autocli-llm 0.2.0

Parse CLI arguments from natural language using LLMs

autocli

Parse CLI arguments from natural language descriptions using small LLMs. Write your CLI examples in any format - the LLM understands them all!

Features

• Flexible Input Format - Understands any CLI syntax: --name=value, -n value, name:value, etc.
• 100% Local - No API calls, runs entirely offline on your machine
• Lightweight - Flan-T5-small model (~250MB) runs on CPU, no GPU required
• One-time download - Model downloads automatically on first use, then works offline forever
• Natural language CLI description parsing
• Automatic argument type inference
• Support for positional and named arguments
• Built-in help generation
• Fallback regex parser for reliability
• Returns arguments as NamedTuple-like objects

Installation

pip install autocli-llm

Note: The package imports as autocli after installation.

First Run: On first use, the package will automatically download the Flan-T5-small model (~250MB) from HuggingFace. This requires an internet connection but only happens once - the model is cached locally for all future use.

Quick Start

import autocli

args = autocli.parse(
    """
    This app greets someone with excitement
    
        $ python greet.py --name Alice --excitement 3
        Hello, Alice!!!
    """
)

print(f"Hello, {args.name}{'!' * args.excitement}")

Format Flexibility & Type Intelligence

The LLM understands any reasonable CLI format and intelligently infers data types from context!

Mixed Syntax Styles

import autocli

# All of these formats work!
args = autocli.parse("""
    A config tool that accepts various argument formats
    
    Examples:
        $ mytool --name="Alice" --age=30 --verbose
        $ mytool name:Bob age:25 -v
        $ mytool -n Charlie -a 35 --verbose=true
        $ mytool --name David age 40
""")

# The LLM figures out that these all refer to the same arguments:
# - name (string)
# - age (number) 
# - verbose (boolean flag)

Intelligent Type Inference

# The LLM automatically understands different data types
args = autocli.parse("""
    Scientific tool with various data types
    
    Examples:
        $ analyze --threshold 3.14159 --iterations 1000 --verbose yes
        $ analyze --ratio 0.5 --count 42 --debug true --days Mon,Wed,Fri
        $ analyze threshold:2.71 iterations:500 verbose:1 days:[Monday,Friday]
""")

# Automatically infers:
# - threshold: float (3.14159)
# - iterations: integer (1000)
# - verbose: boolean (yes/no, true/false, 1/0)
# - days: list (comma-separated or bracket notation)

Boolean Values - All Formats Understood

# The LLM understands all common boolean representations
args = autocli.parse("""
    Build tool with boolean flags
    
    Examples:
        $ build --optimize=true --debug=false    # true/false
        $ build --compress=yes --verbose=no      # yes/no  
        $ build --minify=1 --sourcemaps=0        # 1/0
        $ build --production --no-warnings       # flag presence
""")

Lists and Collections

# Various list formats are automatically recognized
args = autocli.parse("""
    Scheduler that accepts multiple values
    
    Examples:
        $ schedule --days Monday,Tuesday,Wednesday,Thursday,Friday
        $ schedule --days "Mon Tue Wed Thu Fri"
        $ schedule --days=[Monday,Wednesday,Friday]
        $ schedule --times 9:00,12:00,15:00,18:00
""")

Examples

Positional args

import autocli

args = autocli.parse(
    """
    This app adds two numbers and prints their sum

        $ python sum.py 1 2
        3
    """
)

print(args[0] + args[1])

Named args and defaults

import autocli

args = autocli.parse(
    """
    This app prints the file with the longest name in the given directory (default: PWD)

        $ ls /tmp
        a.txt ab.txt abc.txt

        $ python longest_filename.py --path /tmp
        abc.txt
    """
)

print(f'finding the longest filename in directory: {args.path})

Named args wth numeric values and allowed ranges

# greet.py

import autocli

args = autocli.parse(
    """
    This app greets the given `name` defaulting to "Earthling" and
    appends `excitement` number of exclamation marks.
    """
)

print(f'{args.name}{'!' * args.excitement}')

$ python greet.py --help
Greets the given name.

Options:

    -n NAME
    --name NAME

        NAME is who is being greeted.

        Default: "Earthling"

    -e EXCITEMENT
    --excitement EXCITEMENT

        EXCITEMENT is a positive integer that contols
        the number of exclamation marks.

        Default: 1

$ python greet.py -e -3
ERROR: The value for EXCITEMENT is too low. The lowest value is 1,
which is also the default.

    -e EXCITEMENT
    --excitement EXCITEMENT

        EXCITEMENT is a positive integer that contols
        the number of exclamation marks.

        Default: 1

Did you mean?
    python greet.py -e 3

$ python greet.py --excitement 42
Greetings, Earthling!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!

How It Works

1. Intelligent Parsing: The LLM understands intent, not just syntax - it recognizes that --name=foo, -n foo, name:foo, and name foo all mean the same thing
2. Local LLM Processing: Runs entirely on your machine - no API calls or internet connection required after initial download
3. Structured Output: Converts any CLI format to structured argument specifications
4. Automatic ArgParse: Generates standard Python argparse configuration
5. NamedTuple-like Access: Returns arguments as an object supporting both attribute and index access

Privacy First: All processing happens locally on your device. Your code and CLI descriptions never leave your machine.

Performance: The small model (~250MB) runs efficiently on CPU - no GPU required. Works on laptops, desktops, and even resource-constrained environments.

Advanced Usage

Custom Model

from autocli.parser import LLMParser

# Use a different model
parser = LLMParser(model_name="google/flan-t5-base")

Accessing Arguments

args = autocli.parse(description)

# Named arguments via attributes
print(args.name)
print(args.port)

# Positional arguments via indexing
print(args[0])  # First positional arg
print(args[1])  # Second positional arg

Requirements

• Python 3.8+
• PyTorch 2.0+ (CPU version is sufficient)
• Transformers 4.30+
• ~250MB disk space for model cache
• Internet connection (first run only, for model download)

License

MIT

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.