The design point is the restriction: predefined commands mean the agent can read `smartctl` output and container logs but cannot run anything you did not build in. That covers exactly the questions platform APIs cannot answer — disk health, real load averages, log correlation — and it means a compromised conversation still cannot touch the box.
An MCP server that runs predefined diagnostic commands on a server over SSH and passes only the results to the model. The credentials and the shell stay on the server side; the agent gets findings, not a prompt.
- 12 tool modules with 79+ actions, loaded according to the platform the server auto-detects at startup — 10 core modules on generic Linux, 12 on Unraid
- Docker — container logs, stats, environment, ports and network topology
- Storage and array — parity checks, SMART data, temperatures and mover logs on Unraid
- System and logs — processes, disk I/O, network connections, and search across container and system logs
- VMs and security — list and inspect VMs with VNC details and libvirt logs; port scanning, login monitoring and permission audits
- Health diagnostics that aggregate status with automatic issue detection
- Read-only by design, so there is no path to an accidental change
A reachable host over SSH and a key. The quickest path is the published image: `docker run` `ghcr.io/jeprecated/mcp-ssh-sre:latest` with `SSH_HOST`, `SSH_USERNAME` and `SSH_KEY_PATH` set and the key mounted read-only, exposing port 3000; then point your client at `http://your-server:3000/mcp`. Stdio is available too. Unraid and generic Linux are fully supported; TrueNAS and Proxmox are untested, though the core tools should work. Package version 2.0.2. ISC.
One command — docker run -d \ -p 3000:3000 \ -e SSH_HOST=server.local \ -e SSH_USERNAME=mcp-readonly \ -e SSH_KEY_PATH=/keys/id_ed25519 \ -v ~/.ssh/id_ed25519_mcp:/keys/id_ed25519:ro \ ghcr.io/jeprecated/mcp-ssh-sre:latest