What is MCP (Model Context Protocol)?

In a short time, MCP has become the de facto standard layer between AI assistants and business systems. Claude, ChatGPT, Google Gemini and Microsoft Copilot all speak it, and the official registry lists thousands of public servers. On this page you will learn what MCP actually is, which problem it solves, who is behind it and — more importantly — what it means in practical terms for a business that wants to put AI to work.

Which problem does MCP solve?

Before MCP, every AI application needed its own bespoke integration with every tool or data source. If you wanted Claude, ChatGPT and your internal agent to all reach GitHub, Slack and your database, you built a separate connector for each combination. That is the classic N×M integration problem: N AI apps times M tools means N×M integrations to build and maintain. MCP turns that into N+M. A tool vendor builds one MCP server, and any MCP-compatible client can use it. An app builder implements one MCP client, and it can talk to any MCP server. Hence the framing that stuck: MCP is the USB-C port for AI applications — one standard connector instead of a separate plug per device.

Technically, MCP is a client-server protocol based on JSON-RPC 2.0 messages. It defines what an AI app can request from a server — tools (executable functions, such as "create a GitHub issue"), resources (readable data and documents) and prompts (predefined workflows) — and what a server can request back from the client, such as asking the user for additional input midway through a task (elicitation). How that architecture fits together, with hosts, clients, servers and transports, is explained on How does MCP work?.

Who created MCP and who governs it now?

MCP was created by Anthropic, the company behind Claude, which announced the protocol on 25 November 2024 and released it as open source right away. Important for anyone evaluating MCP for business use: Anthropic is no longer the owner. On 9 December 2025, Anthropic donated MCP to the Agentic AI Foundation (AAIF), a fund under the Linux Foundation — the same neutral home as Linux and Kubernetes. The AAIF was founded by Anthropic, Block and OpenAI, with backing from Google, Microsoft, AWS, Cloudflare and Bloomberg. That made MCP vendor-neutral infrastructure rather than "Anthropic's protocol", removing a major adoption objection at competing platforms.

Governance is formally arranged: changes are proposed via SEPs (Specification Enhancement Proposals, comparable to Python's PEPs), reviewed by a tiered group of maintainers and developed in working groups covering, among other things, transports, authentication and the registry. Spec versions are dates, not version numbers: the current stable version is 2025-11-25, and the next — the largest revision since launch — is finalised on 28 July 2026 (see below). For a business this means: MCP is not a single vendor's experiment, but an open standard with formal governance, a published roadmap and, as of recently, even a formal deprecation policy with a minimum transition period of twelve months.

How widely is MCP used?

By mid-2026, adoption on the client side is effectively universal. According to the Linux Foundation's press release (December 2025), MCP has first-class client support in ChatGPT, Claude, Cursor, Gemini, Microsoft Copilot and Visual Studio Code, and the official SDKs were already being downloaded more than 97 million times per month — growing, according to ecosystem trackers, to around 110 million per month by April 2026. The same Linux Foundation announcement reported over 10,000 active public MCP servers; the official MCP registry counted some 9,650 unique server records in May 2026. The 2026 Stacklok survey furthermore shows that 41% of software organisations run MCP servers in limited or broad production.

What stands out is how deeply Anthropic's competitors have embraced the protocol. OpenAI built its entire Apps SDK for ChatGPT apps on top of MCP and offers a full MCP client through Developer Mode. Google supports MCP in the Gemini CLI and delivers managed MCP servers for its own services such as BigQuery and Maps via Google Cloud. Microsoft has MCP generally available in Copilot Studio, Microsoft 365 Copilot and VS Code/GitHub Copilot. Perplexity, Mistral, AWS and xAI have joined as well, and companies such as Stripe, Shopify, PayPal, GitHub, Notion and Figma ship official MCP servers. You will find an overview of servers by industry in our MCP servers overview.

How is it different from a regular API or function calling?

MCP does not replace your existing APIs — it is a layer on top of them. An API is a unique interface per service: every integration requires reading its documentation, arranging authentication and writing code. An MCP server is essentially a standardised adapter that translates such a service once into the format every AI platform understands, including making the available functions automatically discoverable. Mature tooling exists to generate MCP servers from OpenAPI specifications, though a caveat is in order: a naive one-to-one translation of API endpoints into tools produces poor results for AI agents — thoughtful tool design remains essential. More on that at Building an MCP server.

