The Block · T3 Tools & protocols

Model Context Protocol (MCP)

Last reviewed 21 Oct 2026by Pratima Chainani

An open standard for connecting AI applications to external tools, data and workflows, so any compliant client can use any compliant server. Its own docs compare it to a USB-C port for AI applications.

MCP specifications are named by release date. The 2026-07-28 release made the core protocol stateless: there is no initialize handshake or session, and each request carries what it needs, so it can hit any server instance.

The same release deprecated Roots, Sampling and Logging, let tool schemas use any JSON Schema 2020-12 keyword, moved Tasks out of the core into an official extension, and tightened authorization. The TypeScript and Python SDKs serve both protocol versions; Go needs a flag.

MCP doesn’t replace your APIs. Most MCP servers describe your endpoints as tools and forward the call, so your API’s auth, rate limits and validation still do the protecting. MCP is governed by the Linux Foundation’s Agentic AI Foundation, alongside A2A.

Common misread

MCP replaces your APIs.

MCP sits on top of them. The server describes your endpoints as tools and forwards the call.

For your code

If your server keeps per-session state, move it to a store keyed per request. Keep API-side auth and limits.

For your roadmap

Worth one sprint if you own more than a couple of MCP servers; plan auth testing.

Sources

  1. MCP specification 2026-07-28, changelog · 28 Jul 2026
  2. MCP docs, What is the Model Context Protocol · 2026-07-28 docs
  3. AAIF, A2A joins AAIF’s open agentic stack · 17 Aug 2026