Deutsch
Agent protocols

MCP (Model Context Protocol)

Also known as: Model Context Protocol

MCP (Model Context Protocol) is an open standard that lets AI applications such as Claude or ChatGPT connect to external data sources, tools, and workflows. A service exposes its capabilities once as an MCP server, and any compatible AI application can use them. The project’s own documentation compares MCP to a USB-C port for AI applications.

At a glance

Publisher
Agentic AI Foundation (Linux Foundation), originally Anthropic
Connects
AI applications with servers for data, tools, and prompts
Status
open standard, current version 2026-07-28
Transport
stdio (local), Streamable HTTP (remote)
Since
Nov 25, 2024
Specification
MCP Specification 2026-07-28

Last checked:

How does MCP work?

The specification defines three roles. The host is the AI application that opens connections, for example a chat assistant or a code editor. The client is the connector inside the host that talks to one server. The server is the service that provides data and functions.

A server can offer three kinds of features:

  • Resources: context and data, such as files or database records.
  • Tools: functions the model can call, for example “run a query”.
  • Prompts: prepared message templates and workflows for users.

Messages are encoded as JSON-RPC 2.0. The specification defines two standard transports, stdio for servers launched locally by the client and Streamable HTTP for remote servers. It also states that users must explicitly consent before a host shares their data with a server or invokes a tool. The protocol itself cannot enforce this, so consent flows are the job of each application.

MCP: host, clients and servers, plus WebMCPAn AI application acting as host starts one MCP client per server. Each server offers tools, resources and prompts. Local servers connect via stdio, remote servers via Streamable HTTP. Below, WebMCP: a web page registers tools in JavaScript, and an agent in the browser calls them.HOSTMCP SERVERSHostAI application, e.g. a chat appMCP client 1stdioServer 1 · fileslocalToolsResourcesPromptsMCP client 2stdioServer 2 · databaselocalToolsResourcesPromptsMCP client 3Streamable HTTPServer 3 · external APIremoteToolsResourcesPromptsWEBMCP · DRAFT OF THE W3C WEB MACHINE LEARNING COMMUNITY GROUPAgent in the browsere.g. the browser’s agentcalls a toolWeb pageregisters JavaScript tools that run in the browserSchematic. As of MCP specification 2026-07-28 and WebMCP draft of Oct 2, 2026.
According to the MCP specification, a host starts one client per server. Servers offer tools, resources and prompts and run locally (stdio) or remotely (Streamable HTTP). WebMCP is a draft by a W3C community group, not a W3C standard. Here the web page provides its tools inside the browser. Sources: Model Context Protocol: Specification 2026-07-28; Model Context Protocol: Architecture; Model Context Protocol: Transports; W3C Web Machine Learning Community Group: WebMCP (draft of Oct 2, 2026).

Where does MCP come from?

Anthropic released MCP as an open standard on November 25, 2024. On December 9, 2025, Anthropic donated the protocol to the Agentic AI Foundation, a directed fund under the Linux Foundation co-founded by Anthropic, Block, and OpenAI. The specification is versioned by date, and newer revisions have changed parts of the base protocol. The July 28, 2026 revision made MCP stateless. It removed the initialize handshake, so every request now carries its own protocol version and client capabilities.

AI agent protocols compared
ProtocolPublisherConnectsStatus and version
MCP (Model Context Protocol) (this term)by Anthropic, since 2025-12-09 at the Agentic AI Foundation (Linux Foundation)AI applications with servers that provide data, tools and promptscurrent version 2026-07-28
A2A (Agent2Agent)by Google, donated to the Linux Foundation; joining the Agentic AI Foundation announced on 2026-08-27AI agents from different vendors with each otherspecification 1.0 (release v1.0.1 of 2026-05-28)
ACP (Agentic Commerce Protocol)OpenAI and Stripebuyers and their AI agents with businesses and payment providersbeta, current version 2026-04-17
AP2 (Agent Payments Protocol)Google; standardisation continues in FIDO Alliance working groupsAI agents, merchants and payment providers for payments on behalf of usersv0.2 (release of 2026-04-28)
UCP (Universal Commerce Protocol)Google with commerce partners (incl. Shopify, Etsy, Wayfair, Target, Walmart)platforms and AI agents with businesses (profile at /.well-known/ucp)stable, version 2026-08-25
WebMCPW3C Web Machine Learning Community Group (editors from Microsoft and Google)websites with AI agents in the browserdraft (Draft Community Group Report of 2026-10-02), not a W3C standard; origin trial in Chrome from version 149
NLWebMicrosoft (announced 2025-05-19), open source under the MIT licencewebsites (Schema.org and RSS content) with users and AI agents; every instance is an MCP serveropen-source project without a published version number

