▌BestMCPTools.org
Back to blog

What Is an MCP Client? Hosts, Clients and the Apps That Support MCP

By Ethan Halfhide · Updated November 22, 2026

What an MCP client is, how it differs from a host and a server, what changed in the 2026-07-28 spec, and which apps (Claude, Cursor, VS Code, ChatGPT) act as MCP clients.

An MCP client is the connector inside an AI application that talks to one MCP server. The application (the "host", such as Claude Code, Cursor or VS Code) creates one client per server, and each client keeps a dedicated connection, discovers that server's tools, resources and prompts, and calls them when the model needs them.

That is the short answer. The longer one matters if you are configuring clients, comparing them, or building your own, because the protocol changed in its 2026-07-28 revision and a lot of what still ranks for "MCP client" describes the older version. This guide covers what a client does, how it differs from a host and a server, which applications ship one today, and what to check before you pick or build one.

Host, client and server: the three roles

The official MCP architecture docs define three participants:

  • MCP host: the AI application that coordinates and manages one or more MCP clients. Claude Code and Claude Desktop are the examples the spec site uses.
  • MCP client: a component that maintains a connection to an MCP server and obtains context from it for the host to use.
  • MCP server: a program that provides context to MCP clients, whether it runs on your machine or on someone else's infrastructure.

The relationship is one to one. When VS Code connects to the Sentry MCP server, it creates a client object for that connection. When it then connects to the filesystem server, it creates a second client. A host with five servers configured is running five clients.

People often say "MCP client" when they mean the whole application ("Cursor is an MCP client"). That usage is common and fine in conversation. Strictly, Cursor is the host and the clients live inside it. If the server side is new to you, start with what an MCP server is and come back.

What an MCP client actually does

Under the current spec, a client's job breaks down into five steps.

1. Connect over a transport. MCP defines two transports. Stdio runs the server as a local process and talks over standard input and output, which is how most local MCP server setups work. Streamable HTTP uses HTTP POST for client-to-server messages, with optional Server-Sent Events for streaming, and is what remote servers use. The MCP docs recommend OAuth for obtaining tokens on remote servers.

2. Discover the server. Every server must implement server/discover, which returns the protocol versions it supports, its capabilities and its identity. Calling it is optional for the client: the 2026-07-28 revision made MCP stateless, so every request carries the protocol version and the client's capabilities in its _meta field anyway.

3. List what the server offers. The client calls tools/list, resources/list and prompts/list to learn what is available. For tools, it gets a name, a title, a description and a JSON Schema for the inputs.

4. Hand tools to the model and run the calls. The host merges tools from every connected server into one registry for the model. When the model asks for a tool, the client sends tools/call to the right server and returns the result, which can be text, images, audio, resource links or structured content.

5. Handle what the server asks of it. Clients can offer one feature back to servers: elicitation, which lets a server ask the user for missing information or confirmation through elicitation/create. Clients can also open a subscriptions/listen stream to get notifications such as notifications/tools/list_changed when a server's tools change.

What changed in the 2026-07-28 revision

If you read an older explainer, three things will look different now:

  • Roots and sampling are deprecated. Older articles list roots (telling the server which folders it may use) and sampling (letting a server ask the client's model for a completion) as core client features. As of protocol version 2026-07-28 both are deprecated. The spec's guidance is to pass directories through tool parameters, resource URIs or server configuration, and to call LLM provider APIs directly instead of sampling. Both still work during the deprecation window (earliest removal is the first revision released on or after 2027-07-28), but elicitation is now the only client feature the spec does not mark as deprecated.
  • No session handshake is required. Requests are self-contained. Discovery replaces the old initialize-then-negotiate flow, and responses can carry caching hints (ttlMs, cacheScope).
  • Notifications are opt-in. A client only receives change notifications for the types it names in a subscriptions/listen request.

Real clients lag the spec. Claude's own connector docs say Claude does not yet support sampling or resource subscriptions, and each app below supports its own subset. Check the app's docs, not just the spec, before you build a server that depends on a feature.

MCP clients you can use today

These are the hosts most people mean when they search for an MCP client. The table only lists what each vendor's own documentation states as of November 2026.

