What Is the Model Context Protocol (MCP)?

The Model Context Protocol (MCP) is an open standard that defines how AI applications connect to external tools and data. A company exposes a system (a database, a CRM, a document store, an internal API) once as an MCP server, and any AI application that supports MCP can then discover and use it. It plays a role for AI tools similar to what USB did for peripherals: one common plug instead of a custom cable for every pairing.

Anthropic introduced MCP in late 2024. In December 2025 it donated the protocol to the Agentic AI Foundation, a fund hosted by the Linux Foundation and co-founded with Block and OpenAI, with support from companies including Google, Microsoft and AWS (Linux Foundation). The protocol’s maintainers continue to run its technical direction under that neutral home.

How MCP Works

The specification describes three roles that exchange JSON-RPC 2.0 messages (MCP specification):

  • Host: the AI application the user works in, such as a chat assistant, an IDE or a company’s own agent.
  • Client: the connector inside the host that talks to one server.
  • Server: a service that exposes capabilities to the AI application.

Servers offer three kinds of capability:

  • Tools: functions the model can call, such as “search orders” or “create ticket”. This is function calling in a standard, portable form.
  • Resources: data the user or model can read, such as files, records or documents.
  • Prompts: reusable templates and workflows offered to the user.

Clients can in turn offer features to servers, such as elicitation, where a server asks the user for more information mid-task. Servers run locally (started by the host on the user’s machine) or remotely over HTTP, where the specification defines authorisation based on OAuth 2.1.

How the Standard Has Changed

The specification is versioned by date and has moved quickly. The current release, dated 2026-07-28, describes stateless, self-contained requests with capabilities negotiated per request, which simplifies running remote MCP servers at scale. Checked in November 2026. Optional extensions add features such as long-running asynchronous tasks and interactive user interface elements inside a conversation. Teams building on MCP should pin the protocol version they support and check the changelog before upgrading SDKs.

When to Use MCP

MCP fits when:

  • the same tools should be available to several AI applications, or to different model vendors,
  • you want to offer your product’s data to customers’ AI assistants,
  • an AI agent needs many integrations that should be maintained separately from the agent itself.

It adds little when one application calls two or three internal functions; plain function calling inside the application is simpler.

Security Considerations

The specification itself is direct about the risks: tools represent arbitrary code execution, tool descriptions should be treated as untrusted unless the server is trusted, and hosts must get explicit user consent before exposing user data or invoking tools. The protocol cannot enforce these principles; implementers must. In practice the main risks are:

  • Prompt injection through tool results and tool descriptions, since both are text the model reads.
  • Over-privileged servers that hold a broad service account instead of acting with the user’s own permissions.
  • Untrusted third-party servers installed without review, which can change behaviour after installation.
  • Data leaving through a combination of tools, for example one that reads private records and another that can post to the internet.

MCP standardises the connection between an agent and its tools. The Agent2Agent (A2A) protocol addresses a different problem: communication between agents. Both are building blocks for agent orchestration and context engineering, since every tool definition and tool result an MCP server returns takes up part of the model’s context.

MCP in Practice

An MCP server is built as a thin layer over an existing API, in Python or TypeScript, and most of the design work goes into choosing which operations to expose and how their descriptions read to a model. A server starts with read operations; a write operation is added for a confirmed use case and authenticates as the signed-in user through OAuth rather than through a shared service account. Each server is tested with the model clients that will actually call it, because a tool that passes its unit tests can still be misused by a model that misreads its description. Learn more about AI agent development and AI workflow automation.

Back to Glossary

Let's talk about your project

We'll help you harness AI and ML technologies to create innovative software solutions

File requirements: pdf, doc, docx, rtf, ppt, pptx — max 15 MB

No file selected