MCP changing

What it is

MCP is a JSON-RPC 2.0 protocol between a client (an agent, an IDE, a chat app) and a server that exposes three kinds of things: tools the model can call, resources it can read, and prompts it can load. Anthropic released it in November 2024. It is now the default way to give any agent access to a capability without writing an integration per agent.

How it works

A server publishes tools/list, resources/list and prompts/list. The client fetches those, hands the tool definitions to the model, and forwards tools/call when the model picks one. Transport is either stdio for local servers or Streamable HTTP for remote ones.

Since the 2026-07-28 revision the protocol core is stateless:

• There is no initialize handshake. Every request carries the protocol version and client capabilities in _meta.
• There are no protocol-level sessions and no Mcp-Session-Id header. Any server replica can answer any request.
• List results no longer vary per connection, so they are cacheable.
• Servers that need state across calls mint an explicit handle and pass it back as an ordinary tool argument.
• A formal extensions framework and hardened OAuth-based authorization ship in the same release.

When to use it

• You own a system and want every agent product to be able to use it. Write one MCP server.
• Your agent needs third-party capabilities. Consume MCP servers instead of hand-coding tool wrappers.

Do not reach for MCP to talk to another agent. That is what A2A is for.

What changed

• 2026-07-28: stateless core, header-based routing, cacheable lists, extensions.
• 2025-06-18: servers became OAuth resource servers; structured tool output; elicitation.
• 2025-03-26: Streamable HTTP replaced the original HTTP+SSE transport.

Watch out for

• Tool descriptions are model input. A malicious server can inject instructions through them. See prompt injection.
• Servers written against pre-2026 SDKs assume a session. They need a migration, not just an upgrade.