Engineering a Production-Ready MCP Server with Prisma 7 and PostgreSQL

In the shift toward agentic AI, the Model Context Protocol (MCP) has emerged as the industry standard for connecting LLMs to local and remote data sources. As a senior developer, you shouldn't just be building "wrappers"; you should be building robust, type-safe interfaces that allow an AI to reason over your business logic.

This guide covers the architecture and implementation of a Kanban management MCP server, utilizing the latest Prisma 7 features and a PostgreSQL backend.

Technical Debt Alert: Prisma 7 introduced breaking changes regarding how database connections are handled. If you are following tutorials from 2024, your implementation will likely fail. We will use the new Driver Adapter pattern specifically for this reason.

The Repository

The full source code for this implementation is available at: github.com/AlNahean/prisma-kanban-mcp-server. I recommend cloning it to follow along with the instructions below.

Technical Setup & Instructions

Initialize the Environment

Modern Node.js development requires ECMAScript Modules (ESM). Ensure your package.json includes "type": "module". This enables top-level await and cleaner imports.

mkdir kanban-mcp && cd kanban-mcp
npm init -y
# Install core dependencies
npm install @modelcontextprotocol/sdk @prisma/client @prisma/adapter-pg pg dotenv
# Install dev tools
npm install -D typescript @types/node @types/pg prisma tsx

Prisma 7 Migration & Configuration

In Prisma 7, the url property is no longer supported directly in the schema.prisma file for most environments. You must move your configuration to a prisma.config.ts file.

import "dotenv/config";
import { defineConfig, env } from "prisma/config";
 
export default defineConfig({
  schema: "prisma/schema.prisma",
  datasource: {
    url: env("DATABASE_URL"),
  },
});

generator client {
  provider        = "prisma-client-js"
  previewFeatures = ["driverAdapters"]
}
 
datasource db {
  provider = "postgresql"
}
 
model Column {
  id    String @id @default(cuid())
  title String
  tasks Task[]
}
 
model Task {
  id       String @id @default(cuid())
  title    String
  columnId String
  column   Column @relation(fields: [columnId], references: [id])
}

Implementing the Server Architecture

The MCP server operates over Standard I/O. We initialize the PostgreSQL pool using the new adapter pattern before connecting the server.

import { Server } from "@modelcontextprotocol/sdk/server/index.js";
import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
import { PrismaClient } from "@prisma/client";
import { PrismaPg } from "@prisma/adapter-pg";
import pg from "pg";
 
const pool = new pg.Pool({ connectionString: process.env.DATABASE_URL });
const adapter = new PrismaPg(pool);
const prisma = new PrismaClient({ adapter });
 
const server = new Server({
  name: "kanban-manager",
  version: "1.0.0"
}, {
  capabilities: { tools: {} }
});

Exposing Tools to the LLM

A senior-level MCP implementation requires descriptive schemas. The AI uses these descriptions to decide which tool to call. If your descriptions are poor, the agent will hallucinate.

server.setRequestHandler(ListToolsRequestSchema, async () => ({
  tools: [
    {
      name: "get_board",
      description: "Retrieve the full Kanban board state, including all columns and tasks.",
      inputSchema: { type: "object", properties: {} }
    },
    {
      name: "move_task",
      description: "Move a task to a different column by ID.",
      inputSchema: {
        type: "object",
        properties: {
          taskId: { type: "string" },
          columnId: { type: "string" }
        },
        required: ["taskId", "columnId"]
      }
    }
  ]
}));

Pitfalls and Critical Issues

1. The `console.log` Corruption

MCP servers communicate via stdout. If you leave a console.log("Connected to DB") in your code, that string will be injected into the JSON-RPC stream. The AI client (like Claude Desktop or Antigravity) will fail to parse the response, throwing an "Unexpected token" error. Fix: Always use console.error() for internal logging; it uses stderr, which the AI client ignores but displays in debug logs.

2. EPERM Errors on Windows

When running npx prisma generate, you may encounter EPERM: operation not permitted. This happens because a process (like the MCP client or a background Node server) is holding a lock on the .prisma/client engine file. Fix: Shut down your MCP client (Claude Desktop/Antigravity) and any running dev servers before generating the client or running migrations.

3. Serialization of BigInt

If your schema uses BigInt (common for Telegram IDs or snowflake IDs), native JSON.stringify will crash. Fix: Implement a replacer function for your tool results:

const result = JSON.stringify(data, (key, value) =>
  typeof value === "bigint" ? value.toString() : value
);

4. Prisma 7 Driver Adapters

Prisma 7 removed the internal Rust-based engine for database connectivity. If you initialize new PrismaClient() without an adapter, it will crash. Fix: You must install pg and @prisma/adapter-pg and pass the adapter explicitly during instantiation as shown in Step 3.

Senior Architectural Advice

Security Note: Do not grant the PostgreSQL user used by MCP "Superuser" status. Use a dedicated user with permissions limited to the public schema. If your MCP tools allow data deletion, ensure you have implemented a confirmation flow or restricted those tools to specific agent roles.

Handling Context Windows

Do not return 500 rows of data in a single MCP call. Large payloads will blow out the LLM's context window, increasing latency and cost.

