Back to Maus

Maus MCP

Your Maus clipboard, inside Claude Code, Cursor and Codex — so the agent can read your history, search it, and hand you text already clean and ready to paste.

Install (Claude Code)

claude mcp add -s user maus -- npx maus-mcp@latest

macOS · Node 20+ · Maus installed and run once.

Restart Claude Code. Done.

Install (Cursor)

Edit ~/.cursor/mcp.json:

{
  "mcpServers": {
    "maus": {
      "command": "npx",
      "args": ["maus-mcp@latest"]
    }
  }
}

What you can ask the agent

• "Draft a reply to this email and put it in my Maus" — the agent saves it clean. You paste into Gmail with no monospace, no markdown asterisks, no LLM em-dashes.
• "Give me a SQL query for users active this week, save it to Maus" — you paste straight into Supabase / TablePlus / psql.
• "What was the wifi password I copied yesterday?" — agent searches your history, including OCR text from screenshots.
• "What have I copied from Cursor today?" — agent filters by source app. Maus Pro.
• "Find the screenshot of the error from this morning" — search by OCR text; agent gets back the image + extracted text.
• "Rename item 1247 to 'wifi password — home router'" — agent organizes your history with meaningful names.

Tokens compound — that's the whole point

An image pasted into a conversation isn't a one-time cost. It joins the context window and the model re-processes it on every subsequent turn. A 2,691-token screenshot pasted at minute 30 of your session gets re-read on every turn for the next two hours.

That's the difference Maus MCP makes — not the cost of one screenshot, but the compounding cost across a long session.

How the agent sees a 1920×1080 screenshot	Per image	4 imgs · 30 turns
Pasted directly	2,691	322,920
get(id, want:"visual") — ≤800px JPEG	~493	59,160
get(id, want:"text") — OCR text only	~150	18,000

Visual tokens = ⌈width/28⌉ × ⌈height/28⌉ per Anthropic's vision docs (Claude Opus 4.7+). OCR is a typical error-screenshot worth of text.

322k vs 18k tokens re-processed across the rest of the session. That's the gap between "the agent has room to keep reasoning" and "Claude Max says no for the next 4 hours." Prompt caching cuts dollar cost; it doesn't reclaim window space. The agent picks the path: visual for layout and design, OCR for error messages and code.

Why this matters

Anything the agent emits in chat — emails, code, queries, drafts — arrives in your destination app dragging monospace font, markdown junk, and awkward line breaks. Maus MCP delivers it clean. The agent calls add_item, the text lands in your clipboard history, you press ⌘⇧V and paste it where you actually wanted it.

It also stops being "the user is the copy-paste cable between Claude and the rest of their Mac." The agent reads your clipboard directly when it needs context — terminal output, an error screenshot, a snippet from yesterday's session.

Privacy

Clipboard content never leaves your Mac. The MCP is a local-only bridge — your Claude / Cursor / Codex talks to your Maus history file directly. Maus Free is bounded to the last 24 hours; Maus Pro unlocks full history and source-app filtering.

Source & bugs

github.com/mnlt/maus-mcp — open source (MIT). Issues, PRs, schema, all tools documented there.