openclaw-health-mcp

MCP server for AI agent deployment health — gateway status, CPU/RAM/swap, recent errors from journalctl/dmesg, skill-registry integrity, upgrade outcomes, cron + disk usage in a single tool call. Each component gets a HEALTHY/DEGRADED/CRITICAL classification, with overall rollup + ranked critical findings. Linux-proc backend works on any Linux/macOS/Windows host; OpenClaw operators get native ~/.openclaw/ parsing as a built-in reference implementation. Keywords: AI agent health, production AI monitoring, deployment readiness, MCP infrastructure observability.

What it does

Anyone running production AI agents needs a single tool that answers "is this deployment healthy right now?" without SSH'ing in to run six separate commands. The HN front-page thread Ask HN: How are you monitoring AI agents in production? (March 2026) made the gap explicit — the most-upvoted comments described:

• "observability and governance cannot live inside the agent framework. They have to live in an independent execution layer" — the framework-level monitoring leaks audit trail when teams use multiple frameworks
• "agent makes 10,000 correct $0.02 decisions that collectively don't make sense" — per-call rate limits miss systemic patterns
• The gap that "actually hurts during post-mortems" — knowing whether a model drifted, context window failed, or tool misbehaved

Existing options (LangSmith, Langfuse, AgentShield, OTEL/LGTM) sit at the framework or proxy layer. openclaw-health-mcp sits one level closer to the agent runtime — read-only, local, MCP-native — surfacing infrastructure-layer health (gateway, CPU/RAM, recent errors, skill-registry, upgrade outcome, cron, disk) to the same Claude conversation that's running the agent. Works on any Linux/macOS/Windows host out of the box via the linux-proc backend; OpenClaw operators get an additional native backend that parses ~/.openclaw/ paths.

> claude: is my OpenClaw deployment healthy?
[MCP tool: health_overview]
overall_health: critical
component_summary:
  gateway: degraded         (bound to 0.0.0.0, 1 crash in 24h)
  resources: degraded       (memory at 78%, swap at 12%)
  skill_registry: critical  (skill 'clawhub-trending-bot-v2' flagged suspicious)
  upgrade: degraded         (last upgrade rolled back)
  cron: degraded            (1 overdue job)
  disk: degraded            (root at 82%, log dir +187 MB/24h)

critical_findings:
  [CRITICAL] Skill 'clawhub-trending-bot-v2' flagged — possible exfiltration. Disable.
  [DEGRADED] Last upgrade 2026.4.23→2026.4.26 rolled back: websocket_stalls, cpu_spike.
  [DEGRADED] Root disk at 82% — set up log rotation before reaching 95%.
  [DEGRADED] 1 cron job(s) overdue. Install silentwatch-mcp for silent-failure detection.

Why openclaw-health-mcp

Three things that existing tools (Datadog, Prometheus, raw top/free/df) don't do for OpenClaw specifically:

1. OpenClaw-aware probes. Detects 0.0.0.0-binding (the default-publicly-exposed misconfig per the 135k exposed-instances stat), parses ClawHub skill-registry diffs, recognizes named upgrade-regression patterns (websocket_stalls, cpu_spike post-2026.4.26), distinguishes intentional restarts from crashes.
2. MCP-native, no integration layer. Claude Desktop, Cline, Continue, OpenClaw agents — any MCP-aware client queries d

... View full README on GitHub