ssh MCP server

ssh-mcp

MCP server providing persistent SSH sessions for AI coding agents via the Model Context Protocol.

The Problem

AI coding agents can't interact with remote machines. You need to manually SSH in, run commands, and paste output back. For fleet management, server debugging, or remote development, this kills the feedback loop.

The Solution

ssh-mcp gives your AI agent direct SSH access to an allowed set of hosts. It resolves connections from ~/.ssh/config, authenticates via ssh-agent/gpg-agent, and exposes a virtual terminal so the agent sees exactly what you'd see.

# Start with allowed hosts
ssh-mcp --allowed-hosts="web1,db1,staging"

Features

• SSH config integration — Resolves hosts from ~/.ssh/config (HostName, Port, User, IdentityFile)
• Agent forwarding — Authenticates via ssh-agent or gpg-agent (SSH_AUTH_SOCK)
• Virtual terminal — Full PTY emulation with ANSI rendering via charmbracelet/x/vt
• File transfer — Bidirectional file and directory transfers via SFTP (recursive)
• Allowlist-only access — Only explicitly permitted hosts can be connected to
• Session management — Multiple concurrent sessions with independent state
• Zero config — Works out of the box if your ~/.ssh/config is set up

Installation

go install github.com/taigrr/ssh-mcp@latest

Requirements

• Go >= 1.26.5 (for installation)
• ~/.ssh/config entries for your hosts
• One of:
  • ssh-agent or gpg-agent running (recommended)
  • SSH private key at ~/.ssh/id_ed25519 or ~/.ssh/id_rsa

Quick Start

1. Ensure your hosts are in ~/.ssh/config:

Host my-desktop
    HostName 10.0.1.50
    User tai

1. Run with allowed hosts:

ssh-mcp --allowed-hosts="my-desktop"

1. Or use a config file (config.json in working directory):

{
  "hosts": ["my-desktop", "web1", "db1"]
}

🔌 MCP Tools

Tool	Description
ssh_connect	Connect to an allowed host by alias (optional width, height)
ssh_exec	Run a command and block until it finishes or times out
ssh_send_command	Send a command (non-blocking); for TUIs / fire-and-forget
ssh_send_keys	Send raw bytes with no newline (Ctrl-C, Esc, arrows, TUI keys)
ssh_get_screen	Get the current terminal screen content
ssh_clear_screen	Reset the local emulator screen + scrollback
ssh_resize	Resize a session's terminal (sends SIGWINCH to the remote shell)
ssh_read_file	Read a remote file over SFTP (clean bytes, no shell)
ssh_write_file	Write a remote file over SFTP (no quoting pitfalls)
ssh_edit_file	Exact find-and-replace edit of a remote file (read it first)
ssh_upload	Upload a local file or directory via SFTP
ssh_download	Download a remote file or directory via SFTP
ssh_list_sessions	List all sessions with active state and size
ssh_list_hosts	List allowed hosts with resolved connection details
ssh_close_session	Close a session

Choosing between ssh_exec and ssh_send_command

Both run in the same persistent shell (cwd and environment carry over), but they have opposite blocking semantics:

• ssh_exec is what you want for almost everything that should finish: builds, git, installs, queries, scripts. It wraps the command with internal sentinels, blocks until it completes or timeout_ms elapses, and returns clean output (ANSI/prompt stripped) plus the exit code. On timeout the command is not killed — it keeps running, and you get a screen snapshot so you can see what it is stuck on, then poll (ssh_get_screen), interrupt (ssh_send_keys Ctrl-C), or wait longer.
• ssh_send_command is non-blocking and exists for interactive TUIs (vim, less, htop, tmux, REPLs) and fire-and-forget launches. It returns the screen after a short settle delay regardless of completion.

The emulator backing every session is a real virtual terminal, so even when an ssh_exec command times out or a TUI is open, the session stays responsive and its state is always inspectable via ssh_get_screen — nothing can hang the server.

ssh_exec's timeout_ms defaults to 60s (matching Crush's bash tool, which auto-backgrounds a synchronous command after 60s) and is capped at 10m. A timeout is not an error: the tool returns a normal result noting the command is still running, plus the current screen, so the model just keeps going with ssh_get_screen / ssh_send_keys.

One shell per session: concurrency

Each session is a single interactive shell with one stdin and one stdout. The server keeps that consistent under concurrent tool calls:

• Byte level — all writes to the shell go through one serialized path, so bytes from different calls never interleave mid-sequence. Output is captured by a concurrency-safe terminal emulator.
• Two ssh_exec at once (same session) — rejected. Only one ssh_exec may be waiting in a session's shell at a time; the loser gets an "already executing" error. This prevents two wrapped commands from racing into the same shell and scrambling each other's captured output.
• ssh_send_command during an ssh_exec — rejected for the same reason (it would inject a second command line mid-exec).
• ssh_send_keys during an ssh_exec — allowed on purpose. This is how you send Ctrl-C to interrupt a command that timed out and is still running.
• Different sessions — fully independent; run them in parallel freely.

