Live on Product Hunt Intraface launched today. Less steering, more vacations. See the launch
Start for free

What is an MCP host?

An MCP host is the application that runs the Model Context Protocol on the user's side: a coding agent, an IDE or a chat app. It creates one MCP client per server, holds the conversation and the model, decides which tools the model sees, and asks you for consent before data leaves or a tool runs.

Part of Glossary: the terms behind MCP, coding agents and shared pages, defined plainly.

How it works

The host app sits between you, the model and its clients; each client reaches one server, and no server sees the conversation.language modelyouhost appthe coordinatorclient 1client 2git serverboard serverservers never seethe conversation1. asks, approves2. the conversation3. picks a tool4. routes to one client5. tools/call
  1. You configure servers in the host: a command to run locally, or a URL. The host creates a client for each and controls when it connects.

  2. The host gathers what every client learned (tools, resources, prompts) into one view, prefixing names if two servers clash, and presents it to the model.

  3. The model picks a tool. The host shows you the call, gets your approval where needed, and routes it to the right client.

  4. Results come back through the client. The host adds them to the conversation; the server never sees the rest of the history.

  5. If a server asks for a model completion or a question for you (sampling, elicitation), the host is what answers, then the client retries the call.

Source: MCP specification: architecture

MCP host versus MCP client

The specification splits the user's side in two. The host is the container and coordinator: it creates and manages client instances, controls their permissions and lifecycle, enforces security policy and consent, handles your authorization decisions, coordinates the model, and aggregates context across clients. A client is one of those instances and talks to exactly one server. So Claude Code, Cursor, Claude.ai and VS Code are hosts. The thing inside them that holds the connection to your filesystem server is a client. When you add a second server you get a second client, not a second host.

Example: one app, two servers

Say your editor is connected to a local server for git and files and a remote server for a shared board. The host has two clients. When the model asks what tools are available, the host hands it both lists in one view. Two servers may each expose a tool called search; the spec notes that whoever aggregates tools should disambiguate, for example by prefixing names with a server identifier, and warns that a server's self-reported name is not unique enough to rely on. When the model calls the board server's publish, the host routes that call to the second client only. The git server never learns that a board exists.

Why it matters

The host is where trust lives. The specification's security principles put the duties on it by name: hosts must obtain explicit user consent before exposing user data to servers, must not transmit resource data elsewhere without consent, and must obtain consent before invoking any tool. The design principle that servers should not read the whole conversation, nor see into other servers, only works because the host keeps the full history and passes each server the minimum it needs. If a host skips the confirmation step, every connected server inherits the user's full authority. That is why the host's permission settings, allow lists and "always ask" modes matter more than any single server's behaviour.

The host is also the model's gatekeeper. Sampling, where a server wants a completion from the model, goes through the host; so does elicitation, where a server wants an answer from you. Both arrive as an input_required result on the client, and it is the host that decides whether to show the question or run the completion.

Common mistakes

  • Calling the whole app a client. Common, and harmless in chat, but the spec's client is the per-server connector. Docs that say "the client must" usually mean the host's code.
  • Assuming servers share context. They do not. If the board server needs a file the git server can read, the host or the model has to carry it across; the servers cannot talk to each other.
  • Turning off confirmations wholesale. The protocol asks for a human able to deny tool calls. Approve specific tools you trust rather than every tool from every server.

FAQ

What is the difference between an MCP host and an MCP client?
The host is the application: it holds the model, the conversation, your consent and all the clients. A client is one connection to one server, created and managed by the host.
Which apps are MCP hosts?
Any app that can connect to MCP servers: Claude Code, Claude.ai, Cursor, Codex, Gemini CLI, VS Code with Copilot, Zed and others. If you can add a server to it, it is a host.
Does the MCP host send the whole conversation to servers?
No. The full history stays with the host. Each server receives only the request in front of it, and the specification says servers should not be able to read the whole conversation or see into other servers.

Stop sending HTMLover Slack.

Create an account in a minute, connect your coding agent, and make your first board.