Model Context Protocol (MCP)

Before MCP, connecting an AI agent to your private data or specific software tools was a fragmented, custom-built process. Every developer had to write unique “glue code” to make an AI talk to a database, a file system, or a project management tool. The Model Context Protocol (MCP) was introduced on November 25, 2024, by Anthropic to solve this by providing a universal language for these connections. Now hosted by the Agentic AI Foundation (AAIF) under the Linux Foundation, MCP acts as the standard “plug” that allows any AI application to communicate with any data source or tool, regardless of who built them.

Think of MCP like the USB standard for the AI era. Before USB, computers had a chaotic array of proprietary ports for printers, mice, and keyboards. USB standardized the connection, meaning you could plug any device into any computer and it would just work. MCP does the same for AI: it ensures that an AI application (the client) can reliably “plug into” a data source or tool (the server) without needing custom integration for every single combination.

Client-Server Model

The architecture is straightforward: MCP uses a client-server model. The MCP client is the AI application—such as Claude Desktop, ChatGPT, Cursor, or VS Code—that wants to access information or perform actions. The MCP server is a lightweight piece of software that exposes specific data or tools to the client. For example, you might use an MCP server for GitHub to let your AI search your repositories, or one for Postgres to let it query your database. Because it is a standardized protocol, you don’t have to rebuild these connections when you switch from one AI tool to another.

Stateless Architecture (July 2026)

On July 28, 2026, MCP underwent its largest specification revision since launch. The protocol became fully stateless: the initialize/initialized handshake and the Mcp-Session-Id header were retired (SEP-2575, SEP-2567). This change enables horizontal scaling on ordinary HTTP load-balanced infrastructure without session affinity, making MCP practical for enterprise-grade deployments.

Core Primitives

To make this work, MCP defines six core “primitives” that govern how information flows:

• Server-side primitives: Tools (functions the AI can call, like “send email”), Resources (read-only data the AI can read, like a file or database record), and Prompts (pre-defined templates to guide the AI’s behavior).
• Client-side primitives: Sampling (the server asks the AI to complete a task), Roots (defining the boundaries of the file system the AI can access), and Elicitation (the server asks the user for input).

The protocol supports two transport mechanisms: stdio for local processes (like an AI running on your laptop) and Streamable HTTP for remote connections. Header-based routing (Mcp-Method/Mcp-Name, SEP-2243) enables efficient multiplexing. There are 10 official SDKs across major languages including Python, TypeScript, Go, and Rust, with beta SDKs for the July 2026 spec shipped for TypeScript, Python, Go, and C#.

New Capabilities (2026)

The July 2026 spec added several capabilities:

• Multi Round-Trip Requests: Servers can now request multiple sequential LLM completions within a single tool call, enabling complex workflows.
• Cacheable List Results: Resource and tool listings can be cached, reducing redundant server queries.
• Extensions Framework: A formal mechanism for protocol extensions beyond the core primitives.
• OAuth/DCR Hardening: Improved security for remote MCP servers with dynamic client registration.

Ecosystem and Governance

Adoption has been rapid. By December 2025, there were over 10,000 active public MCP servers, and 97 million monthly SDK downloads across Python and TypeScript alone. By May 2026, the official MCP registry listed 9,652 records. The 2026 Stacklok State of MCP report found that 41% of surveyed software teams were already using MCP in production. Major platforms including ChatGPT, Cursor, Gemini, Microsoft Copilot, and Visual Studio Code have all integrated MCP support.

In December 2025, MCP moved under the Agentic AI Foundation (AAIF), a Linux Foundation project. The MCPA certification program, launched September 13, 2026 under AAIF, provides vendor-neutral certification for MCP implementations. The ID-JAG enterprise authorization preview (September 9, 2026) offers Keycloak-based authorization with upcoming EMA support.

Whether you are connecting to Google Drive, Slack, GitHub, or local Git repositories, the goal remains the same: creating modular, interoperable agent infrastructure where AI agents can access the context they need to be truly useful.

Frequently Asked Questions

How does MCP differ from a standard API?

A standard API is a specific interface defined by one provider for one service. MCP is a meta-protocol—a standard way for any AI application to discover and use any compatible server’s tools and data, regardless of who built the server.

Do I need to be a developer to use MCP?

End users interact with MCP indirectly. When you connect tools like GitHub or Slack to an AI application like Claude Desktop, you are often using MCP under the hood. Building your own MCP server does require programming skills.

Can I build my own MCP server?

Yes. There are 10 official SDKs (Python, TypeScript, Go, Java, Kotlin, C#, PHP, Ruby, Rust, Swift) and pre-built templates to help developers expose their own data or tools via MCP.

Is MCP limited to specific AI models?

No. MCP is model-agnostic. Any AI application that implements the client-side protocol can connect to any MCP server. Claude, ChatGPT, Gemini, and Microsoft Copilot all support it.

What changed in the July 2026 stateless update?

The July 28, 2026 spec (SEP-2575, SEP-2567) made MCP fully stateless by removing the initialize/initialized handshake and Mcp-Session-Id header. This enables horizontal scaling on standard HTTP load balancers without session affinity, making MCP practical for enterprise deployments. The update also added header-based routing, multi round-trip requests, cacheable list results, an extensions framework, and OAuth/DCR hardening.