Claude systemd

Read-only systemd inspection for Claude Desktop on Linux. Lets Claude answer the most common Linux ops question — "why is this service broken?" — without you copy-pasting systemctl and journalctl output by hand.

Seven tools cover unit listing, per-unit status, journal tailing, failed-unit hunting, unit-property inspection, raw unit-file display (including drop-in overrides), and queued-job inspection.

Strictly read-only in v0.1. No restart, enable, disable, or daemon-reload. Safe to enable on production hosts and shared systems. If you want destructive control, run those commands manually via claude-terminal-mcp.

Requirements

• Linux with systemd (any modern Ubuntu/Debian/Fedora/Arch).
• systemctl and journalctl on PATH (standard with systemd).
• Read access to the journal for the units you care about. By default, system-wide journal access requires being in the systemd-journal or adm group; user-scope journals are always readable by their owner.

Install (Claude Desktop)

1. Download the latest Systemd.mcpb from the Releases page.
2. Settings → Extensions → Extension Developer → Install Extension → pick the .mcpb.
3. Enable.

Tools

Tool	What it answers
list_units	What's running? What's loaded? Filter by type/state.
failed_units	Anything broken right now?
unit_status	Detailed status for one unit: active/sub state, enabled state, main PID, restart count, recent journal lines, fragment path, drop-in override paths.
unit_show	Full property dump (Environment vars, ExecStart, Restart policy, dependencies — 100+ properties).
unit_cat	Raw unit-file text plus drop-in overrides — equivalent to systemctl cat <unit>.
tail_journal	Recent log entries with optional unit / time / priority filters.
list_jobs	Currently-queued or running systemd jobs (rare; useful for stuck activations).

All tools support scope: "user" to inspect --user units. Default scope is system.

Example prompts

"Is anything in failed state right now?" → failed_units

"Why is ollama.service not using the GPU? Show me its drop-in overrides and last 50 log lines." → unit_cat ollama then tail_journal unit=ollama lines=50

"What did the journal say about errors and worse in the last 10 minutes?" → tail_journal priority=3 since="10 minutes ago"

"Is nginx enabled at boot, and how many times has it restarted since the last boot?" → unit_status nginx

Privacy policy

This extension runs entirely on your local machine and shells out only to systemctl and journalctl. No data leaves your machine. No network I/O. Read-only — cannot modify, restart, enable, disable, or reload anything.

The information visible to Claude includes:

• Names, descriptions, and metadata of every system or user unit you query (or every loaded unit if you call list_units).
• The full content of unit files and drop-in overrides (unit_cat) — these can contain Environment= lines with secrets if you've set them there. Do not store credentials in unit files if you intend to enable this extension.
• Recent journal entries for queried units — including any messages services log, which may contain personally-identifying info, IP addresses, request paths, or stack traces.
• Process IDs, restart counts, and runtime properties of services.

If your unit files or journals contain sensitive data and you don't trust the LLM endpoint, do not enable this extension.

Why no destructive operations?

A restart_unit / enable_unit / disable_unit set was deliberately left out of v0.1 because:

1. They require root for system-scope units, which means either running Claude Desktop as root (bad) or configuring sudo rules (extension-specific friction).
2. They're rarely needed in the workflow this extension actually serves — "why is X broken" is the 95% case.
3. If you need to bounce a service, the existing claude-terminal-mcp gives you run_command sudo systemctl restart <unit> with the same sudo mechanics you'd use anywhere else, plus a confirmation step at the system prompt.

Keeping this extension read-only means it can be enabled on a production host, a shared workstation, or a customer machine without trust concerns. v0.2 may revisit this.

Troubleshooting

"systemctl is not installed" — This server is Linux-only. macOS and Windows do not run systemd.

"Permission denied" when reading journal — Add yourself to systemd-journal or adm: sudo usermod -aG systemd-journal $USER then log out and back in. Or call with scope: "user" to read only your own user-scope journal (no group needed).

Empty tail_journal results — Some distributions have rate-limited or volatile journals. Check Storage= in /etc/systemd/journald.conf — volatile means the journal is wiped on reboot.

Development

Single ~400-line Node.js script, zero npm dependencies. Rebuild the .mcpb:

cd bundle-source
zip -j ../Systemd.mcpb manifest.json package.json server.js README.md LICENSE icon.png glama.json

License

MIT. See LICENSE.

Related

• claude-terminal-mcp — shell + filesystem + background jobs (use this for destructive systemctl actions).
• claude-sessions-mcp — tmux session management for long-running jobs.
• claude-rocm-mcp — AMD GPU monitoring; pairs well for "is the GPU healthy?" alongside "is the GPU service healthy?"
• claude-linux-mcp — X11 desktop control.
• claude-ollama-mcp — local Ollama HTTP-API client.