Model Context Protocol (MCP) Server Development

The Model Context Protocol is an open JSON-RPC 2.0 protocol that lets AI clients call tools, read resources, and fetch prompts from servers. A server is code you write; the client and the model are not trusted to behave. Design every tool as if its arguments were hostile.

Deeper material:

When to use / not use

Use this skill to design, write, harden, test, or migrate an MCP server. Do not use it to wire an existing server into Claude Desktop/Cursor/etc. (client config), and do not use it for REST/GraphQL APIs that merely sit next to an MCP server. For auditing third-party agent extensions see skill-security-audit; for reviewing server code for vulnerabilities see security-code-review.

Version facts to check first (they change)

The latest spec revision at time of writing is 2026-07-28 (previous: 2025-11-25, 2025-06-18, 2025-03-26, 2024-11-05). Always confirm at modelcontextprotocol.io/specification/latest.
Transports defined by the spec: stdio and Streamable HTTP. The 2024-11-05 HTTP+SSE transport is deprecated since 2025-03-26 and must not be used for new servers. A bare "SSE transport" in older tutorials means that deprecated transport.
Revision 2026-07-28 made MCP stateless: no initialize handshake, no Mcp-Session-Id, no standalone GET stream; each POST carries protocol version and client capabilities in _meta. Servers must implement server/discover. Older clients still speak the 2025-xx shape, so a public server decides deliberately which eras it supports (see the migration reference).
TypeScript SDK: current line publishes @modelcontextprotocol/server (and /client); the older v1 line is @modelcontextprotocol/sdk on the v1.x branch. Python SDK: PyPI package mcp (v2 line; MCPServer in mcp.server; v1 exposed FastMCP in mcp.server.fastmcp). Read the SDK README for the version you install; do not mix v1 and v2 imports.

Core rules

Validate every argument with a schema (Zod/Standard Schema in TypeScript, type hints/Pydantic in Python) and re-check semantic limits (ranges, allowed enums, path containment) in handler code. A JSON Schema alone does not enforce policy.
stdio: stdout is the protocol channel. Any console.log/print corrupts framing. Log to stderr or a file.
Least capability per tool. One narrow, typed tool beats run_command/execute_sql. Never build shell strings from arguments; call executables with an argument array.
Pick the right primitive. Tools are model-invoked actions; resources are application-controlled read-only context; prompts are user-invoked templates.
Tool failures are results, not crashes. Return isError: true with a short, non-secret message for execution failures; use protocol errors for malformed requests. Never leak stack traces, tokens, or absolute server paths.
Annotations are hints, not enforcement. readOnlyHint/destructiveHint help clients prompt users, but a client may ignore them and a malicious server can lie. Enforce authorization in the server.
Treat tool output as untrusted data on the client side and as a prompt-injection channel on yours. Do not return fetched web or file content that you have not bounded in size, and never put secrets into tool results.
Remote servers need transport security (below). Local convenience defaults are not safe on a network.

Workflow

Define the surface: list the 3-10 verbs users need. For each, decide tool vs resource vs prompt, inputs, side effects (read-only, idempotent write, destructive), and who may call it.
Choose transport: stdio for a local process launched by the client; Streamable HTTP for anything shared, remote, or multi-user. Decision rule: if more than one user or machine will connect, you need HTTP plus authentication.
Implement with the high-level server class (McpServer.registerTool in TypeScript; @mcp.tool() in Python). Use the low-level request-handler API only when you need custom protocol behavior.
Harden: path containment, allow-lists, size/time limits, secret handling, authz, rate limits (checklist below).
Test: unit-test handlers directly; run the server under the Inspector; test one hostile input per tool (traversal, oversized, wrong type, injection string).
Document the tool contract (description precise enough that the model picks the right tool; state side effects and limits).

Minimal TypeScript server (SDK v2 line)

import { McpServer } from "@modelcontextprotocol/server";
import { StdioServerTransport } from "@modelcontextprotocol/server/stdio";
import * as z from "zod/v4";

const server = new McpServer({ name: "calculator", version: "1.0.0" });

server.registerTool(
  "calculate_sum",
  {
    description: "Add two finite numbers and return the sum. No side effects.",
    inputSchema: z.object({ a: z.number().finite(), b: z.number().finite() }),
    annotations: { readOnlyHint: true },
  },
  async ({ a, b }) => ({ content: [{ type: "text", text: String(a + b) }] }),
);

await server.connect(new StdioServerTransport());
console.error("calculator MCP server running on stdio"); // stderr only

On the v1 SDK the same shape is import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js" with StdioServerTransport from @modelcontextprotocol/sdk/server/stdio.js. Confirm exact imports against the README of the installed version.