Versions and status as stated in each specification or on the publisher's site, checked on 8 October 2026. These protocols change quickly. MCP, ACP and UCP use date-based versions (YYYY-MM-DD), A2A and AP2 use version numbers. Sources: Model Context Protocol: Specification 2026-07-28; Model Context Protocol: Versioning; Anthropic: Donating the Model Context Protocol and establishing the Agentic AI Foundation (2025-12-09); A2A Protocol: Specification; A2A: A new chapter for A2A, joining the Agentic AI Foundation (2026-08-27); A2A releases (GitHub); Agentic Commerce Protocol (GitHub); AP2: Agent Payments Protocol; AP2 releases (GitHub); Universal Commerce Protocol: Specification overview (2026-08-25); Universal Commerce Protocol; WebMCP: Draft Community Group Report (W3C Web Machine Learning CG); Chrome for Developers: WebMCP (updated 2026-10-07); Microsoft: Introducing NLWeb (2025-05-19); NLWeb (GitHub). Last checked: .

Example

An SEO platform offers an MCP server. A user connects it to an AI assistant and asks: “Which URLs changed their status code since the last crawl?” The assistant calls the matching tool on the server, receives the list, and summarizes it. Before the call runs, the assistant asks the user to approve the tool.

Common mistakes

  • Passing tokens through. The specification says an MCP server must not accept tokens that were not issued for it. Forwarding a client’s token unchecked to a downstream API (“token passthrough”) bypasses that API’s security controls.
  • Leaving a local server open. For servers that run on the user’s machine, the specification recommends stdio. A local server on HTTP should require an authorization token or similar restriction, or other processes can reach it.
  • Requesting overly broad scopes. The specification lists wildcard scopes such as * or full-access as a common mistake. A stolen token with broad scopes also unlocks tools that have nothing to do with the task.
  • Using the old transport. The earlier HTTP+SSE transport is deprecated. New servers should use Streamable HTTP.

FAQ

What is the difference between MCP and an API? An API is the interface of a single service, with its own format and authentication. MCP defines one common way for AI applications to discover and call such services. Many MCP servers call an existing API internally.

What is the difference between MCP and RAG? Retrieval-augmented generation (RAG) is a technique for adding retrieved documents to a model’s prompt. MCP is a protocol for connecting an AI application to outside systems. An MCP server can serve as the data source for a RAG setup, and it can also expose actions that go beyond retrieval.

Which applications support MCP? The MCP documentation lists Claude, ChatGPT, Visual Studio Code, and Cursor, among others.

How does MCP relate to websites? MCP connects AI applications to services. WebMCP is a separate proposal that applies MCP’s vocabulary to websites, so a page can offer tools to AI agents running in the browser.

Sources

  1. What is the Model Context Protocol (MCP)? – modelcontextprotocol.iomodelcontextprotocol.io
  2. MCP Specification (version 2026-07-28) – modelcontextprotocol.iomodelcontextprotocol.io
  3. MCP Specification 2026-07-28, Transports – modelcontextprotocol.iomodelcontextprotocol.io
  4. Introducing the Model Context Protocol – Anthropic (Nov 25, 2024)anthropic.com
  5. Donating the Model Context Protocol and establishing the Agentic AI Foundation – Anthropic (Dec 9, 2025)anthropic.com
  6. Lewis et al.: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (2020) – arXivarxiv.org
  7. WebMCP – Draft Community Group Report, W3C Web Machine Learning Community Groupwebmachinelearning.github.io
  8. MCP Specification 2026-07-28, Key Changes – modelcontextprotocol.iomodelcontextprotocol.io
  9. MCP Specification 2026-07-28, Security Best Practices – modelcontextprotocol.iomodelcontextprotocol.io

Last updated: