Claw Eyes

Problem

Many text-only models cannot process images, so screenshots, UI layouts, error messages, or code captures depend on manual description. Claw Eyes splits the task into three steps: capture a clipboard image, ask a vision model for a text description, and return that description to the main model, avoiding a full model switch.

How It Works

• Capture: reads the image from the system clipboard and compresses files over 800KB to at most 1920×1080.
• Analyze: prefers a user-configured OpenAI-compatible vision endpoint; if unavailable, it falls back to an MCP vision tool.
• Return: turns UI elements, errors, documents, tables, or code screenshots into text, letting the main model reason over that output.
  Endpoints, keys, and model names are provided through variables such as CLAW_EYES_API_URL, CLAW_EYES_API_KEY, and CLAW_EYES_VISION_MODEL; the skill does not hard-code vendor details.

Boundaries

It triggers only when the user asks to view an image, such as 'look at this', and mainly handles a single clipboard image. If no vision API is configured, the image should not leave the local machine. API_URL and VISION_MODEL must come from the same provider, and Linux needs xclip or wl-paste.