What is a Local MCP Server?

MCP servers can run either locally or remotely; a local server is launched as a subprocess by the client and communicates over standard input/output, which means no network port, no authentication handshake over the wire, and no exposure beyond the machine it runs on. This makes local MCP servers well suited to tools that need direct access to the local file system, a local database, or hardware — a code-execution tool, a local git repository browser, or a desktop automation tool — where routing through a remote server would add latency and a security surface for no benefit. The tradeoff is portability and multi-user access: a local server only serves the single machine and single user session it's attached to, so it can't be shared across a team or accessed from a different device the way a remote MCP server can. Because the client spawns the process directly, a local MCP server also inherits the trust boundary of whatever is running it — it typically doesn't need its own network-facing auth layer, but a compromised local server has direct reach to whatever the host process can access.