Ask Claude why your Mac is full — you still decide what goes

A question, not an answer

"Your disk is almost full" is a question, not an answer. Where did 180 GB go? Is ~/Library/Developer safe to empty? What is System Data? An AI assistant is good at explaining that, but only if it can see your disk — and you should never have to hand it a delete button to get there.

Dropit 0.2 gives your agent the eyes and keeps the delete button for you.

What it does

Switch on Settings → Agents & CLI and Dropit runs a small, local Model Context Protocol (MCP) server. Claude Desktop, Claude Code and Cursor can then:

• scan your home folder, a folder, a volume or the whole Mac;
• read the results: biggest folders, big files nobody has opened in years, bytes by file type, Quick Wins like Xcode caches and old installers, duplicate files, installed apps and what an uninstall would remove;
• explain what is using your Mac right now;
• stage items into Dropit's cleanup list, with the same safety tier and plain-English reason you see in the app.

What it cannot do

Your agent cannot delete anything. There is no delete tool, no uninstall tool and no "quit process" tool — and a test in our build fails if anyone adds one.

When an agent stages items, Dropit marks them "Added by Claude" in the cleanup sheet. The Clean button stays disabled until you tick "I have reviewed the items the agent added", and Dropit checks again, in its own engine, that you reviewed exactly those items in a window that is in front of you. If the agent adds one more item afterwards, you review again. Protected locations — /System, your home folder itself, volume roots — cannot be staged at all.

Everything still goes to the Trash by default, with Finder's Put Back.

Set it up in a minute

Open Dropit → Settings → Agents & CLI and switch on agent access. The panel shows the snippets below with the exact path on your Mac, each with a Copy button.

Claude Desktop: add this to claude_desktop_config.json, then restart Claude. Cursor takes the same JSON in ~/.cursor/mcp.json.

{
  "mcpServers": {
    "dropit": {
      "command": "/Applications/Dropit.app/Contents/MacOS/dropit-mcp"
    }
  }
}

Claude Code, or let the app do it

Claude Code needs one command. Or let the bundled helper write the client configuration for you; it keeps a backup of the file it changes.

If Dropit is not running when your agent needs it, it starts quietly in the background.

claude mcp add dropit /Applications/Dropit.app/Contents/MacOS/dropit-mcp
/Applications/Dropit.app/Contents/MacOS/dropit-mcp install claude-desktop

Things to ask

A few prompts that work well:

• "Why is my disk full? Don't stage anything yet."
• "Find the biggest space wasters in my home folder and explain which are safe."
• "Which Xcode and simulator data can I remove? Stage only the safe ones."
• "Review my cleanup list and tell me if anything looks risky."
• "What is making my Mac hot right now?"

Built-in prompts — find space wasters, review cleanup queue and why is my disk full — are available in clients that show MCP prompts.

Privacy, plainly

The connection between your agent and Dropit is a private file on your Mac that only your user account can open — not a network port. Dropit itself still sends nothing anywhere: no telemetry, no file names, no scan results.

Your agent is a different matter. If it is backed by a cloud model, the results it reads (folder names, sizes, paths) go to that model's provider, under that provider's terms. If that matters for a particular folder, ask the agent to scan somewhere else, or leave agent access switched off — it is off until you turn it on.

How the review gate works

Every item in the cleanup list remembers who put it there. Items you add yourself behave exactly as they always have. Items an agent adds carry its name, and while even one of them is in the list, the engine refuses to clean until you have acknowledged that exact set.

The acknowledgement is not a flag the agent can set. It can only come from the app window, it has to name every agent-added path that is queued at that moment, and the window has to be in front of you when you press Clean. If the agent stages another item after you tick the box, the tick is withdrawn and the new item is shown to you before anything happens.

The safety tiers do not change either. An item Dropit would mark as risky when you stage it is marked risky when Claude stages it, with the same reason in plain English. Protected system locations are refused for everyone, and the agent is told why, so it can explain the refusal instead of trying again.

• Agent-added items are labelled with the agent's name in the cleanup sheet.
• Clean stays disabled until you tick the review box for exactly those items.
• Adding one more item withdraws the tick until you look again.
• Trash remains the default, and permanent deletion still asks you to type the item count.

When it helps, and when to leave it off

Agent access is most useful when the question needs judgement as well as numbers: which of three simulator runtimes you still build against, whether a 40 GB Docker disk is a project you abandoned, or why a folder doubled since last month. The agent can read the scan, ask you what a project is for, and stage only what you agree to remove.

If you just want to see where the space went, you do not need an agent at all. The map, Quick Wins and the System Data breakdown answer that in a few seconds, without anything leaving the app.

If you work with folders whose names you would rather not share with a model provider, leave agent access off, or point the agent at a specific folder instead of your whole home folder. Switching it off in Settings disconnects every client immediately and removes the local socket.