MCP Host

An MCP Host is the AI-powered application where the language model lives and where the user interacts. The Host is responsible for creating and managing MCP Client instances, enforcing security policies, handling user consent, and coordinating context aggregation across all connected Servers. Per the official spec, a Host can run many Clients in parallel, each with its own 1:1 session to a different Server. The Host also coordinates LLM integration and sampling, meaning it decides when to ask the model to pick a tool and routes the model's requests through the right Client.

Concrete examples of MCP Hosts in production today: Claude Desktop, Claude Code, Visual Studio Code, Cursor, Continue.dev, and JetBrains AI Assistant. The Host is the only component that talks to the LLM directly. Servers never see the model. This separation is intentional: it lets a Server expose data and tools without ever needing access to the user's LLM credentials, and lets a Host enforce one consent flow across many Servers.

You build a Host when you are creating the AI-facing application itself. Most engineers will never build a Host from scratch. They will run their integration as a Server inside an existing Host like Claude Desktop or Cursor.

• MCP architecture (official)
• Spec: client-host-server architecture

Related Terms