Agencies.

Building with AI

Run two MCP servers in one session

Two MCP servers in the same session let an agent cross a boundary it otherwise cannot: read a design file on one side, write the matching Prismic model on the other. This guide covers the setup and the one failure mode worth knowing about.

Connect both servers

Each server is declared separately in your agent configuration. Give them distinct names, since tool names collide more often than you would expect once two servers are loaded.

JSON
{
  "mcpServers": {
    "prismic": {
      "url": "https://mcp.prismic.io/mcp"
    },
    "design": {
      "url": "https://mcp.example.com/mcp"
    }
  }
}

Both servers authenticate as you. An agent with write access to a production repository can publish, so point it at a sandbox while you are still testing.

Give it one instruction, not two

The common mistake is asking for the design first and the model second, in separate turns. The agent then summarises the design in its own words and builds from the summary rather than the source. Ask for the whole crossing in a single prompt.

Prompt

Read the Pricing Table frame in the design file, then create a matching shared slice in Prismic. Use a group for the plan rows rather than repeated fields, and tell me which design properties you could not express in the model.

Where it breaks

Design tools describe layout; Prismic models describe content. Anything purely visual has no home in the model, and an agent left to its own devices will invent fields to carry it. Naming that gap in the prompt is what keeps the output clean.

Worth checking by hand

  • Colour and spacing values leaking into the model as text fields
  • Repeated rows modelled as field_1, field_2, field_3
  • Image fields with no constraints