jsonargparse

jsonargparse is a Python library that extends the standard argparse to build command-line interfaces and configuration systems with minimal boilerplate. It derives arguments automatically from function signatures, dataclasses, and type hints, and can populate them from the command line, YAML/JSON/TOML/Jsonnet config files, and environment variables in a single unified parser.

Beyond argument parsing, jsonargparse supports nested namespaces, class and function instantiation from config (a form of dependency injection), argument linking, sub-commands, and validation via type hints and Pydantic. It is the configuration engine behind tools like PyTorch Lightning’s LightningCLI, where its ability to instantiate objects directly from config makes complex applications configurable without custom parsing code.

What You Get

• An ArgumentParser that derives arguments from type hints, signatures, and dataclasses
• Unified parsing from CLI, YAML/JSON/TOML/Jsonnet config files, and environment variables
• Class and function instantiation from config for dependency-injection-style wiring
• Nested namespaces, sub-commands, and argument linking between parameters
• Type-based validation with support for Pydantic and attrs models

Common Use Cases

• Building configurable CLIs for machine-learning training scripts and research code
• Loading layered configuration from files and environment variables into typed objects
• Instantiating application components (models, optimizers, data modules) directly from config

Under The Hood

Architecture - The public ArgumentParser in jsonargparse/_core.py builds on argparse and delegates to focused modules: _signatures.py and _parameter_resolvers.py introspect callables to add arguments, _typehints.py validates and coerces values, _loaders_dumpers.py/_jsonnet.py/_jsonschema.py handle config formats, _instantiation.py constructs objects from config, and _link_arguments.py and _subcommands.py wire parameters and commands together.

Tech Stack - Pure Python with PyYAML for config parsing and optional integrations (Jsonnet, jsonschema, Pydantic, attrs, fsspec, omegaconf) enabled through extras. It ships py.typed for full typing support and targets a wide range of Python versions.

Code Quality - The library is modular with one responsibility per underscore-prefixed module, is fully type-annotated, and is developed with strong CI including coverage on Codecov and static analysis on SonarCloud. Deprecations are isolated in _deprecated.py, signaling careful backward-compatibility management.

API Design - It intentionally mirrors the familiar argparse API, so add_argument, parse_args, and sub-parsers work as expected, while higher-level methods like add_class_arguments, add_dataclass_arguments, and instantiate_classes remove large amounts of boilerplate. Deriving a CLI from a typed function is often a single call.