The AI update that isn’t about a model - a new spec for MCP is about to land.
On 28 July, the update worth paying attention to won't come from a model release. It's a spec update to MCP - the Model Context Protocol, the standard that lets AI agents connect to tools and data.
The update formalises something that's been quietly shipping in pieces since January - agents that don't just tell you what they found, but show it to you, rendered directly inside the conversation you're already having with them.
Spec documents don’t come with big marketing budgets, so this is going pretty un-noticed, but if you're the person who has to sign off on letting an agent touch your systems, this is the update that really counts. This is every bit a governance story, as it is a UX one.
January was a long time ago - remind me what happened?
Until 26 January 2026, every conversation with an AI agent worked the same way: you typed something, it typed something back. Text in, text out. If the agent needed to show you a table, a form, or a chart, it described it in words.
On that date, “MCP Apps” shipped as an official extension to MCP, a hugely significant change. A tool can now hand the agent an actual interface, rendered directly inside the conversation.
For example, in January Microsoft's rollout into Copilot included an expense workflow where receipts sitting in someone's inbox get automatically matched to pending credit card transactions, with the matching interface itself appearing inline in the chat for the user to check and confirm. No exporting to a spreadsheet. No switching to the expenses portal to reconcile it. Find, match, confirm - all in one window.
Microsoft, OpenAI, and Anthropic have all now shipped this into their agents. If your workflows still involve an AI describing what it found rather than showing you and letting you confirm it, then you're not getting full value.
Going beyond the UX story. What stops this going wrong?
Rendering a live interface inside an AI conversation raises the question: if a third-party tool can put an interface in front of a user, what stops that interface from doing something it shouldn't?
The answer is built into the standard itself, in three stages.
The interface is declared and reviewed before it ever renders.
Tools have to register their UI templates in advance, so a host can inspect and approve what's actually going to be shown to a user - rather than trusting a live, dynamically generated page.
It runs sandboxed.
The rendered interface sits inside an iframe, walled off from the rest of the page and from anything it wasn't explicitly given access to. It can't reach into your session, your other tabs, or your data outside what the tool call already provided.
Every action still goes through the front door.
When a user clicks "approve" or "confirm" inside that rendered interface, the click doesn't take a shortcut. It travels back through the exact same audited request-and-response path - JSON-RPC - as every other action the agent takes. There's no privileged backchannel a UI gets that a plain tool call doesn't.
Put plainly: adding a UI to an agent doesn't add a new way for things to go wrong. It gets held to the same accountability as everything else the agent does, making it much easier for compliance teams to be comfortable with.
Who's actually responsible for keeping that promise?
If a single AI vendor owned the standard, every governance guarantee would really just be a commercial promise from one company, on their own timeline - subject to change whenever their incentives shift.
Anthropic built the Model Context Protocol and open-sourced it in late 2024. But on 9 December 2025, Anthropic donated it outright to the Linux Foundation. It's now stewarded through the Agentic AI Foundation - more than 170 member organisations, with contributors and speakers from Anthropic, AWS, Microsoft, OpenAI, GitHub, Uber and PwC sitting in the same governance process.
This matters because no single vendor can quietly bend it to favour their own product. A business creating its infrastructure on MCP today isn't gambling with one company's continued goodwill - it's supported by a standard the entire industry now has a stake in maintaining.
Where this leaves you and your team
The tooling to collapse multi-system workflows e.g. approvals, reconciliation, exception handling, into a single conversational surface now exists, and it comes with a control model designed to meet enterprise-grade production standards.
The question worth asking internally is
“Which of our multi-tab, multi-system workflows would benefit most from MCP?”
Our advice: Watch closely how the release candidate (26-07-28 spec) - the largest revision since MCP launched - changes between now and final publication on 28 July 2026.
We're already thinking through what this means for the approval, reconciliation, and onboarding flows we build for clients. If you're looking at the same thing, or curious to find out more, we'd be glad to compare notes.