Ssh Tool MCP Server

mcp-ssh-tool

Production-grade MCP SSH automation for operators, developers, and AI clients. mcp-ssh-tool opens persistent SSH sessions and exposes safe, structured tools for command execution, file operations, transfers, tunnels, package/service management, metrics, resources, and guided prompts.

v2 is secure by default: strict host-key verification is on, root login is off, raw sudo is policy-gated, destructive commands and filesystem mutations are denied unless policy allows them, and remote HTTP starts on loopback only unless bearer auth and allowed origins are configured.

Why This Server

• Trust: central policy engine, structured audit events, redacted logs, strict host keys, and machine-readable errors.
• MCP quality: stdio for local clients, Streamable HTTP for remote clients, legacy SSE only behind an explicit compatibility flag.
• AI-friendly tools: stable output schemas, structuredContent, annotations for read-only/destructive/idempotent behavior, resources, and curated prompts.
• Operations: session TTL/eviction, command timeouts, transfer checksum verification, real SSH forwarding, Prometheus metrics, and OpenTelemetry hooks.
• Portability: SFTP first, POSIX/BusyBox-aware shell fallbacks for basic file operations, and explicit support boundaries.

Quick Start

npm install -g mcp-ssh-tool

Add a stdio MCP server to your client:

{
  "servers": {
    "ssh-mcp": {
      "type": "stdio",
      "command": "mcp-ssh-tool",
      "args": []
    }
  }
}

Use it from your MCP client:

Open a safe SSH session to prod-1 as deploy, inspect host capabilities, then show disk usage.

Requirements

• Node.js 22.22.2+ or 24.14.1+ (LTS only)
• SSH access to target hosts
• A populated known_hosts file for strict host verification, or an explicit per-session host-key policy

Transports

Non-loopback HTTP startup is refused unless both --bearer-token-file and allowed origins are configured.

Secure Defaults

Per-session policyMode: "explain" returns a plan/verdict without executing. Use it before mutations when an AI client needs to summarize risk.

Policy Example

Set SSH_MCP_POLICY_FILE=/etc/mcp-ssh-tool/policy.json:

{
  "mode": "enforce",
  "allowRootLogin": false,
  "allowRawSudo": false,
  "allowDestructiveCommands": false,
  "allowDestructiveFs": false,
  "allowedHosts": ["^prod-[0-9]+\\.example\\.com$"],
  "commandAllow": ["^(uname|df|uptime|systemctl status)\\b"],
  "commandDeny": ["rm\\s+-rf\\s+/", "shutdown", "reboot"],
  "pathAllowPrefixes": ["/tmp", "/var/tmp", "/home/deploy"],
  "pathDenyPrefixes": ["/etc/shadow", "/etc/sudoers", "/boot", "/dev", "/proc"]
}

Simple deploys can use environment overrides such as SSH_MCP_ALLOW_RAW_SUDO=true, SSH_MCP_ALLOWED_HOSTS=prod-1.example.com, or SSH_MCP_PATH_ALLOW_PREFIXES=/tmp,/home/deploy.

Core Tools

• ssh_open_session, ssh_close_session, ssh_list_sessions, ssh_ping, ssh_list_configured_hosts, ssh_resolve_host
• proc_exec, proc_sudo, proc_exec_stream
• fs_read, fs_write, fs_list, fs_stat, fs_mkdirp, fs_rmrf, fs_rename
• file_upload, file_download
• ensure_package, ensure_service, ensure_lines_in_file, patch_apply
• os_detect, get_metrics
• tunnel_local_forward, tunnel_remote_forward, tunnel_list, tunnel_close

All tools return text plus stable structuredContent. Tool metadata includes titles, output schemas, and annotations that disclose read-only, destructive, idempotent, and external side-effect behavior.

Resources And Prompts

Resources:

• mcp-ssh-tool://sessions/active
• mcp-ssh-tool://metrics/json
• mcp-ssh-tool://metrics/prometheus
• mcp-ssh-tool://ssh-config/hosts
• mcp-ssh-tool://policy/effective
• mcp-ssh-tool://audit/recent
• mcp-ssh-tool://capabilities/support-matrix

Prompts:

• safe-connect
• inspect-host-capabilities
• plan-mutation
• managed-config-change

Support Matrix

Client Examples

ChatGPT or Claude Desktop:

{
  "mcpServers": {
    "ssh-mcp": {
      "command": "npx",
      "args": ["-y", "mcp-ssh-tool"]
    }
  }
}

VS Code or Cursor:

{
  "servers": {
    "ssh-mcp": {
      "type": "stdio",
      "command": "mcp-ssh-tool"
    }
  }
}

MCP Inspector over HTTP:

printf '%s' 'super-secret-token' > .mcp-token
mcp-ssh-tool --transport=http --host 127.0.0.1 --port 3000 --bearer-token-file .mcp-token

Configuration

High-value environment variables:

Deprecated aliases STRICT_HOST_KEY_CHECKING and SSH_MCP_STRICT_HOST_KEY are still accepted for one v2 compatibility cycle. Prefer SSH_MCP_HOST_KEY_POLICY.

Development

Use the exact local runtime from .nvmrc / .node-version, then run:

npm ci
npm run check

Live SSH suites are opt-in:

RUN_SSH_INTEGRATION=1 npm run test:integration
RUN_SSH_E2E=1 npm run test:e2e

Local quality gates are layered:

• pre-commit: formats staged files and lints staged TypeScript only
• pre-push: runs npm run check:push
• manual/full parity: npm run check

CI/CD Ownership

The personal repository https://github.com/oaslananka/mcp-ssh-tool is the main source repository. Automatic CI/CD and trusted npm publishing run from the organization mirror https://github.com/oaslananka-lab/mcp-ssh-tool; Azure, GitLab, and personal GitHub release actions are manual-only.

See docs/ci-cd-topology.md for mirror secrets, release flow, and manual fallback guidance.

Documentation

• ARCHITECTURE.md
• SECURITY_DECISIONS.md
• MIGRATION.md
• docs/ci-cd-topology.md
• docs/configuration.md
• docs/security-model.md
• docs/troubleshooting.md
• docs/enterprise-deployment.md

License

MIT License. See LICENSE.