farscry

See through any image. Converts screenshots into structured context that any AI agent can understand. No vision API. No cloud. One binary.

farscry (n.) — a magical artifact that reveals what is hidden at a distance.

Problem: Devin CLI, Claude Code, and Cursor struggle with images. Give them a screenshot, they guess. Give them farscry output, they understand.

Benchmark

Tool	Time	Cost/image	Works offline
farscry	~80ms	$0	✅
Mistral OCR	~1.2s	$0.001	❌
Claude Vision	~1.8s	$0.003	❌
Tesseract	~300ms	$0	✅

farscry wins on cost + agent-readiness. Tesseract wins on nothing useful for agents.

Install

# npm
npm install farscry

# pip
pip install farscry

# curl (any platform)
curl -fsSL https://farscry.dev/install | sh

Quick start

import { extract } from 'farscry'

const context = await extract('screenshot.png')
// → structured TOON context, ready for any agent
console.log(context)

How it works

[image] → [binarize] → [layout detect] → [OCR per region] → [classify] → [TOON output]

Detects: error messages · UI fields + values · terminal output · conversations · config screens

Output (TOON format)

TOON uses 40% fewer tokens than JSON with higher LLM accuracy.

screen_type: error_screenshot
error: "Payment limit exceeded"
fields:
  max_value | 1500
  status    | Active
  period    | Monthly

Integrations

Works with any agent that accepts text:

• Devin CLI: pipe output into your message
• Claude Code: farscry screenshot.png | claude
• Zendesk / Intercom: inject into ticket body before AI reads it
• MCP server: farscry serve --mcp

License

Apache 2.0