Minimal Python server (SDK v2 line)

from mcp.server import MCPServer

mcp = MCPServer("Demo")

@mcp.tool()
def add(a: int, b: int) -> int:
    """Add two integers and return the sum. No side effects."""
    return a + b

Run it with the SDK CLI (mcp[cli] extra): uv run mcp run server.py for stdio, or uv run mcp run server.py --transport streamable-http for HTTP. On v1, import FastMCP from mcp.server.fastmcp instead.

Path containment (the most common real bug)

import { realpath } from "node:fs/promises";
import path from "node:path";

export async function resolveInside(baseDir: string, userPath: string): Promise<string> {
  const base = await realpath(baseDir);
  const target = await realpath(path.resolve(base, userPath)); // follows symlinks
  const rel = path.relative(base, target);
  if (rel.startsWith("..") || path.isAbsolute(rel)) {
    throw new Error("path escapes the allowed directory");
  }
  return target;
}

realpath fails for files that do not exist yet; for create operations resolve the parent directory and re-join the validated file name. A plain startsWith(base) string check is wrong (/data also prefixes /data-secret).

Remote (Streamable HTTP) hardening checklist

Validate the Origin header on every request and answer 403 when it is present and not allow-listed (DNS-rebinding defense; required by the spec).
Local development servers bind to 127.0.0.1, not 0.0.0.0.
Authenticate every connection. For HTTP the spec's authorization model treats the MCP server as an OAuth 2.x resource server: publish protected-resource metadata, validate access tokens (signature, expiry, issuer, and that the audience is this server), and never forward a client's token to an upstream API (no token passthrough).
Do not trust identifiers a client supplies for tenancy; derive the tenant from the validated token.
Set request body size limits, per-call timeouts, and rate limits.
Behind nginx or similar, send X-Accel-Buffering: no on SSE responses so events are not buffered.
If you must serve pre-2026 clients, treat Mcp-Session-Id as an untrusted correlation id, never as authentication.

Testing

# Interactive inspector (stdio server)
npx @modelcontextprotocol/inspector node build/index.js

Verify: tool list renders, a valid call succeeds, each invalid call returns isError or a protocol error without crashing, and nothing but JSON-RPC appears on stdout. For HTTP servers, start the server separately and connect the Inspector to its URL. Inspector UI details change between releases; follow its README rather than memorized menus.

Anti-patterns

Anti-pattern	Why it fails	Instead
`console.log`/`print` in a stdio server	Corrupts JSON-RPC frames; client drops the connection	stderr or file logging
New server built on the HTTP+SSE transport	Deprecated since 2025-03-26; clients migrate away	Streamable HTTP
`startsWith(baseDir)` containment check	Prefix and symlink bypass	`realpath` + `path.relative`
Building a shell command string from arguments	Command injection	`execFile("git", [subcommand, ...])` with allow-listed subcommands
Relying on `readOnlyHint` for safety	Hints are advisory	Server-side authz and read-only DB roles
Passing the client's bearer token to a downstream API	Confused deputy; breaks audience binding	Separate server credential or proper token exchange
Returning raw upstream error bodies	Leaks internals and secrets	Short mapped message; log detail to stderr
Copying a 2025-era tutorial that depends on sessions or `initialize`	Wrong for the stateless revision	Check the revision your SDK targets

Verification checklist

Every tool has a schema and handler-side semantic checks.
Nothing but protocol frames on stdout in stdio mode.
Failure paths return errors without crashing or leaking secrets.
Path, command, and URL arguments are allow-listed or contained.
HTTP servers validate Origin, authenticate, and bind to localhost in dev.
Spec revision and SDK version recorded in the README of the server.
Hostile-input test exists for each tool.

Sources

MCP specification (latest), transports and changelog: https://modelcontextprotocol.io/specification/latest
TypeScript SDK: https://github.com/modelcontextprotocol/typescript-sdk
Python SDK: https://github.com/modelcontextprotocol/python-sdk

mcp-server-development

Files4

SKILL.mdskills/mcp-server-development/SKILL.md