Advanced: run it on another machine

Most people run the MCP server on the same computer as their AI assistant: that is the Quick start, and it needs nothing on this page. Read on if the parts should live somewhere else, many agents should share one server, or you want the strongest isolation.

Which mode?

Mode	Use it for	Login	Files come back by
Local (stdio)	everything on one machine (the normal setup)	none needed	they are already on your disk
SSH	a home server and a few of your own machines	your SSH key, one line per machine in authorized_keys	download_part gives the b123d-print pull command (the whole part into parts/<name>/), or export(embed=true)
HTTP	many agents at once, eval runs, a shared service	a named token per tool, TLS from a reverse proxy	download_part: one link to a zip of the whole part, unzipped into parts/<name>/; workspaces unused for 30 days are removed

Docker runs the server in a container, either locally over stdio (the strongest isolation, and the way to get one on Windows) or as the HTTP service with Compose.