MCP just got its biggest update yet: what it means for you
AI / GenAI·5 min·30 July 2026

MCP just got its biggest update yet: what it means for you

Anthropic shipped a new version of MCP, dated 28 July 2026. It is the biggest change since the standard launched. There is nothing for you to install, but the connectors you use in Cowork, Claude.ai and Claude Code will behave differently.

Quick recap: what is MCP?

MCP stands for Model Context Protocol. It is the open standard Claude uses to talk to the rest of your software: Slack, Notion, Gmail, GitHub, your accounting tool, your CRM. Install a plugin or connector in Cowork and MCP is what sits underneath. You never see it. You only notice it when something breaks.

That is exactly what this release is about.

Three things you will notice

1. Connectors drop out less often

Until now MCP ran on sessions. Claude and the server introduced themselves first, got a session ID, and everything after that hung off that session. Lose the connection, or land on a different server instance, and the session was gone. That is why connectors sometimes died halfway and asked you to sign in again.

The handshake is gone. Every request now stands on its own and can land on any instance. For you that means fewer dropped connections and fewer tools stalling mid-task.

2. Tools can ask you something mid-task

The new piece is MRTR, Multi Round-Trip Requests. A tool can now send a question back to you while it is working.

Say you ask Claude to clean up old records in your database. Before, the tool either went ahead or got stuck on missing details. Now the server can ask: are you sure you want to delete these 400 rows? Or: which of these two accounts do you mean?

That is the difference between a tool that plows ahead and one that checks first. For anything touching money, customers or data, that difference is large.

3. Long-running work becomes reliable

Tasks used to be experimental. In this version it is an official extension with its own name, io.modelcontextprotocol/tasks, and a fixed way to poll for status.

In practice: an agent working on something for twenty minutes can now report how far along it is and where it left off. Reports, migrations, bulk processing. Work you used to have to slice into pieces.

For those who want the mechanics

The rest is infrastructure, and it explains why the above is possible now. If you only use Claude through Cowork or Claude.ai, skip this block.

Stateless. The initialize/initialized handshake and session IDs are gone. An MCP server can now sit behind an ordinary load balancer, without sticky routing and without a shared session store. That is why providers can scale their servers with far less effort.

Header-based routing. Two new headers, Mcp-Method and Mcp-Name. A gateway can route on the header instead of parsing the JSON body. That saves work on every hop.

Cacheable lists. Results from tools/list, prompts/list and resources/list now carry a ttlMs and a cacheScope. The list of available tools no longer gets refetched on every request.

Authorization hardening. Issuer validation per RFC 9207 is required. Dynamic Client Registration is deprecated and replaced by CIMD, Client ID Metadata Documents. Both matter if you build servers yourself.

The TypeScript, Python, Go and C# SDKs already ship 2026-07-28.

What is deprecated, and what to do about it

Four pieces are on their way out: Roots, Sampling, Logging, and the legacy HTTP+SSE transport. A new policy comes with them: at least twelve months of warning before anything is removed for real.

If you only use off-the-shelf connectors, there is nothing to do. Your provider updates their server and you notice nothing. If you run your own MCP server, this is the year to move off those four. Put it on the calendar for the autumn, not for next week.

The scale behind this

Two numbers from the announcement. The official SDKs are downloaded close to half a billion times per month combined. At Honeycomb, nearly twenty percent of all queries now come from agents rather than people.

That explains the direction. MCP was designed as a protocol between one model and one tool. It is being used as infrastructure now, and infrastructure has to be boring and stable.

For you it stays the same set of connectors, with less friction around them.

The full changelog is at blog.modelcontextprotocol.io. To see which connectors exist for Claude, check claude.com/connectors