The host is the program that loads your MCP server and exposes its tools to a model — Claude Desktop, Cursor, ChatGPT, an in-product widget, or a custom agent loop. The server is your code; the host is whatever calls it.

Can my server work in multiple hosts at once?

Yes, that's the point of MCP. A server you write once works in Claude Desktop, Cursor, ChatGPT (where supported), and your own widget without changes. Hosts differ in UI, transport defaults, and which capabilities they implement.

Which host should I pick for in-product AI?

If you're embedding AI inside your own SaaS, the host is your widget — full control over UX, auth, and metering. Use Claude Desktop or Cursor as a host for internal/dev workflows, not customer-facing surfaces.

Do hosts charge to run my MCP server?

Self-hosted (you embed a widget) is your own infra cost. Third-party hosts like Claude or Cursor don't charge for connecting MCP servers, but model inference cost still falls on whoever owns the API key.

Choosing an MCP host: widget, Claude, Cursor, or custom

An MCP server is your code; an MCP host is whatever loads and runs it — Claude Desktop, Cursor, ChatGPT, an in-product widget, or a custom agent. The same server works in all of them. Picking the right host is per use case: customer-facing in-product AI runs in your own widget for full UX and auth control; internal workflows run in Claude or Cursor; programmatic agents run in a custom host using the MCP SDK.

Host vs server, in one paragraph

Your MCP server is the code that exposes tools and resources. The hostis the program that loads it and presents the tools to a model — Claude Desktop, Cursor, ChatGPT (where supported), an in-product widget, or a custom agent loop. Same server, many hosts.

Picking the right host for the job

Customer-facing in-product AI — your own widget. You need full control of UX, auth, metering, and brand.
Internal team workflows — Claude Desktop or Cursor. Good UX, no ship cost.
Programmatic agents / pipelines — a custom host using the MCP SDK in your worker.
Public discoverability — list your server in MCP directories so other hosts can pick it up.

Host capability differences

Not every host implements every part of the spec. Check before relying on:

Resources — most hosts support tools; resource support is patchier.
Prompts — only some hosts expose your prompts in the UI.
Sampling — server-initiated model calls; rarely supported, design around it.
Streamable HTTP — newer transport; older hosts may still require SSE.

Auth across hosts

For a customer-facing widget, mint short-lived JWTs per end-user and pass them in. For third-party hosts like Claude Desktop, use OAuth (the spec's Authorizationsection) so users grant access through a standard flow. Never share a long-lived API key across hosts.

When Command+K is your host

For customer-facing in-product AI, Command+K is the host: it embeds your MCP server (or one you converted from a REST API), runs the assistant as the logged-in user, meters usage per end-user, and feeds events into your Stripe billing. You write tools once; Command+K handles transport, auth, metering, and UI.

See Build MCP for the host options.

FAQ

What's an MCP host?: The host is the program that loads your MCP server and exposes its tools to a model — Claude Desktop, Cursor, ChatGPT, an in-product widget, or a custom agent loop. The server is your code; the host is whatever calls it.
Can my server work in multiple hosts at once?: Yes, that's the point of MCP. A server you write once works in Claude Desktop, Cursor, ChatGPT (where supported), and your own widget without changes. Hosts differ in UI, transport defaults, and which capabilities they implement.
Which host should I pick for in-product AI?: If you're embedding AI inside your own SaaS, the host is your widget — full control over UX, auth, and metering. Use Claude Desktop or Cursor as a host for internal/dev workflows, not customer-facing surfaces.
Do hosts charge to run my MCP server?: Self-hosted (you embed a widget) is your own infra cost. Third-party hosts like Claude or Cursor don't charge for connecting MCP servers, but model inference cost still falls on whoever owns the API key.