SSH MCP Server

A Model Context Protocol (MCP) server that gives LLM clients safe, persistent SSH access to remote machines.

Table of Contents

1. Overview

2. Key Features

3. Architecture

4. Installation

5. Running the Server

6. Client Integrations

   • Claude Desktop

   • Claude Code

   • Codex

   • Cursor

7. Host Configuration

   • Host Storage

   • Adding Hosts

   • Listing Hosts

   • Editing Hosts

   • Removing Hosts

8. Session Management

   • Starting a Session

   • Listing Sessions

   • Executing Commands

   • Closing Sessions

9. Authentication Modes

10. Timeouts & Inactivity Handling

11. Directory Structure

12. Using the MCP Tools

13. Testing

14. Troubleshooting

15. Security Considerations

16. Contributing

17. License

Overview

ssh-mcp lets MCP-compatible clients (such as Claude Code, Cursor, or custom MCP inspectors) control remote machines through SSH. Once hosts are registered, the server maintains persistent shell sessions that retain environment state between commands—ideal for multi-step workflows, long-running processes, or interactive diagnostics.

The server is implemented in TypeScript on top of the official @modelcontextprotocol/sdk.

Key Features

• MCP-compliant: exposes functionality through standard MCP tool definitions.

• Persistent sessions: keep shell sessions alive, preserving working directory, environment variables, and process state across multiple commands.

• Stored host profiles: manage SSH targets through durable JSON configuration (~/.ssh-mcp/hosts.json).

• Flexible authentication: supports passwords, private keys, and SSH agent forwarding (fallback).

• Timeout & cleanup safeguards: sessions auto-close after prolonged inactivity; commands are marked and monitored for completion.

• Structured listings: query active sessions and saved hosts directly from the MCP client.

Architecture

Each active session maintains:

• SSH connection via ssh2

• Interactive shell (conn.shell) to support multi-command pipelines

• Buffered output with unique UUID markers to detect command completion

• Inactivity timer (defaults to 2 hours)

Installation

Install the published package from npm so the ssh-mcp-sessions executable is available in your environment.

Global install (preferred)

Launch the server anywhere by running ssh-mcp-sessions.

Project-local install

Inside your project:

Run the server with npx ssh-mcp-sessions (or reference ./node_modules/.bin/ssh-mcp-sessions).

Building from source is only required for contributing. See the Contributing section if you plan to work on the codebase itself.

Claude Desktop

Add an entry to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or the appropriate config path on Windows/Linux:

Restart Claude Desktop after saving the file. You can now use the MCP Inspector or the command palette (Cmd/Ctrl+Shift+O) to call tools like add-host and start-session.

Claude Code (VS Code extension)

Update the Claude Code workspace settings (.vscode/settings.json or global settings) with:

Reload the window. The MCP panel will list ssh-mcp, and commands are available via the command palette (Ctrl/Cmd+Shift+P → “Claude: Run MCP Tool”).

Codex (OpenAI GPT-4o/5 with MCP)

Create or edit ~/.config/openai-codex/mcp.toml (the path may differ per platform—use the location documented by the client). Add:

Restart Codex or re-open the MCP inspector. The ssh-mcp tools will appear under the configured servers list.

Cursor IDE

Open Cursor settings → “Model Context Protocol” (or edit ~/Library/Application Support/Cursor/mcp.json directly) and include:

After saving, reload Cursor. The MCP sidebar exposes the server; you can invoke tools via chat or the command palette (Cmd/Ctrl+Shift+L).

Tip: If you prefer an explicit path instead of relying on npx, replace the command/args with the absolute path to the executable (node_modules/.bin/ssh-mcp-sessions for local installs or /usr/local/lib/node_modules/ssh-mcp-sessions/build/index.js for global installs).

Running the Server

The server is purely stdio-based. Once running it prints:

You can register it with Claude Code or any other MCP client by pointing to the executable installed from npm:

If you prefer to rely on PATH resolution (e.g., after a global install), you can simplify the entry to { "command": "ssh-mcp-sessions" }.

Note: The server no longer accepts CLI arguments for host/user/password. Everything is configured dynamically via MCP tools.

Host Configuration

Host Storage

• Hosts are persisted in ~/.ssh-mcp/hosts.json.

• The directory is created automatically if it does not exist.

• File format:

Fields:

• id (string) — unique identifier used by all session commands.

• host (string) — hostname or IP.

