clipboard-mcp

An MCP (Model Context Protocol) server that reads and writes your system clipboard — tables, plain text, code, JSON, URLs, images, and more. Preserves structure when possible (e.g. spreadsheet row/column layout) and returns non-tabular content cleanly.

The Problem

When you copy cells from Google Sheets or Excel and paste into Claude's chat input, the tabular structure (rows and columns) is destroyed — it arrives as a flat string with no delimiters. And for non-tabular content, there's no easy way to say "look at what I copied" without manually pasting.

The Solution

Instead of pasting, tell Claude to "read my clipboard". This MCP server reads the clipboard directly and handles any content type: spreadsheet tables (returned as Markdown, JSON, or CSV), images (returned as image content), code snippets, JSON data, URLs, rich HTML text, and plain text. It can also write to the clipboard, so Claude can copy results back for you to paste elsewhere.

Tools

Setup

Prerequisites

Python 3.11+ and one of: uv (recommended), pipx, or pip.

You also need a platform-specific clipboard tool:

Platform status: Linux with Wayland is tested on real hardware. X11 has unit test coverage but has not been verified on a live X11 session. macOS and Windows implementations are complete but untested on real hardware — they should work but may have edge cases. Bug reports and PRs are very welcome.

Configure Claude Desktop

Add one of the following to your Claude Desktop config file:

• Linux: ~/.config/Claude/claude_desktop_config.json

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

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

Option A: uvx — no clone needed (recommended)

Requires uv. Installs and runs directly from PyPI:

{
  "mcpServers": {
    "clipboard": {
      "command": "uvx",
      "args": [
        "--from", "clipboard-mcp-server",
        "clipboard-mcp"
      ]
    }
  }
}

Option B: pipx — no clone needed

Requires pipx. Same idea, different tool manager:

{
  "mcpServers": {
    "clipboard": {
      "command": "pipx",
      "args": [
        "run", "--spec", "clipboard-mcp-server",
        "clipboard-mcp"
      ]
    }
  }
}

Note: The PyPI package name is clipboard-mcp-server but the command it installs is clipboard-mcp.

Option C: Local clone with uv

git clone https://github.com/cmeans/clipboard-mcp.git
cd clipboard-mcp
uv sync

{
  "mcpServers": {
    "clipboard": {
      "command": "uv",
      "args": [
        "run",
        "--directory", "/path/to/clipboard-mcp",
        "clipboard-mcp"
      ]
    }
  }
}

Option D: Local clone with pip

git clone https://github.com/cmeans/clipboard-mcp.git
cd clipboard-mcp
pip install -e .

{
  "mcpServers": {
    "clipboard": {
      "command": "clipboard-mcp"
    }
  }
}

Note: Option D depends on clipboard-mcp being on your PATH. Since Claude Desktop launches outside a shell, this may not work reliably. Prefer options A–C.

Restart Claude Desktop after editing the config.

Environment variables

Environment variables can be passed via the "env" key in the Claude Desktop config.

Wayland auto-detection: The server scans for wayland-* sockets in $XDG_RUNTIME_DIR (itself falling back to /run/user/<uid>). Neither WAYLAND_DISPLAY nor XDG_RUNTIME_DIR need to be set manually in most cases. Override them if auto-detection doesn't work (e.g. multiple compositors, non-standard socket path, or a containerized environment):

{
  "mcpServers": {
    "clipboard": {
      "command": "uvx",
      "args": [
        "--from", "clipboard-mcp-server",
        "clipboard-mcp"
      ],
      "env": {
        "WAYLAND_DISPLAY": "wayland-0",
        "XDG_RUNTIME_DIR": "/run/user/1000"
      }
    }
  }
}

Usage

1. Copy anything — spreadsheet cells, code, text, a URL, JSON, an image, etc.

2. In Claude, say something like:

   • "Paste my clipboard"

   • "Read my clipboard"

   • "What's on my clipboard?"

   • "Read what I copied"

   • "I copied some data, take a look"

   • "Show me the table I copied"

   • "Give me that data as JSON" (uses output_format=json)

   • "Convert my clipboard to CSV" (uses output_format=csv)

3. Claude will call clipboard_paste and return the content.

Writing to the clipboard — Claude can also copy results back:

• "Copy that to my clipboard"

• "Put the cleaned-up JSON on my clipboard"

• "Copy just the SQL query"

Tips for reliable triggering

The server includes instructions that tell Claude when to check the clipboard, but host models vary in how consistently they follow MCP server instructions. If Claude isn't picking up on your intent, try these approaches:

Be explicit — these phrases work most reliably:

• "Paste my clipboard"

• "Read what I copied"

• "What's on my clipboard?"

Reference the clipboard when data is missing — Claude is instructed to check the clipboard when you refer to data that isn't in the conversation, but this doesn't always fire. If Claude asks you to provide the data instead of reading it, nudge it:

• "It's on my clipboard — paste it"

• "Check my clipboard"

Add a system prompt hint — If you have access to a custom system prompt (e.g. in a Claude Desktop project), you can reinforce the behavior:

When the user references data not present in the conversation, check the clipboard using clipboard_paste before asking them to provide it.

