Skip to content

MCP host ​

The AI application, such as a chat app, an IDE or an agent runtime, that creates MCP clients, connects them to servers and decides what the model may use.

MCP has three roles. The host is the one a person actually uses. It owns the model, the full conversation and every consent decision. For each server it wants to use, it creates one MCP clientMCP client: The part of an AI application that holds one connection to one MCP server: it negotiates the protocol version and capabilities, then sends that server's requests and receives its replies., a connection to exactly one serverMCP server: A program that offers tools, resources or prompts to AI applications over the Model Context Protocol, locally over stdio or remotely over HTTP.. Among the host's jobs, the 2025-11-25 architecture page lists that it "Creates and manages multiple client instances", "Enforces security policies and consent requirements", "Handles user authorization decisions", "Coordinates AI/LLM integration and sampling" and "Manages context aggregation across clients".

The last job matters most for safety. One design principle of the protocol is that servers "should not be able to read the whole conversation, nor "see into" other servers". Only the host sees everything, so only the host can decide that data from one server may flow into a call to another.

Example ​

A coding assistant in an IDE is connected to a git server and an issue tracker. The IDE is the host and runs two clients. When the model asks to call create_issue with a summary of a diff, the host:

  1. checks whether that tool is allowed at all, and shows its arguments to the developer;
  2. sends the call through the issue tracker's client only;
  3. adds the result to the conversation, which the git server never sees.

If the issue tracker server then asks for user input (elicitationElicitation (MCP): An MCP server asking the person for input through the client, usually while it handles a request such as a tool call: a message with a simple form, or a URL to open for sensitive steps.) or for a model call (samplingSampling (MCP): An MCP server asking the client to run a model call on its behalf, so a tool can use the client's model without its own API key.), the request arrives on its client, but the host decides whether a person sees it and which model answers.

Common confusion ​

People often call a whole chat app "an MCP client". As shorthand that is fine, but the specification splits the duties: protocol rules (versions, headers, token audience) belong to the client connection, while consent screens, approvals and choosing the model belong to the host. A host is also not a server: a gateway can be both, serving tools to its callers while acting as a host with its own clients toward upstream servers.

In L2M ​

In a workflow, the Agent Orchestrator and Supervisor nodes play the host: the node holds the conversation, uses the model on its chat_model port and offers the model the tools of the MCP Tool nodes on its tool port. Host decisions are where L2M's controls sit, all optional: tool policy hides tools before the model sees them or holds a call for approval, a server's elicitation form goes to a person in runs someone is watching and is declined in unattended runs, and sampling stays off until an admin allows it per server (MCP server registry). The MCP gateway is L2M in the server role.

Further reading ​

This website (not the L2M product) uses Google Analytics, only after you accept cookies. Cookie settings