MCP is the de facto standard for AI agents

What seemed like a niche protocol pushed by Anthropic until a year ago has officially become the de facto standard for connecting LLMs and AI agents to tools, data, and external systems: Model Context Protocol (MCP).

In recent weeks, OpenAI, Google, and Meta have all announced native support for MCP in their respective APIs. This means that an MCP server written once works with Claude, GPT, Gemini, and Llama without modifications — a monumental shift for anyone building integrations.

The ecosystem is exploding: the mcp.so public registry counts over 4,000 servers, with coverage ranging from Salesforce and HubSpot to more vertical integrations such as CAD tools, video editing pipelines, or trading systems.

For builders, the advice is clear: stop writing custom wrappers for each LLM and adopt MCP as an abstraction layer. Portability between models has become a requirement, not a nice-to-have, especially in a market where prices and capabilities change every quarter.

The next expected step is the standardization of MCP at the IETF, which is already at an advanced stage: once finalized, the protocol will become part of the web infrastructure just like HTTP or WebSocket.