Host appLocal (stdio) serversRemote serversWhere you add serversDocumented extras
Claude CodeYesYes (HTTP recommended, OAuth)claude mcp add, .mcp.json, ~/.claude.jsonLocal, project and user scopes
Claude (web, desktop, mobile)Desktop app, packaged as a desktop extensionYes, as connectorsCustom connector URL or the connector directoryMCP Apps; no sampling or resource subscriptions yet
CursorYesYes (SSE and Streamable HTTP, OAuth).cursor/mcp.json or ~/.cursor/mcp.jsonTools, prompts, resources, roots, elicitation, MCP Apps
VS Code (GitHub Copilot)Yes (command)Yes (HTTP or SSE)MCP: Add Server, server discovery, pluginsServer input requests, MCP Apps (experimental)
Gemini CLIYesYes (SSE and HTTP streaming, OAuth)mcpServers in settings.json, gemini mcp addPrompts as slash commands, resources with @
ChatGPTNot documentedYesChatGPT Plugins, then "Add custom MCP server"Also usable through the Responses API

Claude Code

Who it's for: developers who want MCP servers available in a terminal coding agent.

Key facts: add a remote server with claude mcp add --transport http <name> <url> and a local one with claude mcp add <name> -- <command>. Anthropic recommends HTTP for remote servers and marks the SSE transport as deprecated. Servers can be scoped to you in one project (local, the default), shared with your team through a .mcp.json file in the repo (project), or available in all your projects (user). Claude Code asks for approval before using project-scoped servers from a .mcp.json file. OAuth sign-in runs through the /mcp command.

Trade-offs: tool results are capped at 25,000 tokens by default (configurable with MAX_MCP_OUTPUT_TOKENS), so servers that dump large payloads need pagination. Our list of MCP servers to add to Claude Code covers the ones worth installing first.

Claude on the web, desktop and mobile

Who it's for: people who want MCP tools in the Claude apps rather than a coding environment.

Key facts: Claude connects to remote servers as connectors, using Streamable HTTP (the legacy HTTP+SSE transport still works but is being deprecated). A user or organization Owner can add a custom connector by URL, or pick one from Anthropic's reviewed directory. Local servers run in the desktop app when packaged as a desktop extension (MCPB).

Trade-offs: Anthropic documents a tool result limit of about 150,000 characters and a 240 second timeout per tool call on claude.ai and Desktop. Sampling and resource subscriptions are not supported yet. For picks that work well here, see MCP servers that work with Claude.

Cursor

Who it's for: developers who live in an AI code editor.

Key facts: put project servers in .cursor/mcp.json and global ones in ~/.cursor/mcp.json. Cursor supports stdio, SSE and Streamable HTTP, handles OAuth with the authorization code flow and PKCE, and its docs list tools, prompts, resources, roots, elicitation and MCP Apps as supported. It asks for approval before running MCP tools by default.

Trade-offs: Cursor still lists roots, which the spec has now deprecated, so expect that support to change over time. Our Cursor MCP servers list covers what to install.

VS Code with GitHub Copilot

Who it's for: teams already standardized on VS Code and Copilot.

Key facts: the MCP: Add Server command can import a server from another app or create a config from an npm package, pip package or Docker image. Servers run as a command (stdio) or over HTTP. Copilot picks relevant tools without you naming the server, servers can ask for missing information (such as a filter or date range) through input requests, and MCP Apps support is marked experimental.

Trade-offs: VS Code's docs do not mention resources, prompts, sampling or roots on the current setup page, so do not assume them.

Gemini CLI

Who it's for: developers using Google's terminal agent.

Key facts: servers go in the mcpServers block of settings.json, or you add them with gemini mcp add <name> <commandOrUrl>. It supports stdio, SSE and HTTP streaming, OAuth 2.0 for remote servers through /mcp auth, prompts exposed as slash commands, and resources referenced with @.

ChatGPT and the OpenAI API

Who it's for: people building for ChatGPT users, or calling MCP servers from code.

Key facts: OpenAI's MCP docs cover remote servers only. In ChatGPT you add one under ChatGPT Plugins with "Add custom MCP server", then enter the URL and auth details. In the API, the Responses API accepts a tool entry of type mcp pointing at a server URL.

Trade-offs: the docs do not say whether ChatGPT can run local servers, so treat it as remote only.

How to choose an MCP client

Start from where you work, then check the features your servers need.

