Model Context Protocol (MCP)

For an AI assistant to reach your calendar, your files or your company database, a separate integration used to be written for every application and every system. MCP reduces those connections to one shared language: a system is prepared once as an MCP server, and every AI client that supports MCP can connect to it in the same way. In a commonly used analogy, MCP is like a USB-C port for AI applications.

Anthropic announced the protocol as open source in November 2024, and during 2025 it gained support from other AI providers, developer tools and enterprise software. The protocol is based on JSON-RPC messages and is updated periodically through dated specification versions. RAG systems can also use MCP to reach their sources.

• Host and client: The AI application the user works in (the host) and, inside it, the client that manages the connection with each server.
• Server: A service that exposes a data source or system through MCP; it can run locally or remotely.
• Tools: Operations the model can call, such as running a search, fetching a record or starting a task.
• Resources and prompts: Content the model can read (resources) and ready-made workflow templates offered to the user (prompts).
• Transport layer: stdio for local connections and Streamable HTTP for remote servers; remote servers use OAuth for authorisation.

globalmeta.net itself is an example. A read-only MCP server runs at https://globalmeta.net/mcp. It responds over Streamable HTTP without keeping sessions and requires no authentication, because it only serves public content and performs no write operations. The server has four tools: search_site, which searches the site; list_pages, which lists pages; get_page, which returns a page as Markdown; and get_contact, which provides contact details.

The server also lists llms.txt, llms-full.txt and the Markdown version of every page as resources. Its server card is published at /.well-known/mcp/server-card.json. An assistant that supports MCP can therefore read the site's information directly and up to date, without imitating a browser.

Implementation starts with scope: which data will the server reach, and will it only read or also perform write operations such as creating records? Tools are then defined with clear names, descriptions and input schemas; whether the model chooses the right tool depends on the quality of these descriptions. stdio is chosen for local use and Streamable HTTP for remote access.

For servers with write access or access to personal data, authorisation, least privilege, audit logging and user confirmation are essential. The server is tested with tools such as MCP Inspector, and the protocol versions it supports are stated explicitly.

An API is the interface a specific service offers on its own terms; every API has different endpoints, parameters and authentication, and every integration is written separately. MCP is a shared layer placed in front of APIs: when a client connects to a server, it learns at run time which tools exist, what they do and which inputs they expect. MCP does not replace APIs; it puts existing APIs into a form that AI models can discover and use. Designing that layer is part of our AI-native system design and solutions service.