The battle for agent connectivity: Can MCP survive the enterprise?

Model Context Protocol, a protocol that lets AI applications connect to various data sources and tools, promises to be the USB-C for AI. As such, it is top of mind for the AI agent community. And despite concerns with its shortcomings and lack of enterprise readiness, it seems to be well on its way to becoming the official standard.

But in an industry moving at the speed of AI, it’s not a sure bet. Even within the agent community, views differ on whether Model Context Protocol (MCP) will prove to be the universal standard for agentic integration.

“MCP aims to organize how agents reach data, but our deployments show that regulated enterprises already follow a different pattern,” said Nuha Hashem, co-founder and CTO at Cozmo AI, a conversational commerce platform on WhatsApp.

Ironically, a protocol designed to end fragmentation is hitting a wall because enterprise AI is itself fragmented.

For some, to be sure, MCP is crucial.

“MCP is the UI for agents. The future of asking ChatGPT to book an Uber and have a pizza available when you arrive at the hotel only works if we have the connectivity,” said Dag Calafell III, director of Technology Innovation at MCA Connect, an IT consultancy for manufacturers.

But while seamless connectivity might be the Holy Grail for consumer apps, critics argue that it is irrelevant — or even dangerous — for the enterprise.

Related:Accenture, Anthropic and the quiet rise of AI integrators

“To me, the entire conversation around a universal public crawling protocol is quickly becoming an irrelevant point. The real, high-value information that AI agents will need for professorial activity, is not exposed on the public internet,” said Mark Friend, director of Classroom365, a provider of IT support for schools across the U.K.

In the education sector, Friend said, the data that matters is not on a public-facing website.

“It is locked into a school’s management information system, which holds all pupil, staff and safeguarding data. No headmaster is ever going to let an AI agent crawl that database, they will demand the AI authenticate via a secure third-party API provider,” Friend added.

This “locked door” reality brings the argument back to Hashem’s take above: Regulated companies are going their own way, and MCP and other protocols just aren’t part of their path. In their enterprises, access isn’t about a universal plug; it’s about satisfying internal governance.

“The main issue is not a missing protocol; the issue is the need for each request to match a defined rule inside the company,” Hashem explained.

Related:What tidings will AI provide for workers in 2026?

“In banking and insurance, we see this in production. An agent can reach data only through a path that ties the step to a policy the company has already reviewed and approved. That approach keeps behavior visible and keeps decisions linked to context,” she said.

Even so, there are plenty of companies that are not so heavily regulated, and millions of individuals who would like agents to be able to retrieve relevant data and tools, whether they reside on the internet or elsewhere. To them, MCP appears to be the perfect fit.

But if not, they also wouldn’t be particularly resistant to the use of a competitor instead. So where does that leave MCP?

Competing protocols and changing approaches

Notably, MCP has significant backing from prominent companies, including Google, OpenAI, Microsoft and its creator, Anthropic. Indeed, Calafell argued that while there are competitors out there, “MCP is winning” precisely because it has seen significant adoption by large software providers.

Still, MCP clearly has significant issues — mostly because it’s in its infancy. MCP’s rapidly evolving specification, uneven tooling, unclear security and governance controls, and lack of standardized memory, debugging, and orchestration make it better for experimentation than reliable enterprise use today.

Related:3 ERP experts on AI’s impact on finance: Why finance will never be the same

Which, of course, leaves an opening for competing protocols.

“I think there are two valid substitutes for MCP,” said Xiangpeng Wan, product lead at NetMind.AI. “One is called the Universal Tool Calling Protocol (UTCP), which allows direct API calls, no server overhead, low latency, and leverages existing tool security features. The other is called Agent-to-Agent Protocol (A2A), which focuses on agent collaboration rather than tool access, and is backed heavily by Google and Microsoft.”

Wan said UTCP has good odds for replacing MCP, A2A has the best odds, and function calling (the AI capability to access and execute external tools or APIs) has the least chance. That said, he added that he doesn’t see a single protocol dominating soon.

“What we’ll most likely see is a combined ecosystem of MCP, UTCP, A2A, function calling and managed solutions in the near future,” Wan said.

Others feel the point is moot, as MCP will likely improve over time.

“MCP began as a protocol created by one AI vendor, Anthropic, and the rest of the ecosystem had to adjust rapidly,” said Yossi Pik, co-founder and CTO at AppSec and vibe coding security company Backslash Security. With that type of fast rollout, it is natural that the first version will need time to mature, he added, and if MCP is not fully adopted, “it simply means the community will try to iterate toward something stronger,” he said.

There is also the possibility that MCP will inspire other developments.

“There is room to innovate with a security-first ‘MCP-like’ standard that is resource aware, with trusted catalogues, privileges, scopes, etc. These would either be built on top of MCP, a sort of MCP v2, or introduced as part of a new protocol,” said Liav Caspi, co-founder and CTO at Legit Security.

And, of course, there remains an evolving trend that the AI industry will take an entirely different direction.

“The alternative to MCP is not another open web standard; it is the authenticated, pay-to-play, private API economy,” Friend said. If MCP is not accepted as the standard, well, frankly, who cares?

“There will not be another open protocol that takes its place. It will be replaced by thousands of private, authenticated, vendor-specific APIs,” he said.

That, in turn, will inevitably lead to vendor lock-in — a scenario that is indeed lurking in the shadows already, as model builders seek to gain and secure market share. But that is another story for another day.

For now, expect MCP to remain relevant — though not the ultimate answer — for many AI applications in retrieving the data and tools that they need.