If you’ve ever paid a developer to “just make the CRM talk to the accounting system,” you already know the real cost of integration isn’t the software, it’s the custom glue code holding it together. Every vendor speaks a slightly different language, every API changes on its own schedule, and every connection someone builds becomes one more thing that can quietly break at 2 a.m. On July 28, 2026, the technical standard increasingly used to build that glue, called the Model Context Protocol, or MCP, reached its most significant milestone yet: a finished, production-ready specification, backed by every major cloud and software vendor. It’s a plumbing story, not a headline-grabbing one, but it’s worth five minutes of a business owner’s attention.

What actually shipped on July 28

MCP is an open standard, originally introduced by Anthropic in late 2024, that gives AI applications a common way to connect to business systems, data sources, and workflows, instead of requiring a custom-built connector for every pairing of AI tool and business system. The maintainers describe the July 28 release as the largest revision since MCP launched. The core change is that the protocol is now “stateless,” meaning a server handling one of these connections no longer has to keep a fragile, sticky session open for each conversation. In practical terms, that means the software running these connections can now be spread across ordinary, off-the-shelf servers instead of specialized infrastructure, which tends to make it cheaper to run and less prone to the kind of outage that happens when one overloaded server falls over.

The release also locks in a formal set of rules for how long features can stick around before they’re retired, and it hardens how these connections handle logins and permissions, aligning more closely with the login standards banks and enterprise software already use. Coverage from The Register framed the shift as MCP “breaking with its stateful past” to become boring, dependable infrastructure, which, for anything a business depends on every day, is exactly the compliment you want.

Why this lands squarely in ERP and CRM territory

MCP isn’t an abstract developer concern. Analysis from ERP Today, a trade publication read by enterprise resource planning professionals, put it plainly: MCP is becoming the connective layer for “agentic ERP,” giving AI agents a standard way to reach into finance, procurement, manufacturing, payroll, and customer service systems rather than requiring a bespoke integration for each one. SAP, Salesforce, Oracle NetSuite, Microsoft, and Snowflake have all built MCP support into their platforms over the past year, according to the same reporting, alongside Databricks and a fast-growing list of CRM vendors.

That same analysis raised a fair warning alongside the opportunity: making a connection easier to build doesn’t automatically make it safe. A poorly designed connection can hand an AI agent access to sensitive data, pricing, payroll, customer records, at a level no individual employee would normally be granted through the regular software interface. The article’s recommendation for ERP and CRM leaders was to treat every one of these connections as a governed piece of infrastructure, with a clear owner, an audit trail, and defined limits on what it’s allowed to touch, not a shortcut to skip past.

The adoption numbers suggest this isn’t a fad

It’s easy to be skeptical of any new technical standard; plenty never leave the whiteboard. This one appears to have already cleared that bar. Independent tracking published in early July put MCP adoption at 78% of enterprise AI teams running it in production, with 28% of Fortune 500 companies operating their own MCP-based connections, and monthly developer downloads of the underlying toolkit near 97 million. Every major cloud platform and most large business software vendors now support it natively. For a standard that only became widely used in the past two years, that’s a fast climb, and it mirrors the pattern IT buyers have seen before with things like universal plug types or common file formats: once enough vendors agree on one way of connecting, switching or adding systems gets measurably cheaper for everyone downstream.

What this means if you’re planning a system integration project

None of this requires a business owner to understand a protocol specification. It does mean the conversation with your next software partner should include a few pointed questions.

1. Are you building us a custom connector, or using an existing, standard one? If most of the major platforms you already run, your CRM, your accounting software, your e-commerce platform, already support a common connection standard, a fully custom integration may cost more and take longer to maintain than it needs to. There can be good reasons for custom work, but you should know which one you’re paying for and why.

2. Who can see what, and who’s watching? Before any AI tool gets a live connection into systems holding customer data, financials, or employee records, ask exactly what it can read, what it can change, and how you’d find out if something went wrong. “It’s connected” and “it’s governed” are two different sentences, and only the second one should let you sleep at night.

3. What happens when the standard changes again? This release included breaking changes for anyone who built against the older version. A software partner who keeps your integrations current as standards evolve is worth more than one who quietly leaves you on unsupported infrastructure two years from now.

The takeaway

The plumbing connecting your business systems to each other, and increasingly to AI tools, just got a widely agreed-upon rulebook. That’s good news for cost and reliability over the next few years, but it shifts the real work from “can we connect these systems” to “should this connection exist, and who’s accountable for it.” Both are questions worth asking before, not after, a project starts.

If you’re evaluating an ERP or CRM integration project, or wondering whether your current connections are built on solid ground, we’re glad to talk it through. Email us at info@kodevox.com or reach out through our contact page.

Sources and further reading:

• Model Context Protocol Blog — The 2026-07-28 MCP Specification Release Candidate (final spec published 28 July 2026)
• The Register — Model Context Protocol prepares to break with its stateful past (23 July 2026)
• ERP Today — MCP Is Becoming the Plumbing of Agentic ERP (2 July 2026)
• andrew.ooo — MCP Enterprise Adoption: The July 2026 State of Play (1 July 2026)

— The Kode Vox Team