Merged into anymd (2026-09-25). anymd reads PDFs, Office files, EPUB, web pages, images and video into clean Markdown for AI agents: npx -y @sylphx/anymd. This repository is archived.

Iris

Image facts with pixel-level proof.

Dimensions, format, and metadata by default. Local Tesseract OCR only when requested.

npm @sylphx/iris · bin iris · MCP io.github.SylphxAI/iris

The problem

A screenshot is not a paragraph. An agent that guesses the text, the size, or the pixels it never measured will cite something it cannot show.

The difference

The ask	Iris returns
“Read this screenshot.”	dimensions, format, metadata, and trust warnings. No OCR unless you ask.
“Read the text.”	OCR lines with boxes, only when `include_ocr` is true or `profile` is `quality`.
“What changed between these two captures?”	changed pixels and a changed box, when both images are the same size.
“Crop this region.”	pixel bounds and a hash. PNG bytes only if `include_region_image` is true.

Iris does not run a generative vision model. A missing Tesseract binary is a gap, not a made-up transcription.

Install

Terminal

npx -y @sylphx/iris

That starts a stdio MCP server. No API key. OCR needs the tesseract binary on PATH.

Your client	Setup
Any agent / CLI	`npx -y @sylphx/iris`
Claude Code	`claude mcp add iris -- npx -y @sylphx/iris`
Claude Desktop / Cursor / VS Code / Codex	`"command": "npx", "args": ["-y", "@sylphx/iris"]`

config.json

{
  "mcpServers": {
    "iris": { "command": "npx", "args": ["-y", "@sylphx/iris"] }
  }
}

A read, then text only if you ask

config.json

{ "path": "/absolute/path/to/screenshot.png" }

fast, or no profile, returns dimensions, format, metadata, and trust warnings. It does not run OCR.

config.json

{
  "path": "/absolute/path/to/screenshot.png",
  "include_ocr": true
}

include_ocr: true, or profile: "quality", runs local Tesseract. Explicit include_ocr: false wins over quality. If Tesseract is missing, status stays ok and gaps includes OCR_UNAVAILABLE.

Reference: tools · defaults

Tools

Tool	When to call it
`read_image`	One local image. Geometry and metadata. OCR only when requested. An optional `region` crops inside this read.
`image_probe`	Format, dimensions, pixel count, and source hash. No OCR and no crop.
`crop_region`	One citeable region: bounds and hash. Set `include_region_image` for PNG bytes. No OCR.
`compare_images`	Changed pixels between two images of equal dimensions. No OCR.

What it will not pretend

GPS fields are always redacted. There is no switch that puts them back.
The default file cap is 32 MiB (33,554,432 bytes). A cropped region may not exceed 67,108,864 pixels.
compare_images requires equal dimensions. It does not resize either image.
OCR lines are {text, bbox, confidence}. The default language is eng.
Oversized or undecodable files fail with an error. They are not returned as a successful read.

Companion MCP tools

Product	Job
Citra	PDF answers with page-level proof
Cue	Video timelines and timestamp evidence
Spine	Repository architecture and impact
Locus	Exact code-chunk retrieval
Lookout	Web research with source excerpts

Each product is independent. Install only the tools your agent needs.

Documentation


Website	sylphxai.github.io/iris
Quickstart	Install and first call
Compare	What Iris does, and what it refuses

Development

bash

bun install
bun test
bun run check
bun run docs:build
cargo test

License

MIT