MCP SSH Orchestrator

What Problem Does This Solve?

Imagine this: Your AI assistant (Claude, ChatGPT, etc.) can access your servers, but you're terrified of what it might do. rm -rf /? Delete your databases? Change firewall rules?

Now imagine this: Your AI has governed, auditable access to your infrastructure. It can check logs, restart services, and manage your fleet, but only if your security policies allow it.

That's exactly what MCP SSH Orchestrator provides: the power of AI-driven server management with deny-by-default access control, IP allowlists, host key verification, and comprehensive audit logging backed by declarative YAML policy-as-code (.

Why This Matters

Zero-Trust Security Model

• Deny-by-default: Nothing runs unless explicitly allowed

• Network controls: IP allowlists prevent lateral movement

• Command whitelisting: Only approved commands can execute

• Declarative policy-as-code: Versioned YAML files define hosts, credentials, and allowed commands

• Comprehensive audit trails: Every action is logged in JSON

Prevents Common Attack Vectors

• Dangerous commands blocked: rm -rf, dd, file deletions

• Network isolation: Servers can't access external internet

• No privilege escalation: Runs as non-root in containers

• Resource limits: CPU and memory caps prevent DOS

Production-Ready Audit & Security

• OWASP LLM Top 10 protected: Mitigates LLM07 (Insecure Plugin Design), LLM08 (Excessive Agency), LLM01 (Prompt Injection)

• MITRE ATT&CK aligned: Prevents T1071 (Application Layer Protocol), T1659 (Content Injection)

• Structured JSON audit logs: Complete audit trail with timestamps, hashes, and IPs

• Forensics ready: Command hashing, IP tracking, detailed metadata

• Real-time monitoring: Progress logs for long-running tasks

Who Is This For?

Homelab Enthusiasts

• Automate routine server maintenance with AI

• Safely manage Proxmox, TrueNAS, Docker hosts

• Get help troubleshooting without losing SSH security

Security Engineers

• Audit and control AI access to infrastructure

• Implement zero-trust principles with declarative policy-as-code configs

• Meet compliance requirements with structured logging

DevOps Teams

• Let AI handle routine tasks: log checks, service restarts, updates

• Manage fleets of servers through conversational interface

• Reduce manual toil while maintaining security standards

Platform Engineers

• Enable AI-powered infrastructure management

• Provide secure self-service access to developers

• Bridge the gap between AI and infrastructure securely

Real-World Use Cases

Scenario 1: Homelab Automation (Homelab Enthusiasts)

You say: "Claude, my Proxmox host is running slow. Can you check disk usage and memory on all my VMs?"

What happens

• Policy allows df -h and free -m on Proxmox hosts

• Network check: Private IP allowlist permits access

• Tag-based execution checks all hosts tagged proxmox

• Commands execute safely with no destructive operations

• Complete audit trail stored in JSON logs

Scenario 2: Incident Response (DevOps Teams)

You say: "We're seeing 500 errors. Check nginx logs across all production web servers and show me the last 100 error lines."

What happens

• Tag-based execution: tail -n 100 /var/log/nginx/error.log runs on all web-prod servers

• Network isolation enforced: No external API calls or egress allowed

• Real-time progress logs stream via MCP context events

• Structured output aggregates results for quick triage

• Full audit trail with timestamps for post-incident review

Scenario 3: Fleet-Wide Maintenance (Platform Engineers)

You say: "Update system packages on all staging servers, but show me what would change first before running the upgrade."

What happens

• Use ssh_plan to preview apt list --upgradable across staging tagged hosts

• Review dry-run output to see pending updates

• Policy validates apt update && apt upgrade -y is allowed on staging

• Tag-based execution runs upgrade on all staging servers in parallel

• Audit logs track which servers were updated and when

Quick Start

1. Prepare local configuration (one-time)

If you prefer to lay things out manually, follow the steps below.

2. Launch the orchestrator container

Restart later with docker start mcp-ssh-orchestrator. Prefer disposable containers? Use docker run -i --rm ... instead.

3. Connect your MCP client

• Cursor: Add to ~/.cursor/mcp.json

• Claude Desktop (macOS): Update ~/Library/Application Support/Claude/claude_desktop_config.json

(Windows path: %APPDATA%\\Claude\\claude_desktop_config.json.)

More examples (Docker Desktop, multi-environment, SDK usage) live in the Integrations guide.

4. Test the connection

Cursor/Claude should now show the orchestrator as connected. Jump to the Usage Cookbook for guided scenarios.

