argklass 2.2.0

Argparse generator from dataclass

Inspired by Simple Parsing, simplified and extended to build extensive, extendable command line interface without much code.

pip install argklass

Features

• Automatic cli discovery and command plugin

  # Folder structure
  project/cli/
  ├── __init__.py         <= empty
  ├── editor/
  │   ├── __init__.py     <= ParentCommand(editor)
  │   ├── cook.py         <= Command(cook)
  │   ├── client.py       <= Command(client)
  │   ├── game.py         <= Command(game)
  │   └── open.py         <= Command(open)
  └── uat/
     ├── __init__.py     <= ParentCommand(uat)
     ├── localize.py     <= Command(localize)
     └── test.py         <= Command(test)
  
  #  editor/__init__.py
  from argklass.command import ParentCommand
  
  class Editor(ParentCommand):
     name = "editor"
  
  
  COMMANDS = Editor
  
  # cook.py
  from argklass.command import Command
  
  class Cook(Command):
     name = "cook"
  
     @staticmethod
     def execute(args) -> int:
        print("cook")
  
  COMMANDS = Cook
  
  #
  cli = CommandLineInterface(project.cli)
  cli.run()
  
  # or
  cli.run(["editor", "cook", "--help"])
  
  #
  cli editor cook --help
  cli uat localize --help

• New Argument format

  • able to show the entire command line interface with all its subparsers

  • new format mirror dataclass syntax

  editor                                           Set of commands to launch the editors in different modes
     server                                       Parameters added to the Map URL
     game                                         docstring ...
     client                                       docstring ...
     resavepackages                               docstring ...
     cook                                         docstring ...
     ml                                           Launch unreal engine with mladapter setup
     editor                                       Other arguments
     open                                         docstring ...
     localize                                     docstring ...
     worldpartition                               Convert a UE4 map using world partition
     -h, --help                                   Show help
  engine                                           Set of commands to manage engine installation/source
        add                                          docstring ...
        update                                       Update the engine source code
  format                                             docstring ...
        --profile: str                               docstring ...
        --file: str                                  docstring ...
        --fail_on_error: bool = False                docstring ...
        --col: int = 24                              docstring ...

• Compact argparse definition

  def workdir():
     d = os.getcwd()
     if os.access(d, os.W_OK):
        return d
     return None
  
  
  @dataclass
  class MyArguments:
     a  : str                                                    # Positional
     b  : int                = 20                                # My argument
     c  : bool               = False                             # My argument
     d  : int                = choice(0, 1, 2, 3, 4, default=1)  # choices
     e  : List[int]          = argument(default=[0])             # list
     f  : Optional[int]      = None                              # Optional
     p  : Tuple[int, int]    = (1, 1)                            # help p
     g  : Color              = Color.RED                         # help g
     s  : SubArgs            = SubArgs                           # helps group
     cmd: Union[cmd1, cmd2]  = subparsers(cmd1=cmd1, cmd2=cmd2)  # Command subparser
     de : str                = deduceable(workdir)
  
  parser = ArgumentParser()
  parser.add_arguments(MyArguments)
  args = parser.parse_args()

• Save and load arguments from configuration files

  parser = build_parser(commands)
  
  # load/save defaults before parsing
  save_defaults(parser, "config.hjson")
  apply_defaults(parser, "config.hjson")
  
  args = parser.parse_args(["editor", "editor"])
  
  # load save arguments after parsing
  save_as_config(parser, args, "dump.hjson")
  apply_config(parser, args, "dump.hjson")

• Lower level interface, that gives you back all of argparse power

  @dataclass
  class SubArgs:
     aa: str = argument(default="123")
  
  
  @dataclass
  class cmd1:
     args: str = "str1"
  
  
  @dataclass
  class cmd2:
     args: str = "str2"
  