• port (number, default 22) — SSH port.

• username (string) — SSH user.

• password (optional string) — password auth.

• keyPath (optional string) — private key path; tilde expansion supported.

  • If neither password nor keyPath is provided, the server attempts to use the local SSH agent via SSH_AUTH_SOCK (with agent forwarding enabled).

The MCP tools ensure this file remains well-formed; never edit it manually unless you know what you’re doing.

Adding Hosts

Tool: add-host

• host_id: new identifier. Must be unique.

• host: hostname or IP.

• port: optional (defaults to 22); provide integer > 0.

• username: SSH user.

• password or keyPath: optional; configure one or rely on agent.

Example (Claude Code command palette or inspector):

Listing Hosts

Tool: list-hosts

Returns text with one host per line:

auth values:

• password — password field present

• key — keyPath present

• agent — neither password nor keyPath; agent fallback active

Editing Hosts

Tool: edit-host

Only supply the properties you want to change. Omitted fields remain unchanged; providing null to a field is not supported—set an empty string or remove the host instead.

Removing Hosts

Tool: remove-host

Deletes the entry from hosts.json. Active sessions using that host must be closed manually.

Session Management

Starting a Session

Tool: start-session

Optionally you can supply sessionId; otherwise, a UUID is returned.

Example response:

Listing Sessions

Tool: list-sessions

Shows all active sessions with metadata:

Executing Commands

Tool: exec

• Commands are sanitized (trimmed, length-limited).

• Output is captured from the persistent shell and returned as plain text.

• Non-zero exit codes raise McpError with stderr in the message.

Example output:

Closing Sessions

Tool: close-session

Note: session IDs for close-session use sessionId (camelCase) to remain backwards compatible with the underlying tool definition.

Legacy Helper

Function execSshCommand(hostId, command, sessionId?) remains exported for programmatic use and simply delegates through the session machinery described above.

Authentication Modes

1. Password — stored in hosts.json; transmitted to ssh2 during connection.

2. Private key — keyPath read at runtime; supports encrypted keys (prompt user to set SSH_MCP_KEY_PASSPHRASE before launch if needed).

3. SSH agent (fallback) — if neither password nor key is set and SSH_AUTH_SOCK is present, the agent is passed to ssh2 (agentForward: true).

Timeouts & Inactivity Handling

• Each session has a global inactivity timeout (default 2 hours). Timer resets whenever a command executes successfully.

• If the timer elapses, the session cleans up the SSH connection, shell, and resolver buffer, and removes itself from activeSessions.

• Command completion uses a UUID marker: printf '__MCP_DONE__{uuid}%d\n' $?. Output before the marker is returned; numeric code after the marker becomes the exit status.

Directory Structure

Using the MCP Tools

Below is a typical workflow using Claude Code (commands start with /mcp), but the same JSON payloads apply to any MCP inspector.

1. Add host

   /mcp mcp-remote-ssh add-host {"host_id":"host","host":"host.local","username":"user"}

2. Start session

   /mcp mcp-remote-ssh start-session {"host_id":"host"}

   → returns session_id

3. Run commands

   /mcp mcp-remote-ssh exec {"session_id":"<id>","command":"pwd"} /mcp mcp-remote-ssh exec {"session_id":"<id>","command":"ls -la"}

4. Inspect

   /mcp mcp-remote-ssh list-sessions /mcp mcp-remote-ssh list-hosts

5. Close session

   /mcp mcp-remote-ssh close-session {"sessionId":"<id>"}

Testing

Unit tests (Vitest):

Integration smoke tests for SSH are not included by default because they require external infrastructure. You can manually validate with the workflow above.

Troubleshooting

Security Considerations

• Treat ~/.ssh-mcp/hosts.json as sensitive; it may contain passwords or key paths.

• Prefer key-based or agent authentication where possible.

• Limit hosts.json permissions: chmod 600 ~/.ssh-mcp/hosts.json.

• Sessions inherit all privileges of the configured SSH user.

• Long-running sessions can be closed manually or rely on the inactivity timeout.

Contributing

1. Fork the repo and create a branch.

2. Make your changes with tests and documentation updates.

3. Run npm run build and npm run test before submitting a PR.

4. Follow the Code of Conduct.

Issues and feature requests are welcome via GitHub.

License

MIT

Happy automating! If this project improves your workflow, please star the repository or share feedback. Your contributions help make remote development safer and simpler for everyone.