MCP (Model Context Protocol): what it is, who runs it and which agents use it

MCP, the Model Context Protocol, is an open standard for connecting AI applications to external systems. A host such as Claude, ChatGPT, VS Code or Cursor connects to MCP servers, and each server exposes tools, resources and prompts the agent can call. Anthropic announced MCP on November 25, 2024 and donated it on December 9, 2025 to the Agentic AI Foundation, a directed fund under the Linux Foundation co-founded with Block and OpenAI.

#What changed in the 2026-07-28 spec?

The current revision makes the core stateless. The initialize handshake and the Mcp-Session-Id header are gone, so every request now carries its own protocol version and client capabilities in _meta. A new server/discover method, which servers must implement and clients may call before any other request, advertises the supported versions, capabilities and identity a client would otherwise have learned from a handshake.

Roots, Sampling and Logging are deprecated, as are the legacy HTTP+SSE transport and OAuth Dynamic Client Registration, which gives way to Client ID Metadata Documents. A new feature lifecycle policy holds a deprecated feature in the spec for a twelve-month minimum before it becomes eligible for removal.

#What are MCP Apps and Skills over MCP?

MCP Apps let a tool point at a ui:// resource that the host renders as an interactive HTML view inside a sandboxed iframe in the conversation. The extension client matrix, read on 20 September 2026, lists support in Claude on web and Desktop, VS Code GitHub Copilot, Microsoft 365 Copilot, Goose, Postman, MCPJam, ChatGPT, Cursor, Archestra.AI and PostHog Code.

Skills over MCP is a separate extension. SEP-2640 is Final, and it lets a server publish Agent Skills through skills/list, skills/get and the existing Resources primitive. Host support trails the spec: the same matrix marks ChatGPT, fast-agent and MCP Inspector as partial and shows no client with full support.

#How do you add an MCP server to an agent?

A local server runs on your own machine. Claude Desktop reads those from claude_desktop_config.json, a file Claude on the web and Claude Cowork do not offer. A remote server is added instead as a custom connector: go to Customize > Connectors, click + and Add custom connector, then paste its URL.

Claude's help centre puts custom connectors on the Free, Pro, Max, Team and Enterprise plans, limits Free users to one, and lists Claude on web, Desktop, Cowork and mobile. The MCP Apps build guide still says custom connectors are for paid plans only, so treat the help centre as the source this page follows. For a coding agent that loads servers per project, see Claude Code, and for a worked setup, MCP servers on OpenClaw.