AI Analysis

Claude reads your server state and proposes structured fixes. Every action requires a backup and explicit confirmation — nothing is auto-applied.

What it does

AI Analysis collects a diagnostic snapshot of the target server and asks Claude to analyze it. The model returns a structured JSON response with a status, issue list, recommended actions, and free-form recommendations.

What the snapshot contains

• System diagnostics (uptime, memory, disk, load, top processes)
• OpenClaw status output
• Recent API errors from gateway logs (last 30 min)
• Agents list with models
• Sessions with token usage percentages
• Gateway config (bind, port, auth mode)
• If RAG is installed: health state and connections, plus the full RAG_ARCHITECTURE_DOC safety brief

What Claude returns

{
  "status": "OK" | "WARNING" | "CRITICAL",
  "summary": "One-sentence summary in Russian",
  "issues": [
    {
      "severity": "critical" | "warning" | "info",
      "title": "Short title",
      "description": "Detailed explanation",
      "action": {
        "type": "command",
        "label": "Button label",
        "command": "exact shell command",
        "requiresRestart": false
      }
    }
  ],
  "recommendations": ["Free-form suggestion"]
}

Safety rules (built into the prompt)

The system prompt explicitly instructs Claude to:

• Never propose destructive commands without a backup. Session resets, for example, must copy the file before touching it.
• Use mv instead of rm when possible.
• Return ONLY valid JSON — no markdown, no text before or after.
• If RAG is installed, treat it as read-only. No rm/mv on agent rag dirs, no direct index.db writes, never read config.json.

User confirmation

Every proposed action shows three buttons in the UI:

• Details — show the full command about to run
• Execute — run it, with automatic backup-before-execute
• Skip — dismiss this issue

If execution fails, the UI exposes a Rollback button that restores from the backup.

Free-form Q&A

Below the structured analysis there's an Ask about this server input. Type any question — Claude answers based on the same diagnostic snapshot. No persistence; each question is independent.

API key resolution

The Anthropic key used for analysis is resolved in this order:

1. Server's main agent auth profile (if present)
2. Global ANTHROPIC_API_KEY env var

If neither is set, AI Analysis is disabled on that server with a clear error.