Iris — competitive positioning
Job
Image facts for agents: dimensions, format, metadata, opt-in OCR boxes, crops, and equal-size diffs.
Wedge
The default call measures the file and does not run OCR. Text is local Tesseract, and only when you set include_ocr or profile: "quality". A missing binary is a gap. Geometry stays citeable either way.
Local-first
Decode, crop, and diff run locally. OCR uses the tesseract binary on PATH. No API key. No generative vision model.
Peer anchors (learn; do not clone)
| Peer | Gap we exploit |
|---|---|
| Local image-search indexes | Retrieval of similar pictures, not dimensions, OCR boxes, or a citeable crop |
| Tesseract MCP servers that only dump text | A transcription with no file hash, no trust warnings, and no crop |
| Vision-model tools | A generated description: not deterministic, and not a measured pixel box |
Non-goals
- A cloud account as the default path
- A multi-product repo whose only purpose is to pool attention
- A generated caption as the evidence
Install
bash
npx -y @sylphx/iris@sylphx/iris is the install. It speaks MCP over stdio.