macroni 0.1.9

Macroni interpreter

Macroni

DSL for GUI automation with OCR, template matching, and screen interaction. Randomness is baked into all operations, such as mouse movement and playback.

Installation

1. Install Python Dependencies

pip install macroni

Note that pytorch for CPU is installed by default. If your NVIDIA GPU supports CUDA, you will have to manually reinstall pytorch for GPU.

2. System Permissions

macOS: System Preferences → Security & Privacy → Accessibility

Grant permission to Terminal/Python to control your computer

3. VSCode Extension

For syntax highlighting, please add the extension to vscode: Macroni Language Support

4. Usage

Basic Command

Interactive

macroni

Execute file

macroni --file script.macroni

Note: the first time running will take time because pytorch is a massive library.

Debug Mode

Enable interactive debugging with breakpoints:

# Enable debug mode
macroni --file script.macroni --debug

# Set breakpoints at specific lines
macroni --file script.macroni --debug --breakpoints 10 --breakpoints 25

# Or use short flags
macroni -f script.macroni -d -b 10 -b 25

Debug Commands:

• n (next) - Execute next line
• c (continue) - Continue to next breakpoint
• p <var> (print) - Print variable value
• q (quit) - Exit debugger
• eval <expression> - Evaluate expression

OCR Text Search

Find and click text on screen:

# Capture region once, reuse forever (cached)
region = @capture_region("login_area", false);

# Find text in region
results = @ocr_find_text(region, 0.8, "Login", 1.0);

if @len(results) > 0 {
    text, conf, bbox = results[0];
    x1, y1 = bbox[0];
    @mouse_move(x1, y1, 500, true);
    @left_click();
}

Note: This operation is fairly slow without CUDA. If using a CPU, keeping the region as small as possible will improve performance greatly.

OCR Functions:

• @capture_region(key, overwrite) - Interactive region capture with caching
  • Hover top-left → Enter → bottom-right → Enter
  • Returns (x1, y1, x2, y2) tuple
• @ocr_find_text(region, min_conf, filter, upscale) - Find text via OCR
  • region: From @capture_region() or null for full screen
  • min_conf: 0.0-1.0 confidence threshold
  • filter: Text substring to search for (case-insensitive)
  • upscale: 1.0 = no scaling, 0.5 = faster, 2.0 = tiny text
  • Returns [(text, conf, [[x1,y1],[x2,y2],[x3,y3],[x4,y4]]), ...]

Template Matching

@set_template_dir("./templates");
x, y = @find_template("login_button");

if x != null {
    @mouse_move(x, y, 1000, true);
    @left_click();
}

Templates folder structure:

templates/
  └── login_button/
      ├── ex1.png
      └── ex2.png

Language Basics

# Variables & types
x = 10;
name = "test";
coords = (100, 200);
items = [1, 2, 3];

# Booleans
enabled = true;   # true = 1
disabled = false; # false = 0

# Destructuring
x, y = (100, 200);
text, conf, bbox = results[0];

# Control flow
if x > 10 {
    @print("yes");
} else {
    @print("no");
}

while x < 100 {
    x = x + 1;
}

# Functions
fn click_button(x, y) {
    @mouse_move(x, y, 500, true);
    @left_click();
}

Key Functions

Mouse/Keyboard:

• @mouse_move(x, y, speed, human_like)
• @left_click()
• @press_and_release(delay_ms, ...keys)

Screen:

• @get_coordinates(label, use_cache) - Interactive coordinate capture
• @get_pixel_at(x, y) - Returns (r, g, b)
• @check_pixel_color(x, y, radius, r, g, b, tolerance)

Timing:

• @wait(ms) or @wait(ms, random_range)
• @time()

Recording:

• @record(name, start_btn, stop_btn, squash_distance) - Record mouse/keyboard (squash_distance defaults to 50 pixels)
• @playback(name, stop_btn) - Replay recording

Lists:

• @len(list), @append(list, item), @pop(list, index), @shuffle(list)

Human-Like Randomness

Macroni incorporates randomness to avoid detection and mimic natural user behavior:

Mouse Movement:

@mouse_move(x, y, 1000, true);  # human_like=true enables randomness

• Uses smoothstep for natural acceleration/deceleration
• Adds big random arcs (bulge peaks randomly between 25-75% of path)
• Mixes in sin wave wobble with random phase drift
• Each movement takes a unique path, even to the same destination

Wait Times:

@wait(1000, (100, 300));  # Base delay + random 100-300ms

• Optional random range parameter adds variability to timing
• Prevents predictable patterns in automation

Recording Playback:

• Mouse coordinates are compressed during recording (50ms buckets)
• Playback connects positions using smoothstep with randomized arcs and sin waves
• Each playback generates a different mouse path, never the exact same trajectory
• Maintains original timing while applying human-like movement between recorded points

Cache Files

Auto-created in working directory:

• regions_cache.json - OCR regions
• coordinates_cache.json - Captured coordinates
• pixel_colors_cache.json - Captured colors
• recordings_cache.json - Recorded macros