MCP SSH Proxy

A human-in-the-loop SSH bridge for AI agents. It exposes an MCP server with two tools — ssh_list_hosts and ssh_exec — but every command an agent wants to run is held for your review in a desktop app before it touches a server. You see the exact command and the target host, then click Run, edit it, or Reject it.

Servers are configured only in the app's UI — nothing is hardcoded, and credentials are encrypted at rest with the OS keystore (Windows DPAPI via Electron safeStorage).

Sibling project to MCP SQL Proxy — same approval-loop architecture, applied to SSH instead of SQL.

Features

• 🔐 Human-in-the-loop approval — every agent command is shown for review (Run / edit / Reject) before it runs; the result is returned to the agent automatically.

• ⚙️ Per-command confirm modes — a global confirm-each ⟷ auto-run switch, overridable per server (always confirm / auto).

• 🖥️ Built-in interactive terminals — full PuTTY-style xterm.js shells, multiple servers at once in tabs; type freely, see colors/vim/live output. Agent commands are echoed into the matching server's terminal as a labelled blue bar.

• 🗂️ Multi-server aware — pending requests are scoped to the selected server with a ! badge on the others; idle target servers are auto-connected on demand.

• 📤 SFTP file transfer — ssh_upload / ssh_download move files to/from a server (inline content or a local path), each approved in the app.

• 🔑 Any key format — OpenSSH, PEM/PKCS#8, and PuTTY .ppk (v2 & v3) are detected and converted automatically; key, password, or ssh-agent auth.

• 📥 PuTTY session import — pull saved sessions (host/port/user/key) straight from the Windows registry.

• 🕑 Persistent per-server history — every executed command is stored across sessions.

• 🛑 Danger mode — destructive commands (rm, dd, shutdown, sudo, systemctl stop, redirects into /, …) flash the window red.

Related MCP server: TunnelAdmin

Why

Letting an agent run arbitrary shell commands on a production server is risky. This proxy keeps a person in the loop: the agent proposes, you approve. Destructive commands (rm, dd, shutdown, sudo, systemctl stop, redirects into /, …) light the whole window up in danger mode so a careless approval is hard.

Architecture

Claude / MCP client ──stdio──► MCP server ──WebSocket(127.0.0.1:52346)──► Electron app
   (ssh_exec)                  (dist/mcp/server.js)                       (approval UI + ssh2)

• MCP server (src/mcp/server.ts) — speaks MCP over stdio, forwards each request to the desktop app and waits for the result. Auto-launches the app if it isn't running.

• Electron app (src/electron/main.ts) — holds the server list, renders the approval queue, and runs approved commands via ssh2. Single-instance: many MCP clients share one window.

• Renderer (src/renderer/index.html) — the UI: server list, per-server approval panel, tabbed interactive terminals (xterm.js), and all-server history.

Nothing binds to anything but 127.0.0.1.

Tools

Install & build

npm install
npm run build

Run the desktop app standalone

npm start

Open Servers → + Add and configure a host:

• Name — the handle the agent uses (e.g. web-01)

• User / Host / Port

• Authentication — Private Key (with optional passphrase), Password, or SSH Agent

Hit Test Connection to verify. Secrets you type are encrypted before they are stored; the UI never reads them back in clear text (leave a secret field blank when editing to keep the stored value).

Register as an MCP server

Point your MCP client at the built server entry (dist/mcp/server.js). For Claude Code:

claude mcp add ssh-proxy -- node /absolute/path/to/McpSshProxy/dist/mcp/server.js

Or in a client config:

{
  "mcpServers": {
    "ssh-proxy": {
      "command": "node",
      "args": ["/absolute/path/to/McpSshProxy/dist/mcp/server.js"]
    }
  }
}

The first ssh_exec call auto-launches the approval window.

Security notes

• All traffic is loopback-only (127.0.0.1:52346).

• Passwords and key passphrases are encrypted with the OS keystore (safeStorage). On platforms without an available keystore they fall back to obfuscated-but-not-encrypted local storage — prefer key/agent auth there.

• The agent can only target servers you have configured by name. An ssh_exec for an unknown host is refused.

• This is an approval tool, not a sandbox: once you click Run, the command executes with the configured user's privileges. Review before approving.

License

MIT — see LICENSE.