Implement pagination in your List tools.
Return only the IDs and Titles for broad queries; let the AI request "Details" by ID for a specific item.

Conclusion

Building an MCP server is the most efficient way to turn a generalist AI into a specialist for your team's workflow. By following the Prisma 7 adapter pattern and respecting the Standard I/O communication model, you can create a production-grade backend for AI agents.

Check out the full implementation and contribute to the logic over at the repository.

Engineering a Production-Ready MCP Server with Prisma 7 and PostgreSQL

This guide covers the architecture and implementation of a Kanban management MCP server, utilizing the latest Prisma 7 features and a PostgreSQL backend.

The Repository

The full source code for this implementation is available at: github.com/AlNahean/prisma-kanban-mcp-server. I recommend cloning it to follow along with the instructions below.

Technical Setup & Instructions

Initialize the Environment

Modern Node.js development requires ECMAScript Modules (ESM). Ensure your package.json includes "type": "module". This enables top-level await and cleaner imports.

mkdir kanban-mcp && cd kanban-mcp
npm init -y
# Install core dependencies
npm install @modelcontextprotocol/sdk @prisma/client @prisma/adapter-pg pg dotenv
# Install dev tools
npm install -D typescript @types/node @types/pg prisma tsx

Prisma 7 Migration & Configuration

In Prisma 7, the url property is no longer supported directly in the schema.prisma file for most environments. You must move your configuration to a prisma.config.ts file.

import "dotenv/config";
import { defineConfig, env } from "prisma/config";
 
export default defineConfig({
  schema: "prisma/schema.prisma",
  datasource: {
    url: env("DATABASE_URL"),
  },
});

generator client {
  provider        = "prisma-client-js"
  previewFeatures = ["driverAdapters"]
}
 
datasource db {
  provider = "postgresql"
}
 
model Column {
  id    String @id @default(cuid())
  title String
  tasks Task[]
}
 
model Task {
  id       String @id @default(cuid())
  title    String
  columnId String
  column   Column @relation(fields: [columnId], references: [id])
}

Implementing the Server Architecture

The MCP server operates over Standard I/O. We initialize the PostgreSQL pool using the new adapter pattern before connecting the server.

import { Server } from "@modelcontextprotocol/sdk/server/index.js";
import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
import { PrismaClient } from "@prisma/client";
import { PrismaPg } from "@prisma/adapter-pg";
import pg from "pg";
 
const pool = new pg.Pool({ connectionString: process.env.DATABASE_URL });
const adapter = new PrismaPg(pool);
const prisma = new PrismaClient({ adapter });
 
const server = new Server({
  name: "kanban-manager",
  version: "1.0.0"
}, {
  capabilities: { tools: {} }
});

Exposing Tools to the LLM

A senior-level MCP implementation requires descriptive schemas. The AI uses these descriptions to decide which tool to call. If your descriptions are poor, the agent will hallucinate.

server.setRequestHandler(ListToolsRequestSchema, async () => ({
  tools: [
    {
      name: "get_board",
      description: "Retrieve the full Kanban board state, including all columns and tasks.",
      inputSchema: { type: "object", properties: {} }
    },
    {
      name: "move_task",
      description: "Move a task to a different column by ID.",
      inputSchema: {
        type: "object",
        properties: {
          taskId: { type: "string" },
          columnId: { type: "string" }
        },
        required: ["taskId", "columnId"]
      }
    }
  ]
}));

Pitfalls and Critical Issues

1. The `console.log` Corruption

2. EPERM Errors on Windows

3. Serialization of BigInt

If your schema uses BigInt (common for Telegram IDs or snowflake IDs), native JSON.stringify will crash. Fix: Implement a replacer function for your tool results:

const result = JSON.stringify(data, (key, value) =>
  typeof value === "bigint" ? value.toString() : value
);

4. Prisma 7 Driver Adapters

Senior Architectural Advice

Handling Context Windows

Do not return 500 rows of data in a single MCP call. Large payloads will blow out the LLM's context window, increasing latency and cost.

Implement pagination in your List tools.
Return only the IDs and Titles for broad queries; let the AI request "Details" by ID for a specific item.

Conclusion

Check out the full implementation and contribute to the logic over at the repository.

Engineering a Production-Ready MCP Server with Prisma 7 and PostgreSQL

Engineering a Production-Ready MCP Server with Prisma 7 and PostgreSQL

The Repository

Technical Setup & Instructions

Pitfalls and Critical Issues

1. The console.log Corruption

2. EPERM Errors on Windows

3. Serialization of BigInt

4. Prisma 7 Driver Adapters

Senior Architectural Advice

Handling Context Windows

Conclusion

Engineering a Production-Ready MCP Server with Prisma 7 and PostgreSQL

Engineering a Production-Ready MCP Server with Prisma 7 and PostgreSQL

The Repository

Technical Setup & Instructions

Pitfalls and Critical Issues

1. The console.log Corruption

2. EPERM Errors on Windows

3. Serialization of BigInt

4. Prisma 7 Driver Adapters

Senior Architectural Advice

Handling Context Windows

Conclusion

1. The `console.log` Corruption

1. The `console.log` Corruption