MCP Server

The MCP server is the real unlock. Register claude-guardian as a local MCP server and Claude can self-monitor mid-session — checking health, fixing logs, capturing bundles, and managing concurrency.

Setup

Add to your ~/.claude.json:

{
  "mcpServers": {
    "guardian": {
      "command": "npx",
      "args": ["claude-guardian", "mcp"]
    }
  }
}

Tools

Tool	What it returns
guardian_status	Disk, logs, processes, hang risk, budget, attention level
guardian_preflight_fix	Runs log rotation/trimming, returns before/after report
guardian_doctor	Creates diagnostics bundle (zip), returns path + summary
guardian_nudge	Safe auto-remediation: fix logs if bloated, capture bundle if needed
guardian_budget_get	Current concurrency cap, slots in use, active leases
guardian_budget_acquire	Request concurrency slots (returns lease ID)
guardian_budget_release	Release a lease when done with heavy work
guardian_recovery_plan	Step-by-step recovery plan naming exact tools to call

How Claude uses it

With the MCP server registered, Claude can reason about its own health:

1. Call guardian_status to check current conditions
2. If attention level is WARN or CRITICAL, call guardian_nudge for safe auto-remediation
3. Use guardian_budget_acquire before launching heavy parallel work
4. Call guardian_budget_release when done to free slots
5. If something goes wrong, call guardian_doctor to capture evidence

The guardian_recovery_plan tool returns a deterministic step-by-step plan naming exact tools to call. It never auto-restarts or kills processes — it just tells Claude what to do next.

Nudge behavior

guardian_nudge is the “do the safe things” action:

• If logs/disk thresholds are breached → runs preflight fix
• If warn/critical with no bundle yet → captures diagnostics
• Returns what changed and what to do next
• Never kills processes or restarts anything