MCP goes stateless: what the 2026 spec change means for UK businesses

Most UK business owners have never heard of the Model Context Protocol. That is exactly why this change matters.

On 28 July 2026, Anthropic finalised the most significant update to MCP since it launched. The revision rewrites a core piece of plumbing that sits beneath many of the AI agent tools your business may already be using. You will not see it in a product changelog. But if your vendors have not kept up, you might start seeing things break.

What MCP actually is, in plain English

Think of MCP as a universal plug socket for AI tools. It is a shared standard that lets AI assistants and agents connect to external data sources, applications, and services. Rather than every AI tool building its own bespoke connection to every platform it needs to talk to, MCP gives them all a common language.

Anthropic, the company behind the Claude AI platform, designed MCP so developers could build once and connect to many. The protocol has grown quickly. By 2025, developers had downloaded the MCP SDK over 97 million times per month, and at least 10,000 MCP servers had been deployed. That level of adoption means this architectural change reaches far beyond Anthropic's own products.

If your business uses AI agents, AI-connected workflows, or any tool that pulls live data into an AI assistant, there is a reasonable chance MCP is part of the infrastructure underneath it.

What changed on 28 July 2026

The original version of MCP was stateful. That means every connection between an AI tool and a data source maintained a continuous, persistent link, like a phone call that stays open for the entire working day.

This worked well when developers were building and testing on their own laptops. It became a problem when those same tools moved into business environments with many simultaneous users. A persistent connection for every session, across every user, across every data source, adds up fast. Scaling it required either reworking the tooling or accepting that performance would degrade as usage grew.

The 2026 revision removes protocol-level sessions entirely. MCP is now stateless. Each request carries all the information it needs to complete, then closes. Routing information now travels in HTTP headers rather than in the body of each message. This mirrors how most of the modern web already works and makes MCP far easier to deploy behind standard cloud infrastructure.

The revision also deprecates two features that were rarely used in practice: sampling and roots. These remain functional for at least 12 months after the finalisation date, giving vendors time to adjust. A new extensions model replaces the old feature set, allowing businesses and developers to adopt only the MCP capabilities relevant to their specific use case.

For businesses using vendor-built tools and official MCP SDKs, the disruption should be minimal. For anyone who built a custom MCP engine, substantial rework is required to remain compatible.

Why this matters for UK SMEs right now

The honest answer is that most SME owners will never touch MCP directly. Your IT team, your software vendor, or the consultancy managing your AI setup will deal with the technical side. But the business impact is real, and it lands on your desk in three ways.

First, vendor compliance becomes a due diligence question. Operations managers and IT leads at businesses using AI-connected tools should ask their vendors directly: have you updated to the 28 July 2026 MCP specification? A vendor running the old stateful architecture against a client expecting the new stateless standard is a compatibility failure waiting to happen. The time to ask is now, not after something stops working.

Second, the architectural shift lowers the cost of scaling AI agent workloads. The old stateful design forced a choice between expensive, complex infrastructure or performance that degraded as more people used the system. Stateless architecture removes that trade-off. For SME owners planning to extend AI agents across more of their business, the new spec means those plans do not hit a hidden ceiling.

Third, the new extensions model makes selective adoption more straightforward. Your business does not need to implement every feature MCP offers. The revised spec formalises the ability to take only what you need, which keeps implementations leaner and easier to maintain. This is the 80% of the value for 20% of the complexity, which is exactly the adoption profile that works for a business without a large internal technology function.

Three actions to take this week

These three steps can be completed within five working days and require no technical background to initiate.

1. Ask your AI tool vendors for a compliance statement. Contact anyone supplying you with AI agent tools, AI-connected workflows, or MCP-based integrations and ask them to confirm they are aligned with the 28 July 2026 MCP specification. A reputable vendor should be able to answer this in a short email. If they cannot, treat it as a red flag.

2. Map which of your current AI tools depend on MCP. You do not need to understand the technical architecture to do this. Ask your IT lead or your AI implementation partner to list which tools in your stack use MCP connections, and whether any were built on custom engines rather than official SDKs. This gives you a clear picture of where your exposure sits.

3. Review your AI agent roadmap against the new scaling model. If your business has plans to expand AI agent usage across more departments or users, spend 30 minutes with whoever owns that roadmap. Confirm that the architecture you are building on is aligned with the stateless standard. This is a short conversation now that avoids a significant rework conversation later.

Considerations and limitations

The 2026 MCP revision is not an immediate crisis for most UK SMEs. Deprecated features remain functional for at least 12 months after the finalisation date, which gives vendors reasonable time to update. Businesses using tools built on official SDKs are likely to see changes rolled out automatically through normal software updates.

The genuine risk sits with two groups. Businesses that built custom MCP tooling in-house or through a third party face real rework if that tooling is not updated. And businesses that have not asked their vendors about compliance are operating without visibility, which is a different kind of risk. The spec change is not optional. It is a forced modernisation, and the 12-month grace period is not an indefinite delay.

There are no GDPR or ICO implications specific to this architectural change, but any review of your AI agent infrastructure is a reasonable moment to revisit your data handling and processing records.

Your next step

If your business is running AI agents or planning to expand into them, the MCP spec change is exactly the kind of hidden architectural issue that is worth surfacing before it becomes a live problem. The AI Readiness survey is designed to help SME owners and operations leads identify precisely this type of gap, the infrastructure and vendor questions that do not appear on a product changelog but do affect whether your AI investments hold up at scale.

Completing the survey gives you access to 65+ free resources and a custom AI Readiness report. That report is the starting point for a free 45-minute AI Readiness call where we walk through the findings together.

If you are building or scaling AI agents and want a second opinion on whether your architecture will hold, that is work gecco does directly through our AI Agents service line.

Website · LinkedIn · Case Studies · Newsletter