Local vs remote MCP servers

When is local the right choice?

Anything that touches your working directory, your database, your git history or your local network has to run locally — a remote server simply cannot see those.

Local is also the honest choice when credentials matter. A local server reads your API key from your own environment; a remote one needs you to hand that key to a third party.

When is remote the right choice?

When the data already lives with the provider. A remote server for a SaaS product is talking to that product's own API on its own infrastructure, so routing it through your machine adds nothing.

Remote also wins on maintenance: there is no runtime to install, no version to pin, and fixes arrive without you doing anything.

What should I check before connecting a remote server?

The trade-off is trust, so the questions are about the operator rather than the code.

• Who runs the endpoint, and does the domain match the product it claims to be?
• What authentication does it expect, and what scope does that credential carry?
• Is there a privacy policy covering what it does with what you send?
• Does it need write access, or would read-only do?

常见问题

Is a remote MCP server less secure?

It is a different risk, not strictly a worse one. A remote server cannot read your filesystem, which removes a whole class of problem. But you are sending your requests and credentials to a third party, so the operator's trustworthiness becomes part of your security posture.

Can one server be both?

Yes, and many are. A project can publish an npm package you run locally and also host an endpoint. Entries here list whichever transports the publisher registered.