In depth
An MCP host is the AI application that a user interacts with. It owns the conversation, runs (or connects to) the LLM, and embeds one or more MCP clients to reach external capabilities. The host's responsibilities include: rendering the UI, managing conversation state, performing LLM inference, dispatching tool calls to the correct client, and enforcing user permissions.
Hosts typically support MCP server configuration via a JSON file. Claude Desktop uses `claude_desktop_config.json`, Claude Code uses `.claude.json`, Cursor uses `.cursor/mcp.json`. The file lists servers to spawn along with their commands, arguments, and environment variables.
Hosts also handle security-sensitive decisions: when an MCP server requests filesystem roots, the host prompts the user for approval. When a server wants to invoke the LLM (sampling), the host controls whether to allow it. The host is the trust boundary for the user.
Every popular AI coding tool in 2025 is an MCP host. The ecosystem is rapidly expanding to include custom-built hosts (internal company AI copilots) using the MCP host SDK.