Your AI agent can read code, run tests, edit files — but it can't see your screen. Now it can.
You're staring at a broken UI. Claude is staring at your code. You describe the bug. Claude guesses. You paste a screenshot into the chat. Claude says "I can't view images in this interface."
The problem is simple: your AI agent is blind. It can read your code, run your tests, and edit your files — but it can't see what you see on screen.
Condor Eye fixes that. One overlay, one command, and your agent can see exactly what's in front of you.
Transparent overlay with capture frame, focus box, and pen annotations
A lightweight overlay that turns any screen into a data source for AI agents.
Capture any region via the transparent overlay frame. Win32 GDI for native speed. DPI-aware with automatic scaling.
Screenshots sent to Claude for visual analysis. Get structured JSON extractions or natural language descriptions.
Three endpoints on port 9050. Any tool that speaks HTTP can capture screens, locate windows, and check status.
Four Claude Code tools via Model Context Protocol. Your agent can see the screen without leaving the terminal.
Freehand annotations with pressure simulation. Circle areas, draw arrows, write notes — included in screenshots.
Draggable highlight region that tells the AI exactly where to look. Resize and position to isolate any UI element.
Without focus, the AI describes everything on screen. With the focus box, it zeroes in on exactly what you're pointing at.
I see a stock watchlist showing 6 tickers with prices and daily changes. There's a sidebar with market statistics. AAPL is at 189.42 up 1.23, MSFT at 412.87...
NVDA is down $12.40 (−1.37%) at 891.03 — the only significant red in the list. This is a sharp single-session drop. Previous close was ~903.43.
Rust backend with an axum HTTP server, WebView2 frontend, and a Node.js MCP bridge.
┌─────────────────────────────────────────────────────────────┐ │ WebView2 Frontend (src/) │ │ │ │ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌───────────┐ │ │ │ Capture │ │ Focus │ │ Draw │ │ Results │ │ │ │ Frame │ │ Box │ │ Canvas │ │ Panel │ │ │ └──────────┘ └──────────┘ └──────────┘ └───────────┘ │ ├─────────────────────────────────────────────────────────────┤ │ Rust Backend (src-tauri/src/) │ │ │ │ capture.rs ── Win32 GDI screen capture │ │ claude.rs ── Anthropic Vision API │ │ http_api.rs ── axum server (:9050) │ │ compare.rs ── diff engine │ │ truth.rs ── Redis ground truth │ │ windows.rs ── Win32 window enumeration │ ├─────────────────────────────────────────────────────────────┤ │ MCP Server (mcp/) │ │ │ │ Node.js stdio transport ── wraps HTTP API as │ │ Claude Code tools for terminal integration │ └─────────────────────────────────────────────────────────────┘
An axum server on port 9050. Anything that speaks HTTP can capture screens.
curl -X POST http://localhost:9050/api/capture \ -H "Content-Type: application/json" \ -d '{"prompt": "Describe what you see."}'
curl -X POST http://localhost:9050/api/locate \ -H "Content-Type: application/json" \ -d '{"query": "the Chrome browser"}'
curl http://localhost:9050/api/status {"running":true,"version":"0.1.0","api_key_configured":true,"model":"claude-haiku-4-5-20251001"}
Register once. Your agent gains four tools for screen interaction — without leaving the terminal.
| Variable | Required | Default |
|---|---|---|
| ANTHROPIC_API_KEY | Yes | — |
| CLAUDE_MODEL | No | claude-haiku-4-5 |
| CONDOR_EYE_PORT | No | 9050 |
| REDIS_URL | No | redis://127.0.0.1:6379 |
The repo ships with a Claude Code custom command. Clone it — and /screen works automatically.
Ctrl+3, then capture and analyze.Located at .claude/commands/screen.md — automatically available when you clone the repo.
Requires a Rust toolchain on Windows and an Anthropic API key.
Clone the repo and build with Tauri CLI.
git clone https://github.com/CondorCommodore/condor-eye
cd condor-eye
cargo tauri dev
Create a .env file with your API key.
echo "ANTHROPIC_API_KEY=sk-ant-..." > .env
Give Claude Code access to your screen.
claude mcp add --scope user \
condor-eye -- node ./mcp/index.js
Uses claude-haiku-4-5 by default — the fastest, cheapest Claude model.
A typical capture costs ~$0.003. That's roughly 300 captures per dollar.