Connectors UI, not via a Claude Code config file. This means it works in claude.ai chat / Cowork, not (directly) in Claude Code. (Claude Code can connect to remote MCP servers via SSE config; whether Crafft documents that path is not surfaced.)
• Credit consumption is per-generation and per-model — agency-tier users can buy credit packs anytime to top up.
• “Agent Skills (beta)” sub-section exists in the docs but was not extracted in this pass; investigating it should clarify whether Crafft ships installable Claude Skills as well as the MCP.

Try It

1. Create a free Crafft account. Click “Start for free” on the homepage; you get starter credits and full canvas access on the $0 Starter tier.
2. Build one workflow manually first. Wire image-generation → video-motion → voiceover → music for a single ad. This gives you the mental model before Claude takes over.
3. Save it as a template. Promote your one-off canvas into a reusable Blueprint with input variables (e.g., product description, brand color, voiceover script).
4. Connect Claude via MCP. Add the connector at https://mcp.crafft.ai/mcp in claude.ai Settings → Connectors. Test with the docs’ sample prompt: “Run my '' template with these 5 different background descriptions.”
5. Compare outputs to your existing creative stack. If you’re already using HeyGen Studio Automation or Higgsfield’s Claude Code workflow, pick a recent ad you produced there and rebuild it in Crafft’s canvas — same inputs, same models where possible, then judge whether the visual-canvas-plus-MCP combo is faster than code-driven orchestration for your team’s workflow.

Related

• Refero Styles — peer “expose a curated stack to Claude via MCP” pattern, but for design tokens vs creative production
• MotionSites — peer creative-prompt library, builder-agnostic spec-prompts; complementary input source for Crafft’s image/video nodes
• HeyGen Studio Automation — closest-cousin Claude-Code-orchestrated video pipeline (avatar-focused)
• Higgsfield Claude Code Ad-Agency Workflow — closest-cousin Claude-orchestrated ad creative workflow (image-to-video focused)
• Higgsfield MCP — peer ad-creative MCP server
• Hyperframes — peer Claude-driven motion-graphics pipeline (HTML-based instead of canvas)
• Figma MCP — peer remote MCP for design-context (claude.ai-Connectors-style)
• Essential MCP Servers — broader MCP catalog
• Agent Skills overview — Crafft docs ships a Skills (beta) sub-section
• Skills + MCP synergy — Crafft uses both surfaces
• AI Marketing — the “AI ad creative factory” framing slots cleanly into ad-creative workflows alongside Outcome Kit

Open Questions

• GA pricing. All tiers show $0 during Public Beta. What does Solo and Agencies actually cost at GA? Is there a transparent credit-to-USD conversion, or does the agency tier shift to enterprise pricing?
• Agent Skills (beta) page. What do Crafft’s Skills look like? Do they install as standard Claude Code skills, or is the surface specific to claude.ai Skills (Agent Skills format)? Does this overlap or complement the MCP, and how?
• Claude Code vs claude.ai support. The MCP setup walks through claude.ai Connectors. Does Crafft document a Claude Code (config-file) integration too? Worth confirming for users who live in the Code surface.
• Model coverage scope. Homepage names Seedance 2.0 + GPT Image 2; FAQ adds Kling, Veo, Flux, Minimax. Full catalog (incl. Sora, Runway, Suno, etc.) and their per-credit cost is not surfaced.
• Workspace pricing. Agencies tier mentions team workspaces — seat caps, per-seat pricing at GA, SSO, and audit-log support not documented in extract.
• Brand-memory portability. Co-pilot “learns your brand style and preferences over time” — is this exportable, and does it expose the same brand context to the MCP-connected Claude session, or are they separate memories?
• Export and ownership. Standard but worth flagging — what’s the license on outputs, especially for Agencies-tier multi-seat workspaces and resold creatives.