An SSH MCP server is an MCP server whose tools run commands on a remote machine over SSH. Connect one to Claude, Cursor or another AI client, and the model can ask for df -h on your web server and read the answer without you typing anything.
It's one of two common ways to put an AI agent to work on a server. The other is to run the agent on the server itself and reach it over SSH, which is how Shell Drop is used. This page covers both, the security trade-offs of giving an agent SSH access, and what Shell Drop's own MCP server is (it is not an SSH MCP server).
MCP, the Model Context Protocol, is an open standard for connecting AI applications to outside tools and data. An MCP server offers a set of tools, each with a name, a description and the inputs it takes. An MCP client, such as Claude, Claude Code or Cursor, lists those tools to the model; when the model decides to use one, the client calls the server and hands the result back. A server can run on your own computer or as a remote service at a URL.
An MCP server whose tools reach a machine over SSH. A typical one offers tools along the lines of "run a command on this host", sometimes with file upload and download. The server holds or borrows SSH credentials, opens the connection, runs what the model asked for and returns the output.
The model never touches SSH directly. It sees tool names and results, and the MCP server is where the access really lives: which hosts it can reach, as which user, with which key, and whether it runs any command or only some.
| SSH MCP server | Agent on the server, reached over SSH | |
|---|---|---|
| Where the agent runs | On your computer or in a cloud AI app | On the server, for example Claude Code or Codex inside Tmux |
| How it reaches the machine | MCP tool calls, turned into SSH commands | It's already there, with the permissions of the user it runs as |
| Where you watch it | In the AI app's conversation | In the terminal session, from any SSH client |
| Long jobs | Bounded by the AI app's session and each tool call | Keeps running in Tmux after you disconnect |
| What has SSH credentials | The MCP server | You, to log in and check on it |
An SSH MCP server suits quick, bounded jobs across several machines: checking disk space on five servers, reading a log, restarting a service. Running the agent on the server suits long work, a coding agent that builds, tests and edits for an hour, because the session lives where the work is and keeps going when you close the laptop or lock the phone.
The second model is the one Shell Drop is built around. Claude Code or Codex runs in a Tmux session on a Mac mini, a Linux box or a VPS, and you attach from an iPhone or iPad to read what it did and type the next instruction. Claude Code on iPhone and Codex on iPhone walk through that setup.
Either way, an agent with a shell can do what that account can do. A model that misreads an instruction, or follows one planted in a file or log it reads, acts with the same reach. The general precautions:
authorized_keys, options such as from="10.0.0.5", no-port-forwarding, no-agent-forwarding and command="..." narrow what that key can do. SSH keys on iPhone covers keys per device and per purpose.sudoers rather than all of them.None of this is specific to MCP. The same applies to an agent running in Tmux on the server: it has whatever access the user who started it has.
The Shell Drop MCP connector is not an SSH MCP server. It never connects to any machine and can't run a command anywhere.
It's a free, read-only help server at https://mcp.shelldrop.sh/mcp, with no account and no sign-in. It gives Claude and other AI tools step-by-step help for running coding agents on a Mac or Linux machine and reaching them from an iPhone or iPad. Its four tools:
~/.tmux.conf and how to apply it.Every tool returns text and never connects to or changes any machine. The server answers from fixed text and keeps nothing. Your AI client reads the answer, and you decide what to run. The MCP connector page has the endpoint and the setup lines for Claude, Claude Code and Cursor.
An MCP server whose tools run commands on a remote machine over SSH. An AI client such as Claude or Cursor calls a tool like “run this command on that host”, the MCP server makes the SSH connection, and the output goes back to the model.
No. It is a read-only help server: a setup guide, Tmux how-tos, a tmux.conf generator and connection troubleshooting. Every tool returns text and never connects to or changes any machine.
It can be, with limits: a separate user with only the permissions the job needs, a key used only by the agent, no broad sudo, and approval for anything that changes the system. Treat everything the agent reads on the server as possibly carrying instructions.
No. Claude Code and Codex can run on the server itself, inside Tmux, with the access of the user you start them as. You then reach that session over SSH from a laptop, an iPhone or an iPad.