The problem MCP came to solve

Before MCP, connecting an AI model to an external data source (a database, a CRM, a file system, a third-party API) required writing a custom connector for each combination of model and tool. With N models and M tools, that means N×M different integrations to maintain — the classic "N by M integration problem" already solved before in other software domains.

Anthropic introduced the Model Context Protocol in November 2024, created by engineers David Soria Parra and Justin Spahr-Summers, reusing messaging ideas from the Language Server Protocol (LSP) — the same standard that lets any code editor talk to any programming language without ad-hoc integrations.

How it works, in simple terms

MCP defines a common protocol: any "MCP server" exposes a set of tools, resources and prompts in a standardized way, and any "MCP client" (the AI model or the agent orchestrating it) can discover and use those capabilities without needing specific code for each combination.

In practice, this means that a single MCP server built to expose, for example, access to a Postgres database works just as well with Claude, with GPT, or with any other model that implements the protocol on the client side — without rewriting anything.

Everyday example: at VURAOS, Claude Code runs with multiple connected MCP servers (Supabase, AWS, Stripe) that give it direct access to operations on those systems without each integration having to be coded from scratch — exactly the kind of use case MCP was designed to solve.

Adoption was surprisingly fast for a new standard

  • OpenAI officially adopted MCP in March 2025, integrating it into the ChatGPT desktop app, and by September 2025 it already supported MCP in ChatGPT Apps for third-party access.
  • Google DeepMind joined in April 2025.
  • Microsoft integrated it into Semantic Kernel and Azure OpenAI.
  • Salesforce began routing interactions from its Headless 360 platform through MCP in April 2026, processing 4.5 million calls by the end of May that same year.

In December 2025, Anthropic donated MCP to the Agentic AI Foundation (AAIF), a directed fund within the Linux Foundation — a strategic move signaling that Anthropic prioritizes mass adoption of the standard over keeping exclusive control of it, betting that an open ecosystem grows faster than a closed one.

Why this matters today, not just for developers

In April 2026, the Agentic AI Foundation hosted the MCP Dev Summit in New York, with about 1,200 attendees — a clear sign that MCP stopped being a single lab's technical experiment and consolidated as industry infrastructure, similar to what HTTP meant for the web or SMTP for email.

For any company evaluating building AI agents with access to internal systems, the relevant question today is not "should we use MCP?" but "which internal tools still don't have an MCP server available, and is it worth building one?" — the conversation shifted from whether to adopt the standard to how to make the most of it.