JobsDart LogoJobsDart

How to build an MCP server: step-by-step developer tutorial with TypeScript

The Model Context Protocol standardizes how AI applications connect to external tools and data sources. Here is how to build and test your own MCP server from scratch.

JobsDart Editorial3 min read

Key takeaways

  • The Model Context Protocol (MCP) exposes three primary capabilities to AI clients: Tools (executable functions with side effects), Resources (contextual read-only data), and Prompts (reusable templates).
  • Building an MCP server requires defining strict input schemas using libraries like Zod and choosing a transport layer (stdio for local desktop apps, SSE/HTTP for remote servers).
  • The official MCP TypeScript SDK simplifies server creation, protocol handshake negotiation, and error handling.
  • MCP servers decouple data access from specific AI models, allowing the exact same tool to be shared across Claude Desktop, Cursor, and custom autonomous agents.
  • Production security requires validating all parameters, executing sandboxed read-only database queries, and auditing client invocations.

Why MCP is replacing bespoke tool integrations

Before the release of the Model Context Protocol (MCP), integrating external data into an AI assistant was fragmented. Every developer wrote custom function-calling definitions, glued together bespoke REST wrappers, and rewrote the entire integration whenever switching between Claude, OpenAI, or local open-source models.

MCP acts like the Language Server Protocol (LSP) for artificial intelligence. By creating an MCP server once, any compatible client—whether Claude Desktop, an IDE coding assistant, or an enterprise agentic workflow—can instantly discover your tools, inspect input schemas, and invoke backend functions.

In this build an MCP server tutorial, we walk through building a complete, production-grade MCP server using TypeScript and Node.js that connects an AI agent directly to a live PostgreSQL recruitment database.

  • One standardized protocol connects diverse AI clients to private tools and data.
  • Eliminates rewriting custom API wrappers for every LLM vendor.
  • Enforces typed parameters and structured JSON inputs via client-side discovery.

MCP architecture: Tools, Resources, and Transports

An MCP server communicates with client applications through three distinct primitives and two standardized communication transports.

The Core Primitives of Model Context Protocol
PrimitiveTypeDescriptionReal-World Example
ToolsExecutable ActionFunctions the LLM can invoke with arguments to alter state or fetch dynamic dataschedule_interview, create_job_post
ResourcesRead-Only DataStructured documents, files, or URI endpoints the client reads into contextpostgres://jobs/active, file://resumes/alice.pdf
PromptsInteraction TemplatesPre-engineered workflow prompts presented directly to human usersreview_candidate_code, screen_resume_rubric
stdio TransportProcess I/OLocal inter-process communication using standard input and output streamsClaude Desktop app connecting to local Node/Python script
SSE TransportNetwork HTTPServer-Sent Events over HTTP for cloud-hosted remote multi-tenant serversEnterprise SaaS backend serving remote agent swarms

Step-by-step implementation with the TypeScript SDK

To get started, initialize a new Node.js project and install the official MCP TypeScript SDK alongside Zod for schema validation.

Using the Server class from @modelcontextprotocol/sdk/server, register handlers for ListToolsRequestSchema and CallToolRequestSchema. Input arguments are validated at runtime via Zod schemas, returning structured candidate records directly to the requesting model.

Finally, attach the server to StdioServerTransport, enabling local AI clients like Claude Desktop or Cursor to communicate over standard input and output streams.

Testing and connecting to Claude Desktop

To connect your newly built server to Claude Desktop, open the local configuration file located at claude_desktop_config.json in your application data directory.

Add your server configuration under the mcpServers key, specifying node as the executable command and passing the absolute path to your compiled server.js script along with any required environment variables.

When Claude Desktop boots, it executes the protocol handshake, displays your tool under the chat input, and automatically triggers it when users ask relevant recruitment questions.

  • Place configuration in %APPDATA%/Claude/claude_desktop_config.json on Windows.
  • Pass environment variables like DATABASE_URL directly in the configuration object.
  • Restart Claude Desktop to verify the tool handshake in the conversation window.

Frequently asked questions

What is the difference between an MCP tool and an MCP resource?

Tools are active functions invoked by the model that can perform calculations, query databases, or trigger side effects. Resources are passive data streams (like file contents or system logs) that the client reads into context without execution.

Can MCP servers run on remote cloud servers instead of local machines?

Yes. While stdio transport is used for local tools, MCP supports Server-Sent Events (SSE) over standard HTTP, allowing teams to deploy centralized cloud MCP servers accessible by remote agent swarms.

Is MCP exclusive to Anthropic Claude?

No. MCP is an open-source standard released under the MIT license. Many leading agent frameworks (including LangChain, LlamaIndex, Cursor, and Continue.dev) natively support MCP clients and servers.

Further reading

Check this against your own resume

Scan your CV against a real job description, or build a parse-safe one from scratch. Your first scan costs nothing.

Free ATS checkerAI resume builder

Keep reading

AI EngineeringMCP vs RAG: what is the difference and when should you use each?RAG relies on pre-computed static vector embeddings. MCP provides open, standardized tool calling to connect models directly to live databases, APIs, and file systems.
AI EngineeringHow to build an Agentic RAG system: an engineering blueprintLinear RAG scripts fail on complex queries. Here is a production-tested engineering blueprint for building an Agentic RAG system with state graphs, tool routing, and automated self-critique.
AI EngineeringHow to build an AI database assistant: a complete Text-to-SQL guideGiving an AI direct access to your database is immensely powerful and dangerous if misconfigured. Here is a production-hardened engineering guide to building a secure Text-to-SQL database assistant.
All career guides