MCP won: how agent tooling standardized in eighteen months

The least glamorous problem in agent-building is also the most expensive: every model has to be wired to every tool and data source, and everyone was doing that wiring differently. The Model Context Protocol (MCP) is the boring standard that fixed it — and its adoption curve over the last year and a half is one of the fastest I’ve seen for a piece of infrastructure.

From one lab’s idea to shared infrastructure

MCP started as an Anthropic open standard for connecting AI models to tools and data. In December 2025 Anthropic donated it to the newly formed Agentic AI Foundation under the Linux Foundation — co-founded with Block and OpenAI and backed by Google, Microsoft, AWS, Cloudflare, and Bloomberg. By then it had already gone from internal experiment to something adopted across ChatGPT, Cursor, Gemini, Copilot, and VS Code, with a November 2025 spec update (async operations, statelessness, an official server registry) and more than 10,000 public MCP servers.

Donating a protocol to a neutral foundation is the clearest signal a company can send that it wants the thing to outlive its own roadmap. For a standard whose entire value is that everyone uses it, that neutrality is the product.

The tooling followed. NVIDIA’s NeMo Agent Toolkit is MCP-compatible as both client and server and works across LangChain, LlamaIndex, CrewAI, and Google ADK. Anthropic’s Claude Agent SDK generalized the agent loop behind Claude Code — file access, shell, web, MCP calls — into a library. The primitives converged.

Why a protocol is worth caring about

A standard doesn’t make your agent smarter. What it does is collapse an integration matrix. Instead of M models × N tools bespoke connectors, you build one MCP server per tool and one MCP client per app, and everything composes. It turns “does this agent support my tool?” from an engineering project into a configuration line.

I’ve felt this directly across two projects that sit on opposite ends of MCP:

• Orca exposes its teamwork-graph reasoning as an MCP server, so any MCP-aware agent can query the graph without knowing anything about how it’s built. The graph is the product; MCP is just one of the doors into it.
• Blinq is the other side — an MCP client. It talks to Notion’s hosted MCP server over OAuth to read, search, and update your workspace, with no local server and no token to copy-paste.

Neither of those integrations required me to invent a transport, an auth flow, or a tool schema. That’s the whole point. Standards are invisible when they work, and MCP is getting close to invisible.