plain-sight Handbook

plain-sight is a local, MIT-licensed image describer. Point it at an image and an AI (Florence-2) says what it sees — as a one-line caption, a detailed paragraph, or the text read off the pixels. It ships as an MCP server for Claude and a CLI for shells and scripts, sharing one engine.

When to reach for it

You want…	Use
“What is in this image?” in prose	describe_image / plain-sight describe
.txt caption sidecars for a LoRA training set	describe_batch / plain-sight batch
The text visible inside an image	read_text / plain-sight ocr
“Does this image contain X?” as a score	ai-eyes-mcp — not this tool

The sibling rule

plain-sight is one half of a deliberate pair:

	ai-eyes-mcp	plain-sight
Job	judges images	describes images
Model	SigLIP2 (discriminative)	Florence-2 (generative)
Output	calibrated scores	prose / OCR / caption files
Failure mode	can’t narrate	can hallucinate detail

Descriptions are generative: fluent, usually accurate, and capable of inventing detail. plain-sight makes output reproducible (deterministic decoding), not guaranteed true. When a specific claim matters, verify it with ai-eyes — the two tools are different model families by design, so one can check the other.

Architecture in one glance

engine.py    Florence-2 wrapper — no MCP dependency, lazy load,
             validation BEFORE the model loads.
sidecars.py  The training-data contract, pure stdlib: basename
             pairing, bare concatenation, directory expansion.
server.py    FastMCP layer: 5 tools, structured errors.
cli.py       argparse CLI: 5 commands, real exit codes.

A cloud sibling of the same contract runs on Comfy Cloud as the caption-florence2-v1 workflow (one image per job, metadata-rider use). This tool is the bulk lane: on a modern GPU it captions at roughly one image per second, entirely locally.