How to Build MCP Servers
2026

Before MCP, connecting an AI model to your data required custom connectors for each source, creating what Anthropic described as an "N×M" integration problem. Each new data source needed its own implementation, making truly connected systems impossible to scale. Earlier attempts like OpenAI's 2023 function-calling API and ChatGPT plugins solved similar problems but required vendor-specific connectors.

MCP was announced by Anthropic in November 2024 as an open standard transported over JSON-RPC 2.0, reusing the message-flow ideas of the Language Server Protocol (LSP). The adoption timeline that followed was unprecedented in protocol history.

The Pento 2025 year-in-review describes MCP's adoption as comparable to USB-C standardisation: before MCP, every AI integration needed its own connector. MCP provides the universal cable. For developers, this means building once and connecting everywhere.

Understanding the architecture is essential before writing code. MCP defines three roles that interact through a standardised protocol.

Three Capability Types

MCP servers provide three capability types. Tools are functions the LLM can call with user approval, like fetching data from an API or running a database query. Resources are file-like data that clients can read, such as API responses, file contents, or database records. Prompts are pre-written templates that help users accomplish specific tasks. Most developers start with Tools since they provide the most immediate value.

Transport Modes

MCP supports two primary transports. STDIO (Standard Input/Output) is used for local servers where the client spawns the server process directly, ideal for Claude Desktop integration. Streamable HTTP is used for remote servers accessible over the network, required for production deployments. The official TypeScript SDK includes middleware packages for Express, Hono, and Node.js HTTP to simplify HTTP transport setup.

You need Node.js version 16 or higher. The official SDK is published as @modelcontextprotocol/sdk and requires zod for schema validation. The SDK uses Zod v4 and runs on Node.js, Bun, and Deno.

Update your package.json to include the module type and build script:

For Python developers, the setup is even simpler: pip install mcp and use FastMCP for rapid development. The Python SDK provides the same capabilities with a Pythonic API.

Here is the minimal working MCP server that registers a tool and connects via STDIO transport. This is the pattern used by thousands of production MCP servers in the ecosystem.

Build with npm run build, then configure Claude Desktop to use your server by editing claude_desktop_config.json:

Restart Claude Desktop and you will see a plug icon indicating your MCP server is connected. When you ask Claude a question that matches your tool's capability, Claude will request permission to call your tool, executing the "human in the loop" consent flow.

Stop writing scaffolding code from scratch. Use this generator to produce a working MCP server boilerplate customised to your use case. Select your data source type and the tool generates the complete index.ts file ready for building.

The official MCP Inspector is the primary debugging tool. It provides a web UI to visually inspect your server's tools, resources, and prompts, and to test tool calls interactively.

The Inspector mirrors the runtime behaviour of Claude Desktop and ChatGPT, catching issues before deployment. For HTTP-based servers, use ngrok to tunnel localhost for remote testing before deploying to Cloudflare Workers, Fly.io, or Vercel.

Common Pitfalls

MCP enables powerful capabilities through arbitrary data access and code execution, which introduces security considerations that all production deployments must address. In April 2025, security researchers identified vulnerabilities including prompt injection, tool permission exploits, and lookalike tools that can silently replace trusted ones.

The protocol requires explicit user consent before invoking any tool (the "human in the loop" pattern). For remote HTTP deployments, the official SDK supports OAuth 2.0 authentication and discovery endpoints. Enterprise deployments should use auth gateways such as Pomerium, SGNL, or MCPTotal for centralised access control.

Week 1: Build and Test Locally

Use the Boilerplate Generator above to create your first server. Connect it to Claude Desktop via STDIO. Test with the MCP Inspector. Get comfortable with the tool registration and handler pattern.

Week 2: Add Real Data Sources

Connect your server to actual data: a SQLite database, a local file system directory, or a REST API your team uses daily. Add multiple tools with proper Zod input validation and meaningful tool descriptions that help the AI understand when to use each tool.

Week 3: Deploy Remotely

Switch from STDIO to HTTP transport using the Express or Hono middleware. Deploy to Cloudflare Workers or Fly.io. Add OAuth 2.0 authentication. Share the server URL with your team so multiple people can connect their AI clients to the same data bridge.

Week 4: Scale to Multi-Agent

Build multiple specialised MCP servers (one for databases, one for APIs, one for file systems) and connect them all to your AI client. Explore the new code execution patterns where agents write code to interact with MCP servers rather than making direct tool calls, reducing token consumption for complex workflows.

Published: March 11, 2026  |  Last Updated: March 11, 2026