Content handling

Table output formats

When the clipboard contains tabular data, output_format controls the format:

• Markdown (default) — renders as a table in the conversation

• JSON — array of objects keyed by the header row (single-column: flat array)

• CSV — comma-separated values

Development

# Install with dev dependencies
uv sync --extra dev

# Run tests
uv run pytest

# Run the server directly (stdio mode)
uv run clipboard-mcp

# Run with debug logging (logs backend detection, tool parameters, clipboard reads)
uv run clipboard-mcp --debug

# Test with MCP Inspector
uv run mcp dev src/clipboard_mcp/server.py

Debug logging can also be enabled via the CLIPBOARD_MCP_DEBUG=1 environment variable, which is useful when the server is launched by Claude Desktop (see Environment variables).

Project Structure

clipboard-mcp/
├── src/clipboard_mcp/
│   ├── __init__.py          # Package version
│   ├── server.py            # MCP server, tool definitions & debug logging setup
│   ├── clipboard.py         # Platform-agnostic clipboard backend (Wayland auto-detection)
│   ├── parser.py            # HTML table parser, formatters, content detection
│   └── instructions/        # Tool & server descriptions (loaded at startup)
│       ├── server.md        # Server-level MCP instructions
│       ├── clipboard_paste.md
│       ├── clipboard_copy.md
│       ├── clipboard_read_raw.md
│       └── clipboard_list_formats.md
├── tests/
│   ├── test_parser.py       # Parser & formatter tests
│   └── test_server.py       # MCP server, clipboard backend & Wayland detection tests
├── pyproject.toml           # Project metadata, dependencies & pytest config
├── CLAUDE.md                # Claude Code guidance
├── LICENSE                  # MIT
└── README.md

How It Works

Reading (clipboard_paste)

1. Clipboard read: The server calls the platform's clipboard tool (wl-paste, xclip, pbpaste, or PowerShell) to read the clipboard.

2. Table detection: Tries text/html first — Google Sheets and Excel put <table> markup on the clipboard. Parsed with Python's built-in html.parser (no external dependencies). Falls back to text/plain tab-separated values.

3. Table found? Format as Markdown, JSON, or CSV (per output_format) and return.

4. Non-tabular text: If no table is found, the plain text (or HTML-extracted text) is classified as JSON, URL, code, or plain text and returned with appropriate formatting (pretty-printed JSON, fenced code blocks, etc.). Content is truncated at 50KB.

5. No text content? Check for image formats on the clipboard. If found, read the image data and return it as base64-encoded image content that Claude can see.

6. Audio/video: Detected and reported, but content is not returned.

Writing (clipboard_copy)

The server writes text to the clipboard using the platform's clipboard tool (wl-copy, xclip, pbcopy, or PowerShell). Currently supports text content only.

Features

• Smart Table Parsing — Automatically detects structured data on the clipboard (from Excel, Google Sheets, or web tables). Prioritizes text/html and text/plain (TSV) to preserve formatting, delivering clean Markdown, JSON, or CSV instead of a disorganized text string.

• Multimodal Image Support — Reads image data directly from the clipboard. When a user "pastes" a PNG or JPEG, the server captures the raw bytes and returns a standard MCP ImageContent block, enabling vision-capable models to analyze screenshots or graphics instantly.

• Clipboard Write — Copies text back to the system clipboard so Claude can hand off cleaned-up JSON, reformatted code, SQL queries, or any text for the user to paste elsewhere.

• Automatic Content Classification — Intelligently identifies and formats specific data types:

  • JSON — Pretty-printed for better model readability.

  • Code — Wrapped in fenced code blocks.

  • URLs — Extracted and presented for easy navigation or fetching.

• Cross-Platform Architecture — Uses native system binaries (wl-paste/wl-copy for Wayland, xclip for X11, pbpaste/pbcopy for macOS, PowerShell for Windows) to minimize dependencies and maintain a lightweight footprint.

• Zero-Config Execution — Designed for modern workflows using uvx, allowing the server to run instantly without manual environment management or permanent installation.

• Diagnostic Tools — clipboard_list_formats lets users inspect available MIME types on the clipboard, and clipboard_read_raw returns raw content for any text-readable MIME type (including image/svg+xml, application/json, and application/xml) for troubleshooting and verification.

Limitations

• Audio and video are not supported. If the clipboard contains audio or video, the server will report what format is present but cannot return the content.

• Clipboard write is text-only. clipboard_copy writes plain text. Writing images or other binary data is not yet supported.

• Text content is truncated at 50KB to avoid overwhelming the model's context window.

• macOS and Windows are untested (see Platform status). Implementations are complete but may have edge cases on real hardware.

Acknowledgments

This project was designed and built in collaboration with Claude Code (Anthropic's CLI for Claude). Architecture, design decisions, and release management were driven by the human; implementation, testing, code review, and documentation were delegated conversationally, with Claude writing code, catching stale docs, and filling test coverage gaps across every commit.

License

MIT — see LICENSE.

Copyright (c) 2026 Chris Means