MCP cheat sheet (protocol revision 2026-07-28)

Roles

• Host: the AI app the user talks to
• Client: one connection per server, inside the host
• Server: exposes tools, resources, prompts

Server features

• Tools: functions the model may call
• Resources: read-only data by URI
• Prompts: reusable templates

Client features: elicitation, sampling, roots (optional)

Transports

• stdio: newline-delimited JSON-RPC to a subprocess. Never print to stdout; log to stderr. Credentials from the environment.
• Streamable HTTP: each message is an HTTP POST to one endpoint. Follow the MCP authorization spec.

2026-07-28 request model

• No initialize handshake, no session
• Every request carries protocol version and client capabilities in params._meta
• Optional server/discover returns supported versions and capabilities
• HTTP also sends the MCP-Protocol-Version header

Required _meta keys (seen in a real run)

"_meta": {
  "io.modelcontextprotocol/protocolVersion": "2026-07-28",
  "io.modelcontextprotocol/clientCapabilities": {}
}

Observed with the Python SDK 2.3.0: missing keys return error -32602.

Security rules

• Read-only by default; approve consequences
• No raw shell, SQL or URL tools
• Least privilege; authorise per user and per tool
• Treat tool output as untrusted (prompt injection)
• Bound inputs and outputs; time out and rate-limit
• Never put secrets in client apps

Use MCP when…

• Several AI apps need the same capability
• Tools are naturally model-callable functions

Skip MCP when only your own code calls a fixed API. Use REST or function calling.

Python quickstart (tested)

uv add "mcp==2.3.0"
from mcp.server import MCPServer
mcp = MCPServer("demo")
@mcp.tool()
def add(a: int, b: int) -> int: ...
mcp.run()  # stdio