  • You want MCP inside your coding workflow: Claude Code, Cursor, VS Code or Gemini CLI. All four support both local and remote servers.
  • You want MCP in a chat assistant: Claude or ChatGPT, using remote servers. Claude's desktop app can also run local servers packaged as extensions.
  • Your server relies on elicitation or MCP Apps: check the host's docs first. Cursor, VS Code and Claude document MCP Apps; Cursor documents elicitation by name, and VS Code documents server input requests.
  • Your team shares config: Claude Code's project scope (.mcp.json in the repo) and Cursor's .cursor/mcp.json both live in version control.
  • You need remote servers with sign-in: Claude Code, Claude, Cursor, Gemini CLI and OpenAI document OAuth, while VS Code's setup page points you to each server publisher's authentication instructions instead. Implementations differ. Our remote MCP server guide covers what to check.

Whatever client you use, the server still matters more than the client for safety. The MCP spec says hosts must get explicit user consent before invoking any tool, and that tool descriptions should be treated as untrusted unless they come from a trusted server. Our best MCP servers list and the guide on how to choose an MCP server are the next step once your client is set up.

Building your own MCP client

You need a custom client when you are building an AI application yourself rather than using one of the hosts above. The official SDKs all support building clients as well as servers. As of November 2026 the SDK page lists TypeScript, Python, C#, Go, Rust and Ruby as Tier 1, Java as Tier 2, and Swift, PHP and Kotlin as Tier 3.

The official Python tutorial (it requires Python SDK 2.0.0 or higher) shows the full loop in a few dozen lines:

  1. Open a connection with Client(stdio_client(...)). Entering the block launches the server and agrees a protocol version; leaving it shuts the server down.
  2. Call client.list_tools() and convert each tool into the name, description and input schema your model expects.
  3. Send the user's question to the model with those tools attached.
  4. For each tool call the model makes, run client.call_tool(name, input). A tool that fails does not raise an exception; it comes back with is_error set.
  5. Send the results back to the model and repeat until it answers in text. The tutorial caps this at 10 tool turns.

The same tutorial is available in TypeScript, Java, Kotlin, C#, Ruby and Rust. If you are building servers to go with your client, the MCP Python SDK is the reference implementation for Python.

Frequently asked questions

What is the difference between an MCP client and an MCP host?

The host is the AI application, such as Claude Code or VS Code. The client is the component inside it that holds one connection to one MCP server. A host creates one client per server it connects to.

Is Claude an MCP client?

Claude Code, Claude Desktop and the Claude apps are MCP hosts that create MCP clients. Anthropic's docs say Claude connects to MCP servers from claude.ai, Claude Desktop, Claude mobile and Claude Code, and that Claude implements a subset of the MCP specification.

Does ChatGPT support MCP?

Yes, for remote servers. OpenAI's docs describe adding a server in ChatGPT through "Add custom MCP server" and calling MCP servers from the Responses API with a tool of type mcp.

Which is the best MCP client?

It depends on where you work. Claude Code, Cursor, VS Code and Gemini CLI suit coding work and support local and remote servers. Claude and ChatGPT suit chat use with remote servers. Check which MCP features your servers need, because each client supports a different subset.

Do MCP clients still support sampling and roots?

Some do, but the spec no longer recommends them. Both were deprecated in protocol version 2026-07-28. Cursor's docs still list roots; Claude's docs say sampling is not supported.

Can one MCP client connect to several servers?

No. Each client keeps a dedicated connection to one server. The host runs several clients when you configure several servers. A remote server, on the other hand, typically serves many clients.

What we don't know

  • How quickly each app will drop roots and sampling or adopt the stateless discovery flow from the 2026-07-28 revision. None of the vendor pages we read gives a date.
  • Whether ChatGPT can connect to local stdio servers. OpenAI's MCP page only covers remote servers.
  • Per-app limits on the number of tools. Cursor's and VS Code's current docs do not state one.

How this guide was put together

We compared the published MCP specification and architecture docs with each vendor's own MCP documentation, read on November 22, 2026. We did not test these clients. Feature support changes often, so check the linked docs before you rely on a specific capability.

Sources

About the author

Ethan Halfhide

Ethan Halfhide is an AI technologist who works with AI agents and Model Context Protocol integrations. He curates BestMCPTools' server directory, checking each server's documentation, supported clients and setup steps.