  
  @dataclass
  class MyArguments:
     a: str                  = argument(help="Positional")
     b: int                  = argument(default=20, help="My argument")
     c: bool                 = argument(action="store_true", help="My argument")
     d: int                  = argument(default=1, choices=[0, 1, 2, 3, 4], help="choices")
     e: List[int]            = argument(default=[0], help="list")
     f: Optional[int]        = argument(default=None, help="Optional")
     p: Tuple[int, int]      = argument(default=(1, 1), help="help p")
     g: Color                = argument(default=Color.RED, help="help g")
     s: SubArgs              = group(default=SubArgs, help="helps group")
     cmd: Union[cmd1, cmd2]  = subparsers(cmd1=cmd1, cmd2=cmd2)
  
  
  parser = ArgumentParser()
  parser.add_arguments(MyArguments)
  args = parser.parse_args()

MCP Server

argklass can expose your CLI commands as MCP tools, letting AI agents call them directly.

pip install "argklass[mcp]"

Quick start

The fastest way to run an MCP server is the built-in entry point — just point it at your CLI module:

# stdio (default) — for MCP clients that spawn the process
python -m argklass.mcp mypackage.cli

# SSE — for testing or web-based clients
python -m argklass.mcp mypackage.cli --transport sse

# Streamable HTTP — the newest MCP transport
python -m argklass.mcp mypackage.cli --transport streamable-http

# Custom host/port/name
python -m argklass.mcp mypackage.cli --transport sse --host 0.0.0.0 --port 9000 --name my-tools

Programmatic usage

For more control, use create_mcp_server directly:

import mycommands
from argklass.mcp import create_mcp_server

server = create_mcp_server(mycommands, name="my-tools")

# inspect discovered tools
for tool in server.tools:
    print(tool.name, tool.schema)

# run as a stdio MCP server
server.run()

# or run with SSE
server.run(transport="sse", host="0.0.0.0", port=9000)

The server walks the parser tree, discovers every leaf command, converts its arguments to JSON Schema, and registers them as MCP tools. When a tool is invoked, the arguments are converted back to argv and the command runs normally — stdout, stderr and exit code are returned to the caller.

You can also call tools directly for testing:

output = server.call("editor_cook", {"--dry-run": True})

Configuration (sysconfig)

argklass.sysconfig lets you define configuration as dataclasses, with values resolved from environment variables, a config dict, or defaults (in that priority order).

from dataclasses import dataclass, field
from argklass.sysconfig import ConfigContext

ctx = ConfigContext(prefix="MYAPP")

@dataclass
class DatabaseConfig:
    host: str = ctx.configfield("db.host", str, "localhost")   # MYAPP_DB_HOST
    port: int = ctx.configfield("db.port", int, 5432)          # MYAPP_DB_PORT

@dataclass
class AppConfig:
    debug: bool = ctx.configfield("app.debug", bool, False)    # MYAPP_APP_DEBUG
    db: DatabaseConfig = field(default_factory=DatabaseConfig)

Each field is resolved at instantiation time. Override via environment:

export MYAPP_DB_HOST=db.prod.internal
export MYAPP_DB_PORT=5433

Or programmatically with a config dict:

ctx.set_config({"db": {"host": "db.staging.internal"}})
cfg = AppConfig()   # cfg.db.host == "db.staging.internal"

File I/O supports YAML, JSON and HJSON:

# save / load
ctx.save_config(cfg, "config.yaml")
cfg = ctx.load_config(AppConfig, "config.yaml")

# load and apply as the context's config dict
ctx.load_and_apply("overrides.yaml")

When several libraries use argklass in the same process, each one creates its own ConfigContext with a unique prefix, keeping environment variables and config dicts fully isolated.

Architecture

argklass works by building the argument parser as a tree, adding metadata to each nodes when necessary.

One of the core component is ArgumentParserIterator which traverse the parsing tree. Each features, such as argument grouping into dataclasses or saving/loading configuration, are implemented as a simple traversal.

This enable us to implement each feature independently from each other and make them optional.