Function calling (tool calling) from OpenAI, Anthropic and Google is not a competitor either, but a different layer. Function calling is a model capability: the language model produces a structured JSON call, but the developer has to arrange how tools are discovered, executed and secured — anew for every application. MCP standardises exactly that part: discovery, transport, authentication (OAuth) and the lifecycle. In practice they work together: an MCP client converts a server's tool list into the model's function-calling schema, and then executes the model's call via MCP. The fact that OpenAI now routes its own ecosystem through MCP effectively settles the debate: it is "both, at different layers".

Finally, A2A (Agent2Agent Protocol): created by Google in April 2025 and, since June 2025, also housed at the Linux Foundation. A2A governs communication between AI agents themselves, whereas MCP governs the connection between an agent and its tools and data. The consensus stack taking shape: MCP for tools, A2A for agent-to-agent. According to the Linux Foundation (April 2026), A2A is backed by more than 150 organisations. An earlier attempt at an agent-tool standard, agents.json from Wildcard AI, has been effectively abandoned since early 2025 and superseded by MCP.

What does MCP mean in practice for your business?

For a business, MCP concretely means three things. First: your AI assistant can reach your systems. With an MCP server for your CRM, accounting software or document management, an employee can ask Claude or Copilot questions about live business data and have actions carried out, without building a separate AI integration for every system. Second: you are not investing in a dead end. Because the protocol is governed vendor-neutrally and all major platforms support it, an MCP server remains usable if you later switch AI vendors. Third: you can choose from a growing catalogue. Official servers already exist for a lot of standard software; see for example our sector pages for accounting, insurance or CRM and sales.

What changes on 28 July 2026?

Anyone getting started now is doing so at an interesting moment. On 28 July 2026, the final version of the newest spec revision arrives, described by the maintainers themselves as the largest since launch. At its core: the protocol becomes stateless, so MCP servers behave like ordinary scalable web services behind a load balancer — without sessions or handshakes. In addition, a formal extension mechanism is introduced, with MCP Apps (servers that can render their own interface inside the chat) and Tasks (long-running jobs) as the first extensions. Authorisation is tightened further, and several older components (roots, sampling, logging) are formally deprecated with at least twelve months of transition time. The release candidate has been frozen since 21 May 2026 and the major SDKs are delivering support within the validation window. For businesses, the key message is this: MCP is deliberately maturing towards enterprise requirements — scalability, audit trails and SSO are explicitly on the official 2026 roadmap.

Further reading

Frequently asked questions

What is MCP in one sentence?

MCP (Model Context Protocol) is an open standard that governs how AI applications such as Claude, ChatGPT and Copilot securely connect to external tools and data sources — often summarised as "the USB-C port for AI".

Who created MCP?

MCP was created and released as open source by Anthropic on 25 November 2024. Since 9 December 2025 the protocol has been owned by the Agentic AI Foundation under the Linux Foundation, co-supported by OpenAI, Google, Microsoft, AWS and Block.

Is MCP the same as an API?

No. An API is a bespoke interface per service; MCP is one standardised protocol that sits on top of existing APIs. An MCP server translates a service into the standard format once, after which any MCP-compatible AI client can use it.

Does MCP only work with Claude?

No. Although Anthropic (the company behind Claude) invented the protocol, by mid-2026 all major AI platforms support MCP: ChatGPT, Google Gemini, Microsoft Copilot, Cursor, VS Code and many other clients.

What is the difference between MCP and function calling?

Function calling is a model capability: the language model generates a structured call. MCP standardises everything around it — tool discovery, transport, authentication and security. They work together at different layers, not against each other.

Is MCP secure enough for business use?

MCP carries serious risks such as prompt injection and malicious servers, but the standard now includes hard security requirements (OAuth, resource indicators) and the NSA and CISA published official security guidance in June 2026. Read our security page for the details.

How many MCP servers exist?

The official MCP registry counted roughly 9,650 unique public servers in May 2026; in late 2025 the Linux Foundation reported over 10,000 active public servers. In addition, many companies build internal servers that are not publicly registered.

Last updated: 2026-07-06