ionclaw-org/ionclaw/vision

Vision

Analyze images using the vision tool. When a user uploads an image, you receive an [image attached: path] annotation — use that path as the path parameter.

Parameters

How It Works

The vision tool loads an image from any source, resizes it to an LLM-friendly preview (max 1024px wide, JPEG), and passes it to the model as a visual content block. The model then "sees" the image and can describe, analyze, or answer questions about it — no external vision API required.

Common Workflows

Describe a local image

vision path="public/media/photo.jpg"

Analyze a screenshot

vision path="public/screenshots/page.png" question="What error message is shown?"

Analyze an image from a URL

vision url="https://example.com/chart.png" question="What trends does this chart show?"

Read text from an image (OCR)

vision path="public/documents/scan.png" question="Extract all text from this image"

Analyze a browser screenshot (with output_path)

browser action="screenshot" output_path="public/screenshots/page.png"
vision path="public/screenshots/page.png" question="What is displayed on this page?"

Use with base64 data

vision base64="data:image/png;base64,iVBOR..." question="What is this?"

Supported Formats

JPEG, PNG, GIF, WebP, SVG, BMP, ICO, TIFF, AVIF.

Max image size: 20MB.

Notes

• Provide exactly one image source: path, url, or base64
• MIME type is auto-detected from file extension or Content-Type header
• Use question to focus the analysis on specific aspects
• The model's vision capabilities depend on the underlying LLM (Claude, GPT-4V, Gemini all support vision)