supypowers

Run self-contained Python scripts (with uv script dependencies) as callable, schema-documented functions.

Designed for AI agents: Create, discover, and execute typed Python functions without managing environments.

What you can do with this

• Run a typed function in a script: validate/coerce JSON input with Pydantic, execute, emit JSON output.
• Generate documentation for all scripts in a folder as either JSON (for machines/LLMs) or Markdown (for humans).
• Scaffold new scripts instantly with supypowers new <name>

For AI Agents

👉 See AGENTS.md for the complete agent guide.

Quick start:

supypowers skills                    # What can I do?
supypowers run script:func '{...}'   # Do it
supypowers new name                  # Make something new

All scripts live in the powers/ folder. All output is JSON.

Prerequisites

• uv must be installed and on your PATH (the CLI shells out to uv run).

Install

Using uv (recommended)

# Install as a global CLI tool
uv tool install supypowers

# Or run without installing (ephemeral)
uvx supypowers skills

Using pip

pip install supypowers

Note: uv add supypowers adds it as a project dependency but won't expose the CLI globally. Use uv tool install for a global command, or uv run supypowers within your project.

Quickstart

# 1. Initialize a new powers/ folder
supypowers init

# 2. Create a new script from template
supypowers new my_tool

# 3. Edit powers/my_tool.py to implement your logic

# 4. Run it
supypowers run my_tool:my_tool '{"value": "test"}'

# 5. See all available functions
supypowers docs --format json

The init command creates powers/hello.py (example) and powers/hello.md (instructions for agents).

Try the bundled examples (when running from source):

supypowers run --examples exponents:compute_sqrt "{'x': 9}"
supypowers docs --examples --format md

Core concepts

Folder convention

All scripts live in a powers/ folder in your project root. This is hardcoded by convention to keep the CLI simple for agents.

my-project/
├── powers/
│   ├── hello.py
│   └── math.py
└── ...

Script dependencies (per-script, no venv)

Each script declares its dependencies using uv inline script metadata:

# /// script
# dependencies = [
#   "pydantic",
# ]
# ///

Supypowers reads those dependencies and runs your function in an isolated uv run environment (it does not import your script into the CLI process).

Superpower function contract

A function is considered a "supypower" if it matches this contract:

• Exactly one parameter, named input
• The input type is a Pydantic BaseModel
• Input data is validated/coerced via Pydantic (it must be an object mapping)
• Output can be any type; if it's a Pydantic model, it will be emitted in JSON-friendly form

See examples/exponents.py in the source repo for the canonical pattern.

Media tools convention

Tools that generate media files (images, audio, video, documents) should include a _media list in their output:

{
  "_media": [
    {"path": "/absolute/path/to/file.png", "type": "image"},
    {"path": "/absolute/path/to/audio.wav", "type": "audio"}
  ]
}

• Valid types: image, audio, video, document
• Paths must be absolute so consumers don't need to guess the working directory
• Output directory: Media tools should accept an output_dir parameter and create the directory if needed
• Legacy compatibility: Consumers should also check for _images/images (list of paths) and path with an image extension

Creating new scripts

Syntax

supypowers new <name>

Creates a new script powers/<name>.py from a template with:

• Dependency block already set up
• Input/Output Pydantic models
• Error handling pattern
• TODO comments showing where to customize

Example

supypowers new fetch_weather
# Creates powers/fetch_weather.py
# Edit the file, then run:
supypowers run fetch_weather:fetch_weather '{"city": "London"}'

Running functions

Syntax

supypowers run <script>:<function> <input_data> [--secrets ...]

The CLI looks for scripts in the powers/ folder of the current directory.

Input format (<input_data>)

• Preferred: strict JSON (e.g. {"x": 9})
• Also supported ("YAML-ish"): Python literal objects (e.g. {'x': 9}) via ast.literal_eval

Supypowers will then validate/coerce that object into your Pydantic model.

Secrets (--secrets)

You can pass secrets as:

• a dotenv file path (e.g. --secrets .env)
• inline KEY=VAL pairs (e.g. --secrets API_KEY=abc)

They are exposed to your script via environment variables.

Options

• --root <path>: Use a different root directory (default: current directory)
• --examples: Run from bundled examples instead of local powers/ folder

Examples

supypowers run hello:hello "{'name': 'World'}"
supypowers run math:compute_sqrt "{'x': 9}"
supypowers run math:compute_sqrt "{'x': 9}" --secrets .env --secrets API_KEY=abc

# Run from bundled examples (source install only)
supypowers run --examples exponents:compute_sqrt "{'x': 9}"
supypowers run --examples strings:reverse_string "{'s': 'abc'}"
supypowers run --examples dates:add_days "{'d': '2025-01-01', 'days': 10}"

Generating documentation

JSON docs (for machines / LLM context)

supypowers docs --format json
supypowers docs --format json --output docs.json

The JSON output is a list of:

• script: path to the script
• functions: list of functions with name, description, input_schema, output_schema
• error (optional): if a script failed to load/inspect

Markdown docs (for humans)

supypowers docs --format md
supypowers docs --format md --output docs.md

Options

• --root <path>: Use a different root directory (default: current directory)
• --examples: Document bundled examples instead of local powers/ folder
• --recursive: recurse into subfolders and include **/*.py
• --require-marker: only include functions explicitly marked (currently: decorator named superpower)

Generating skills document (for AI agents)

The skills command generates a comprehensive markdown document that teaches an AI agent everything it needs to know:

supypowers skills
supypowers skills --output SKILLS.md

The output includes:

• What supypowers is and how it works
• All available functions with their input schemas
• Example commands to run each function
• How to create new functions
• Rules and best practices
• How to use secrets

This is the recommended way to give an AI agent context about your supypowers.

Install supypowers on your PATH (so you can run supypowers ...)

If you don't want to prefix every command with uv run, install the CLI as a uv "tool":

Use uv tool install:

uv tool install supypowers
uv tool update-shell

Then open a new shell (or reload your shell config) and run:

supypowers init
supypowers run hello:hello "{'name': 'World'}"

Example scripts

See examples/ (uses uv script metadata + Pydantic models).

Troubleshooting

• uv not found: install uv and ensure it’s on your PATH.
• Docs command returns errors for a script: run uv run <script.py> yourself to verify the script’s dependencies/imports are valid.
• Input validation errors: ensure you pass an object mapping with keys matching your input model fields.

Notes (implementation)

• The CLI executes target scripts only via uv run <script.py> ... (no imports into the CLI process).