If you fire several commands at the same session simultaneously, expect all but one to come back with an "already executing" error rather than corrupting the shell. Sequence them, or use separate sessions.

Terminal size

Sessions default to a 200x50 virtual terminal, large enough for typical ls, tail, ps, log, and progress-bar output without triggering line wraps that pollute scrollback. Clients can override the initial size via width / height on ssh_connect, or resize a live session with ssh_resize. Values are clamped to [20, 500] columns x [5, 200] rows.

Session lifecycle

ssh_list_sessions returns every session the server knows about along with its current state ([active] or [inactive]) and current terminal size. Background goroutines watch each session's stdout and keepalive channel; when they detect that the remote end has gone away (EOF, keepalive timeout) the session is automatically marked inactive and removed from the manager, so a follow-up ssh_list_sessions won't keep advertising a dead entry. Only sessions reported as [active] accept ssh_send_command / ssh_get_screen / ssh_resize; attempts on inactive entries return ErrSessionInactive.

ssh_send_command is non-blocking

ssh_send_command writes the command + newline to the remote shell's stdin, waits a short settle delay (~3.5s) for the terminal to redraw, then returns the screen as-is. It does not wait for the command to finish. This is intentional: it keeps TUIs (vim, htop, tmux, less) interactive across calls.

Practical consequences for clients (especially LLM agents):

• A long-running command (sleep 60, make, training jobs, package installs) keeps running on the remote host after the call returns.
• To wait for completion, sleep on your side (your own timer / bash tool outside this MCP), then call ssh_get_screen to read fresh output. Do not prefix sleep N; to the SSH command expecting it to block here.
• Sending a second command before the first finishes queues bytes in the remote shell's stdin and produces garbled, interleaved output. Poll with ssh_get_screen until you see the prompt first.
• For output you'd rather not stream through the screen, redirect to a file (cmd > /tmp/out 2>&1 &) and read it back with cat.

⚙️ Crush Configuration

Add to your project's crush.json or .crush.json:

{
  "mcp": {
    "ssh": {
      "command": "ssh-mcp",
      "args": ["--allowed-hosts=my-desktop,staging"],
      "type": "stdio"
    }
  }
}

🤖 Claude Configuration

Claude Code

Register the server with the Claude Code CLI (writes to your MCP config):

claude mcp add ssh -- ssh-mcp --allowed-hosts=my-desktop,staging

Or add it manually to ~/.claude.json (global) or .mcp.json (project):

{
  "mcpServers": {
    "ssh": {
      "command": "ssh-mcp",
      "args": ["--allowed-hosts=my-desktop,staging"],
      "type": "stdio"
    }
  }
}

Claude Desktop

Add to your claude_desktop_config.json:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "ssh": {
      "command": "ssh-mcp",
      "args": ["--allowed-hosts=my-desktop,staging"]
    }
  }
}

If ssh-mcp is not on your PATH, use the absolute path (e.g. the output of which ssh-mcp, typically ~/go/bin/ssh-mcp) as the command. Because the server authenticates via ssh-agent, ensure SSH_AUTH_SOCK is available to the Claude process; you can pass it explicitly:

{
  "mcpServers": {
    "ssh": {
      "command": "/Users/you/go/bin/ssh-mcp",
      "args": ["--allowed-hosts=my-desktop,staging"],
      "env": {
        "SSH_AUTH_SOCK": "/path/to/ssh-agent.sock"
      }
    }
  }
}

Restart Claude Desktop after editing the config.

Flags

Flag	Description
--allowed-hosts	Comma-separated list of allowed SSH host aliases

If --allowed-hosts is not provided, falls back to config.json in the working directory.

How It Works

1. Resolve — Looks up host alias in ~/.ssh/config for connection details
2. Authenticate — Tries ssh-agent → key file → password (in order)
3. Connect — Opens a PTY session with xterm-256color terminal
4. Emulate — Captures output in a virtual terminal emulator
5. Expose — Returns rendered screen content (with ANSI) to the MCP client

Architecture

┌──────────────────────────────────────────────────┐
│                   ssh-mcp                        │
│                                                  │
│  ~/.ssh/config ──► Host Resolution               │
│                                                  │
│  SSH_AUTH_SOCK ──► Authentication                │
│                                                  │
│  Sessions:                                       │
│  ├─ tai@10.0.1.50:22 (PTY + VT)                  │
│  ├─ admin@web1.example.com:22  (PTY + VT)        │
│  └─ ...                                          │
│                                                  │
│  MCP Tools ◄──► stdio transport ◄──► AI Agent    │
└──────────────────────────────────────────────────┘

Development

# Build
go build -o ssh-mcp .

# Test
go test ./...

# Lint
staticcheck ./...

Related Projects

• Crush — Charm's AI coding agent
• charmbracelet/x/vt — Virtual terminal emulator
• neocrush — LSP/MCP bridge for Crush + Neovim

License

0BSD © Tai Groot