What Is MCP v2? The 2026 Update, Explained Simply
An update to the Model Context Protocol arrived on July 28, 2026, and most people call it MCP v2. It is the standard that lets AI assistants connect to other software, and this is its biggest revision yet. You do not need to be a developer to understand what changed. Here it is in plain language.
What is MCP, in one minute
MCP stands for Model Context Protocol. It is a shared language between AI assistants and the tools they use. Instead of every AI app building a custom connection to every calendar, CRM, and database, both sides agree on one way to talk. The AI asks "what can you do?", the tool answers with a list, and from then on the AI can search records, file tickets, or draft invoices through that one connection.
If you have ever plugged a charger into a standard socket and thought nothing of it, you already understand why standards win. MCP is that socket for AI.
The big change: from phone calls to letters
The original MCP worked like a phone call. The AI dialed in, both sides said hello, agreed on a session, and kept that line open while they worked. If the line dropped, everything started over. Servers had to remember who was on which line, which gets expensive and fragile once thousands of assistants call at the same time.
MCP v2 works like letters. Every request now carries everything the receiver needs: who is asking, which version of the protocol they speak, what they want. No hello ritual, no session to keep alive, no memory of previous requests required. Any server can answer any letter, even if a different machine answered the last one.
That sounds like plumbing, and it is. But plumbing decides whether things break under load.
| Old MCP | MCP v2 | |
|---|---|---|
| Connection | Open line with a session id | Each request stands alone |
| If a server restarts | Session lost, start over | Nothing happens, the next request just works |
| Scaling to many users | Needs sticky sessions and shared state | Any server answers any request |
| Repeated tool lists | Asked again and again | Cached, with an expiry the server sets |
Tool lists now have a best-before date
Small change, real money. An AI used to re-request the same list of available tools over and over, and every repeat costs tokens, the unit AI usage is billed in. In v2, a server can say "this list stays fresh for an hour". The assistant stops asking, answers come faster, and the bill shrinks.
Tools can ask follow-up questions
Sometimes a tool needs one more detail before it can finish, the way a pharmacist asks for your date of birth before handing over a prescription. Under v2 there is a clean, standard way to do this: the tool pauses, states what it needs, the assistant supplies it, and the job continues. Previously this required exactly the kind of open line that v2 removed, so it rarely worked well across different apps.
Logins got safer
A handful of security fixes ship with v2. The most important one: during a login flow, the server granting access now identifies itself in a way the app must verify. That closes a known trick where a malicious server could sit in the middle of an authorization. You will never see this working. That is the point.
Will anything break?
No, and this is deliberate. Servers that adopt v2 keep answering old-style clients, and new clients fall back to the old handshake when they meet an old server. Features being phased out get at least a twelve-month grace period. Standards only work if updating is boring.
Why this matters if you use AI agents
Fewer repeated questions, fewer tokens, fewer mysterious disconnects, and tools that can ask for what they need instead of failing. If you run agents against many tools at once, the effect compounds. That routing and caching work is exactly what an MCP gateway does for you, and we wrote a plain guide to that too: [What is an MCP gateway?](/blog/what-is-an-mcp-gateway)