datarekha
Agents June 10, 2026

MCP won the tool-integration war — now comes the hard part

The Model Context Protocol became the USB-C of AI tools, and once OpenAI, Google, and Microsoft all adopted it the integration wars ended. 2026 is about making it enterprise-grade: stateless HTTP, a registry, apps, and long-running tasks.

7 min read · by datarekha · mcpagentsprotocolsinteroperabilitytools

For the first two years of the agent boom, every team rebuilt the same plumbing. Want your assistant to read GitHub? Write a GitHub tool. Want it in your IDE too? Rewrite the same tool in the IDE’s plugin format. Slack? A third format. The Model Context Protocol (MCP) ended that by being, in the now-standard phrase, the USB-C for AI tools: one client-server standard so any host can use any server. The intro lesson covers what it is; this post is about why the “wars” are effectively over — and what the genuinely hard 2026 work looks like.

Adoption ended the debate

A protocol only matters if everyone speaks it, and that’s what happened. OpenAI, Google’s Gemini SDK, and Microsoft Copilot Studio all shipped first-party MCP support, alongside GitHub, Vercel, VS Code, and Cursor. The open ecosystem followed: Anthropic reported more than 10,000 active public MCP servers by December 2025, the official registry held 9,652 records by May 2026, and over 15,900 GitHub repositories carry the mcp-server topic. On the usage side, the Python and TypeScript SDKs see 97M+ monthly downloads, and a 2026 enterprise survey found 41% of organizations running MCP in production. When the three biggest model vendors converge on the same wire format, there’s no war left to fight.

The 2026 spec: scale, not features

With interoperability settled, the protocol is maturing into infrastructure. The current spec shipped in November 2025, and the 2026-07-28 revision is the largest since launch — a stateless core, OAuth/OIDC-aligned auth, server-rendered UI, and a long-running Tasks extension. Tellingly, the official 2026 roadmap says they are not adding new transports this cycle — instead they’re evolving Streamable HTTP for horizontal scaling and adding server discovery via .well-known metadata. That’s a protocol growing up: making the existing transport survive load balancers and enterprise networks, rather than chasing novelty. Concretely, Streamable HTTP now carries Mcp-Method and Mcp-Name headers so gateways can route requests without parsing the body.

Two extensions worth watching

Two additions stretch MCP well beyond “call a tool, get text back.” First, MCP Apps, added in January 2026, let a tool return a rich HTML UI rendered in a sandboxed iframe inside the chat — the model can now hand back an interactive widget, not just a string. Second, the Tasks primitive standardizes long-running work with retry semantics and result expiry, and is the centerpiece of agent-to-agent communication, which puts MCP squarely in the same territory as dedicated agent-to-agent protocols — a boundary worth understanding from the agent protocols lesson.

The interesting story of 2026 isn’t a protocol war — it’s the unglamorous work of turning a clever standard into production infrastructure: identity, discovery, scaling, and governance. If you’re building agents, MCP is now table stakes; start with the MCP lesson, then learn where it fits among the other agent protocols.

Skip to content