For Engineering teams building AI agents on Claude Code, Cursor, Windsurf, or custom Agent SDK / OpenAI Agents / LangGraph stacks.

Image and video generation for AI agents

Give Claude, Cursor, and Windsurf five MCP tools that produce brand-locked images, reels, and try-on — callable from any agent task.

Lamina exposes generative-media as five MCP tools — lamina.create, .track, .evaluate, .distribute, .brand_lookup — so an autonomous agent in Claude, Cursor, Windsurf, or Claude Code can generate brand-locked images and video as native tool calls. Run npx @lamina/mcp install once; the agent uses the tools without further wiring.

The problem

Most generative-media APIs were designed for humans clicking a UI. Agents don't have a UI — they have a tool list, a budget, and a non-zero retry rate. Calling a raw model endpoint from an agent fails on long jobs, drifts on brand, and double-bills on retries. You either build the agent-callable layer yourself or your agent quietly burns money producing inconsistent assets.

How Lamina solves it

How it works

The Lamina MCP server registers five typed tools with whichever client the developer uses. Each tool is JSON-Schema typed so the agent sees argument shapes; outputs are structured JSON so the agent can branch on brand-fit score; idempotency keys are generated per agent turn so retries don't double-bill; long jobs stream progress so the agent can update the user without polling.

Why it works for autonomous agents

Agents don't pick models — they pick outcomes. lamina.create takes a creative-app name (productShoot, brandFilm, virtualTryOn) and a brief; the platform handles routing, evaluation, and delivery. The agent's code stays stable across model upgrades, which is critical when the image-model leaderboard reshuffles every quarter.

Steps

  1. 1

    Install the Lamina MCP server

    npx @lamina/mcp install — writes the entry into Claude Desktop, Cursor, or Windsurf config automatically.

  2. 2

    Set your API key

    Add LAMINA_API_KEY to the env block; restart the client so MCP discovery picks up the server.

  3. 3

    Call from any agent task

    Ask the agent: 'Generate a hero for SS26 — sage linen, 16:9, four variants. Push the keeper to Shopify.' The agent will call lamina.create, .evaluate, and .distribute.

  4. 4

    Stream progress to the user

    lamina.track exposes phase events the agent reports back inline. The user sees queued → composing → rendering → done.

What you get

  • •One MCP install gives the agent generate, track, evaluate, distribute, and brand-lookup tools.
  • •Idempotency keys prevent double-billing on retried agent turns.
  • •The same agent code works in Claude Desktop, Cursor, Windsurf, and Claude Code.
  • •Brand consistency is enforced server-side, so the agent doesn't need to manage prompt scaffolding.

Works with