MCP Turns One Year Old: How It Changed Agent Integration

A year ago, Model Context Protocol was a newly published stable spec. Today it's the default way any serious agent connects to the rest of the world.

From Internal Protocol to De Facto Standard

Anthropic open-sourced Model Context Protocol in November 2024, but it was the publication of the stable 1.0 spec in June 2025 — accompanied by the launch of the Public MCP Registry — that marked the real turning point. A year after that milestone, MCP stopped being "Anthropic's protocol" and became the default integration layer that OpenAI adopted in AgentKit, Google built into Antigravity and Gemini, and Microsoft natively incorporated into Copilot Studio and Foundry.

What MCP Solves That Traditional Integration Didn't

Before MCP, every agent framework (LangChain, LlamaIndex, each provider's proprietary SDK) defined its own format for describing tools, forcing teams to rewrite the same integration — say, connecting an agent to Salesforce or a ticketing system — once per framework or model they wanted to support. MCP standardizes that layer: an MCP server describes its tools, resources, and prompts once, and any compatible client (Claude, GPT-5.5, Gemini 3, or a custom agent) can consume it with no intermediate adapters.

Adoption: OpenAI, Google, Microsoft, and the Ecosystem

OpenAI's adoption of MCP in March 2025 and Google's shortly after was the moment the protocol stopped depending on a single provider for its legitimacy. By July 2026, all four major labs (Anthropic, OpenAI, Google, and more recently Microsoft with Foundry) support MCP as a first-class tool format, and protocol governance moved to a technical committee with representation from all four companies plus independent contributors, in a model similar to other open web standards.

The MCP Registry and the Server Explosion

The Public MCP Registry passed 15,000 published servers by mid-2026, covering everything from trivial integrations (weather, web search) to serious enterprise connectors for Salesforce, SAP, Jira, GitHub, and major databases. Adoption accelerated notably once enterprise SaaS providers started officially publishing and maintaining their own MCP servers, instead of leaving that integration to the community.

Security: the Problem That Came With Scale

The ecosystem's growth brought the expected problem: malicious or poorly implemented MCP servers exposing more capability than they declare, "tool poisoning" attacks (hidden instructions in a tool's description that try to manipulate the model), and documented cases of servers requesting excessive permissions. The ecosystem's response was introducing cryptographic signatures for verified servers in the registry and stricter security guidance from Anthropic, OpenAI, and Google on how their clients should validate servers before granting them access to sensitive data.

MCP in Practice: an Example

Connection simplicity remains the protocol's central selling point: declaring a server and its tools takes a few lines, with no need to write a separate adapter for every model the company wants to evaluate.

That same configuration file works, unmodified, in Claude Code, in an agent built with AgentKit, or in Antigravity — the protocol's original interoperability promise, delivered in practice.

What's Next for the Protocol

Active discussions at the MCP technical committee for the second half of 2026 include native support for long-lived bidirectional streaming (useful for agents collaborating with humans in real time) and a standardized tool-usage billing mechanism, which would let MCP server providers charge directly per call, laying the groundwork for a monetization model similar to traditional REST APIs but native to the agent ecosystem.