How Security Works (The Technical Details)

Policy-as-code workflow: config/servers.yml, config/credentials.yml, and config/policy.yml are parsed on startup, enforced during every ssh_* tool invocation, and mirrored in the structured audit logs so the same declarative files you review in Git gate what your AI can execute.

Defense-in-Depth Architecture

What Gets Blocked

What Gets Allowed (Examples)

Protection Against Real Threats

MCP SSH Orchestrator directly addresses documented vulnerabilities in the MCP ecosystem:

• CVE-2025-49596: Localhost-exposed MCP services → Mitigated with stdio-only transport

• CVE-2025-6514: Command injection in MCP servers → Mitigated with policy-based validation

• 43% of MCP servers have command injection flaws → Zero-trust security model

Full Security Model Documentation | Security Risks Analysis

Documentation

Complete Documentation Wiki

Supply Chain Integrity

Signed release artifacts: Every tarball/zip in GitHub Releases ships with a detached GPG signature produced by the maintainer key (openpgp4fpr:6775BF3F439A2A8A198DE10D4FC5342A979BD358). Import the key and verify before unpacking:

Cosign-signed container images: The images under ghcr.io/samerfarida/mcp-ssh-orchestrator are signed via Sigstore keyless signing in the release workflow. Verify the signature (and optional attestations) before deploying:

Image digests and signatures are published with every tag in GitHub Packages so you can pin exact references when promoting builds between environments (package feed).

OpenSSF Scorecard: The repository maintains an automated Scorecard run to track security posture across dependencies, build settings, branch protections, and more (scorecard summary).

What Can AI Do With This? (MCP Tools)

Your AI assistant gets 13 powerful tools with built-in security:

Discovery & Planning

• ssh_list_hosts - See all available servers

• ssh_describe_host - Get host details and tags

• ssh_plan - Test commands before running (dry-run mode)

Execution

• ssh_run - Execute single command on one server

• ssh_run_on_tag - Run command on multiple servers (e.g., all "web" servers)

• ssh_run_async - Start long-running tasks in background

Monitoring & Control

• ssh_get_task_status - Check progress of async tasks

• ssh_get_task_output - Stream output in real-time

• ssh_get_task_result - Get final result when done

• ssh_cancel - Stop a running synchronous task safely

• ssh_cancel_async_task - Stop a running async task safely

Management

• ssh_reload_config - Update hosts/credentials without restart

• ssh_ping - Verify connectivity to a host

MCP Resources + Context

• ssh://hosts – discover sanitized host inventory (alias, tags, description, credential presence)

• ssh://host/{alias} – inspect a single host without exposing credentials

• ssh://host/{alias}/tags – fetch tag-only view for planning tag executions

• ssh://host/{alias}/capabilities – derived policy summary, limits, and sample command allowances per host

Context-aware logging: Streams lightweight ctx.debug / ctx.info events (task start, completion, cancellations) in supported clients for ssh_run, ssh_run_on_tag, config reloads, and async task polling—all without exposing raw commands or secrets.

LLM-friendly hints: Policy/network denials (and ssh_plan previews) include helpful hints so assistants automatically retry with ssh_plan, consult the orchestrator prompts, or ask whether a policy/network update is appropriate instead of looping on blocked commands.

Complete Tools Reference with Examples

Learn More

Key Differentiators

• Production-Ready Security: OpenSSF Scorecard 7.5+ score

• Zero-Trust Architecture: Deny-by-default, allow-by-exception

• OWASP LLM Top 10 Protected: Mitigates insecure plugin design, excessive agency, prompt injection

• MITRE ATT&CK Aligned: Prevents content injection and unauthorized protocol usage

• Security-Focused: Built on security-first principles against real CVEs (CVE-2025-49596, CVE-2025-6514)

• Easy Integration: Works with Claude, ChatGPT, and any MCP client

• Open Source: Apache 2.0 licensed, community-driven

What Users Are Saying

"Finally, I can let Claude manage my Proxmox cluster without fear!" - Homelab Admin

"This is what infrastructure-as-code should have been. Declarative security for AI access." - Platform Engineer

"The structured audit logs make incident response so much easier." - Security Engineer

Contributing

We welcome contributions! See our Contributing Guide for:

• Development setup

• Code of conduct

• How to submit PRs

• Architecture decisions

License

Apache 2.0 - See LICENSE for details.

Links

• GitHub Repository - Star us on GitHub!

• Issue Tracker - Report bugs or request features

• CHANGELOG - Version history and release notes

• MCP Specification - Learn about MCP

• Docker MCP Security Guide - Security best practices