clipboard-vision-mcp

An MCP server that gives clipboard vision to text-only LLMs. Cheap, fast models like DeepSeek V4 and GLM 5.1 are great for code but can't read images — normally that means saving every screenshot to disk and typing out a path just so the model can look at it. This server skips that: copy a screenshot, ask about it, done.

Under the hood it exposes *_from_clipboard tools. When the model needs to "see" something, it calls one of them, the server reads the image straight off the OS clipboard, sends it to a real vision model (Groq's free tier running Llama-4 Scout), and hands the text model back a description to reason over. The text-only model never touches pixels.

• Clipboard-first tools — analyze_clipboard, extract_text_from_clipboard, diagnose_error_from_clipboard, describe_ui_from_clipboard, code_from_clipboard — plus file-path equivalents for images already on disk
• Multi-OS clipboard support — Windows (Pillow/pywin32), macOS (Pillow + pngpaste fallback), Linux X11 (xclip) and Wayland (wl-paste)
• Security-hardened — file-extension allow-list and magic-byte validation so a prompt-injected model can't read arbitrary local files, a 20 MB size cap, and auto-deletion of clipboard temp files right after analysis
• Standard MCP server — works with Opencode, Claude Code, Cursor, Cline, Continue, or any MCP-capable client

Forked from itcomgroup/vision-mcp-server and rewritten around the clipboard-first workflow, cross-platform clipboard extraction, and a one-prompt install flow where you can paste an install prompt into any coding assistant and have it clone, set up the venv, and wire the MCP config end-to-end. Bilingual docs (English/French). Tested on Windows 11 + Opencode + DeepSeek V4 Pro.