1:1 mentoring with Big Tech AI engineers
RAG & MCP

Building MCP Servers

Build a CRM MCP server on the current SDKs (Python mcp 2.x, with a TypeScript twin): the smallest server, type hints as the tool schema, rejecting bad input, resources and prompts, running it under the Inspector, and testing it before a model does.

Last updated

Core12 min readFirst readMCP ArchitectureTool Surface Design

After this section you can

  • Build an MCP server on the current Python SDK and read the tool schema it derives from your type hints
  • Make bad input fail loudly with Literal and Field limits, and say what still needs authorization on the server
  • Run the server under the Inspector and test it with an in-memory client before any model calls it
03

Building MCP Servers: From Eight Lines to Tested

Type hints become the tool schema, bad input fails loudly, and a test finds the first mistake before a model does.

Key idea

An MCP server is ordinary functions with a decorator. The SDK reads each signature and docstring, publishes it as the tool a model sees, and rejects calls that do not fit. This page builds one for Acme’s CRM on mcp 2.x, revision 2026-07-28, and tests it without a model.

Your signature is the tool schema
SERVER.PY · WHAT YOU WRITE from mcp.server import MCPServer mcp = MCPServer("acme-crm") @mcp.tool() def get_account(account_id: str) -> str: """Look up one account by id.""" return "Acme Corp, active" if __name__ == "__main__": mcp.run() nameinputSchemaoutputSchemadescription TOOLS/LIST RETURNS · TRIMMED { "name": "get_account", "description": "Look up one account by id.", "inputSchema": { "properties": {"account_id": {"type": "string"}}, "required": ["account_id"]}, "outputSchema": { "properties": {"result": {"type": "string"}}, "required": ["result"]}}

Captured from mcp 2.2.0 and 2.3.0, trimmed (step 2). The host sends the model name, description and inputSchema (Inside a Tool Call); outputSchema is for the client.

Related

More in RAG & MCP

RAG Architecture

Preview

What RAG is and when to use it instead of a long prompt, the offline and online pipelines, what Claude receives as cited sources, and five numbered gates that locate the stage which lost a wrong answer.

Document Processing

Preview

Turn files into clean, typed elements with provenance: choose a parser tier per file, keep tables as structure, clean without deleting content, dedupe without dropping new versions, and gate the output.

Chunking Strategies

Preview

Choose what to cut on, size chunks in the embedder's own tokens, keep headings, tables and code whole, search small but return big, and measure any change with recall@k on real questions.

Vector Indexes & Stores

Preview

How a vector index stores and searches embeddings: the index as a rebuildable copy, how HNSW finds neighbours and what ef_search changes, RAM arithmetic, and choosing a store from constraints.

Get full access to all 149 sections with code examples, diagrams, and interactive animations.

Unlock Premium