laptop-care

A laptop maintenance agent that runs inside Claude Desktop. It inspects your machine, explains what it found, recommends what to do, and remembers what it saw last time.

Works on macOS and Windows. No API key, it runs on your existing Claude Desktop subscription. Nothing leaves your machine.

npx laptop-care setup

Then restart Claude Desktop and say "run my laptop maintenance".

What you actually get

A 15 point inspection, presented the way a garage hands you a check sheet after a service. Every point is shown, not just the failures, because knowing 10 things passed is as useful as knowing 2 failed.

LAPTOP INSPECTION REPORT
MacBook Pro 16-inch 2021  ·  M1 Pro  ·  32 GB  ·  macOS 26.5.1
First inspection  ·  15 points checked

STORAGE
  Disk space             Good      607 GB free of 926 GB, 65 percent
  Cache composition      Watch     Docker 6 GB of it, rebuilds in weeks

SECURITY
  Disk encryption        Good      FileVault on
  Startup agents         Watch     49 agents and daemons present

MAINTENANCE
  OS updates             Action    26.5.2 available, security release
  Backup                 Ask       No Time Machine, need your answer

  10 good  ·  4 watch  ·  1 needs action  ·  1 question for you

Then a verdict in plain language, a ranked list of what to do, and it acts only on what you approve.

Related MCP server: AI Dev Assistant

Why this is an agent and not a script

A script runs the same 15 commands in the same order and prints the output. This decides what to do based on what it finds and what it remembers.

The agent loop in practice:

graph TD
    A["User: run my laptop maintenance"] --> B[start_maintenance]
    B --> C{First run or returning?}
    C -->|First| D[Introduce, explain, ask permission]
    D --> E{User agrees?}
    E -->|No| F[Stop. Nothing runs.]
    E -->|Yes| G[grant_consent unlocks the scanners]
    C -->|Returning| H[Load unresolved issues from last time]
    G --> I[Run diagnostic checks]
    H --> I
    I --> J[build_report_card verifies real coverage]
    J --> K[analyze_trends compares to this machine's history]
    K --> L[Present findings and ranked recommendations]
    L --> M{User picks what to act on}
    M --> N[Apply approved fixes, verify they worked]
    N --> O[Save report, record metrics, track what was skipped]
    O --> P[Next run resumes from here]

Guardrails that are code, not instructions

The interesting part is what the agent is unable to do. These are enforced by the server, so they hold even if the model misreads its instructions.

Consent lock. On a fresh install every scanning tool is locked. The model cannot read your disk, battery, or security posture until it has introduced itself and you have said yes.

delete BEFORE consent      ->  CONSENT_REQUIRED

Sequencing gates. Destructive tools refuse to run until their prerequisite has completed. "Never clean without showing the user first" is a precondition, not a request.

save_report BEFORE diagnostic  ->  PREREQUISITE_NOT_MET
save_report AFTER  diagnostic  ->  saved

Honest coverage. The inspection card is generated by the server from a log of what genuinely executed. Run 3 checks of 15 and the card says so. The model cannot claim an inspection it did not perform.

Ran 3 of 15 checks deliberately:
  total_points     : 15
  checks_completed : 3
  checks_missing   : 12
  rows locked to "Not checked": 12

The division of labor: the server owns facts, sequencing, and coverage. The model owns judgment and language. Anything that can be enforced is enforced; only what genuinely requires interpretation is left to the prompt.

What it checks

Persistence monitoring is the one most tools skip. Launch agents and daemons are how unwanted software survives a reboot. laptop-care records them on the first run and afterwards reports only what is new, which is the part worth reading.

What it remembers

Everything stays in ~/.laptop-care/, in your home folder, separate from the installed package. Updating or reinstalling never touches it.

Baselines need history to be useful. Trends become meaningful around the fourth run, and persistence changes start reporting from the second.

Architecture

src/
  index.ts      MCP server: consent lock, sequencing gates, execution log,
                report card generation, lifecycle-aware playbook loading
  tools.ts      Tool definitions with safety tiers
  commands.ts   Platform commands keyed by darwin | win32
  runner.ts     exec wrapper with per-tool timeouts
  cli.ts        setup, help, and serve

prompts/
  kernel.md         Always sent on connect. Identity and safety rules.
  first-run.md      Onboarding workflow. Only for new users.
  returning-run.md  Follow-up workflow. Only for returning users.
  shared.md         Judgment, escalation thresholds, tone, report format.

The playbook is split by when it is needed, not by topic. A returning user never loads first-time onboarding. start_maintenance checks whether history exists and hands over only the workflow that applies:

32 tools across three safety tiers: 25 read-only, 6 requiring explicit consent, 1 requiring admin.

Scheduling

On the first run it offers to set up recurring checks, weekly, monthly, or quarterly. It creates a launchd agent on macOS or a Task Scheduler entry on Windows. You can also just say "schedule my maintenance checks" any time.

Without a schedule it still notices when it has been more than 30 days and mentions it.

Manual setup

If npx laptop-care setup cannot find your config, add this to claude_desktop_config.json:

{
  "mcpServers": {
    "laptop-care": {
      "command": "npx",
      "args": ["-y", "laptop-care"]
    }
  }
}

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

• Windows: %APPDATA%\Claude\claude_desktop_config.json

For Claude Code: claude mcp add laptop-care -- npx -y laptop-care

Honest limitations

• It reads surfaces. It does not detect malware, verify that a backup restores, or confirm firmware is current against the vendor. It reports what the OS reports.

• History only grows when a run finishes. A run you interrupt records nothing. There is no partial save.

• Baselines need about four runs before "abnormal for this machine" means anything.

• Tone and judgment are prompt-shaped. The safety gates are code and hold absolutely. How well it explains a finding does not.

• macOS and Windows only. Linux returns a clear unsupported error rather than guessing.

Requirements

Node.js 18+, Claude Desktop or another MCP client, macOS or Windows.

License

MIT