MCP goes mainstream: open protocols reshape enterprise AI
The gist
Open protocols like MCP are finally making enterprise AI integrations fast, cheap, and reliable—after years of clunky, costly custom builds.
What to know
- Anthropic open-sourced MCP in late 2024, and Google plus 50+ partners launched A2A in April 2025, laying the groundwork for interoperable AI agent stacks.
- MCP’s July 28, 2026 overhaul ditched handshakes and session headers, making the protocol stateless and truly production-ready for massive enterprise use.
- Enterprises like Microsoft, Dell, and Datasite saw real savings: one-off integrations became unsustainable as token costs soared and scaling demanded standardized agent stacks.
Stateless Protocols Power Scale
MCP’s stateless overhaul and massive SDK adoption turned open protocols from technical experiments into the backbone of enterprise AI infrastructure.
The stack’s late-2026 production turn rested on foundations laid earlier: on November 25, 2024, Anthropic “publicly released and open-sourced the Model Context Protocol (MCP) to standardize two-way connections between AI assistants and external data sources,” explicitly displacing bespoke point-to-point integrations with a common layer. Google extended that logic on April 9, 2025, when it and “over 50 partners announced the Agent2Agent Protocol (A2A) as an open standard enabling agents from different frameworks and vendors to securely discover, communicate, and coordinate,” establishing interoperable building blocks before the ecosystem was ready to scale.
What changed by September and October 2026 was not the idea of open protocols but their operational maturity: The Tech Buzz tied MCP’s production readiness to usability improvements that made enterprise implementation “significantly less technical,” while July 28, 2026 delivered MCP’s biggest revision since launch. That overhaul made the protocol core stateless—the “2026-07-28 specification removes the initialize/initialized handshake and the Mcp-Session-Id header entirely,” so “any server instance can now handle any request”—and, with close to half a billion monthly SDK downloads, cross-vendor contributions, and a formal 12-month deprecation policy, it looked like managed infrastructure rather than an experiment.
Integration Costs Reach Breaking Point
As token usage soared and custom connectors collapsed under real-world demands, standardized protocols became essential for both technical feasibility and economic survival.
By 2026, the question was no longer whether models could shine in demos, but whether they could survive real enterprise systems. MarTech for Humans said that once AI was connected to data warehouses, CRMs, and workflow tools, things slowed down, permissions got fuzzy, and auditability decreased, while an August interview said broadscale rollout was blocked mainly by security and governance when wholesale access was granted.
The economics pushed the same way, making reusable protocol layers look better than one-off connectors. Microsoft cited a super compelling data point showing Dynamics 365 users could automate anywhere from 600 to 650,000 actions with less development work; Datasite said it built the MCP server in 4 weeks after whiteboarding and discussions; Duco cut a reconciliation build from two days to four hours; and Dell, starting from 900 scattered AI projects, warned that if it spent $12 in tokens to do $0.50 of work, it would go out of business. Monthly consumption also jumped from roughly 3 million tokens to over 150 million in under a year.
