CLI Packaging Patterns Agent Skill

CLI Packaging & Distribution

The Pattern

Problem: You want a CLI tool that installs cleanly from a git URL with zero manual steps — no cloning, no virtual env, no PATH fiddling.

Approach: Combine pyproject.toml [project.scripts] with hatchling build backend, a __main__.py dual entry point, and argparse with a default-action subcommand design so mytool and mytool serve do the same thing.

Pattern proven in production across multiple Python CLI tools and web services.

When this skill is NOT the right answer

This skill assumes a pure Python CLI distributed via uv tool install. If your project does not fit that profile, use a different pattern:

| Project shape | Use instead |

|--------------|-------------|

| Multi-language stack (Python + Node + Docker) | one-line-installer-patterns |

| Raw TS/React app with no Python wrapper | one-line-installer-patterns (or publish to npm) |

| Tool that bootstraps system prerequisites | one-line-installer-patterns |

| Containerized multi-service app | Ship docker-compose.yml; see container-orchestration-patterns |

| Single static binary (Go/Rust) | GitHub releases + curl -L .../bin -o ~/.local/bin/tool |

If the project IS a pure Python CLI, the rest of this skill applies.

Key Design Decisions

1. Hatchling + [project.scripts] — Not setuptools

Use hatchling as the build backend. The entry point declaration is:

[project.scripts]
my-tool = "my_tool.cli:main"

Why hatchling: simpler than setuptools, no setup.py, no MANIFEST.in. The [tool.hatch.build.targets.wheel] section lets you exclude test files from the published wheel:

[tool.hatch.build.targets.wheel]
packages = ["my_tool"]
exclude = ["my_tool/tests", "my_tool/frontend/tests"]

2. __main__.py dual entry — python -m always works

Include a minimal __main__.py so the tool works even when the script entry point isn't on PATH:

# my_tool/__main__.py
"""Allow running as: python -m my_tool"""
from my_tool.cli import main
main()

This matters because uv tool install creates a wrapper script, but during development or in edge cases, python -m my_tool is a reliable fallback. Service management code should use this as a fallback too:

def _resolve_tool_bin() -> str:
    which = shutil.which("my-tool")
    if which:
        return which
    return f"{sys.executable} -m my_tool"

3. Default action + subcommands — tool equals tool serve

Use argparse with shared flags on the root parser AND on the serve subcommand, so the bare command runs the server:

def main() -> None:
    parser = argparse.ArgumentParser(prog="my-tool", ...)
    _add_serve_flags(parser)              # flags on root parser
    sub = parser.add_subparsers(dest="command")
    serve_parser = sub.add_parser("serve", help="Start the server (default)")
    _add_serve_flags(serve_parser)        # same flags on 'serve' subcommand

The dispatch at the bottom falls through to serve() when no subcommand is given:

    else:
        serve(host=args.host, port=args.port, ...)

4. 3-tier config resolution: CLI > file > default

The serve() function resolves every setting with the same pattern — CLI flag wins, then settings file, then hardcoded default:

settings = load_settings()
host = host if host is not None else settings.get("host", "127.0.0.1")
port = port if port is not None else settings.get("port", 8088)
log_level = log_level if log_level is not None else settings.get("log_level", "info")

Using None as the argparse default (not a value like "127.0.0.1") is critical — it distinguishes "user didn't pass a flag" from "user explicitly set it."

5. uv tool install git+https://... compatibility

No special config needed — hatchling + [project.scripts] is all uv requires. The install command is:

uv tool install git+https://github.com/yourorg/your-tool

For tools with plugins as extras, the --with flag adds plugin packages:

uv tool install git+https://github.com/yourorg/your-tool \
  --with 'your-plugin @ git+https://github.com/yourorg/your-plugin@main'

This can be handled programmatically in a reinstall helper:

cmd = [uv_path, "tool", "install",
       "git+https://github.com/yourorg/your-tool", "--force"]
for spec in plugin_specs:
    cmd.extend(["--with", spec])

6. Post-install configuration: init, login, or nothing?

The Amplifier ecosystem convention for first-run configuration is <tool> init. This is internally consistent but not the dominant community convention. Choose deliberately:

| Tool needs | Recommended subcommand | Examples |

|------------|------------------------|----------|

| Auth/credentials only | <tool> login or <tool> auth login | gh, vercel, heroku, fly, supabase |

| Tool-wide config (region, defaults) | <tool> configure | aws |

| Per-project setup (creates files in cwd) | <tool> init | terraform, firebase, npm, cargo |

| Sensible defaults; prompt lazily | (no subcommand) | bun, deno, rustup, pnpm |

In the broader community, init overwhelmingly means "create a new project/workspace in the current directory." Using init for "configure the tool itself" has essentially one major precedent: gcloud init. If you're building a tool the broader community will consume, prefer login / configure / no-command unless you actually mean "scaffold a new project here." If your audience is internal Amplifier-only, the init convention is fine — just know what you're choosing.

See one-line-installer-patterns for the full survey table and the rationale.

Template / Starter Code

# pyproject.toml
[build-system]
requires = ["hatchling"]
build-backend = "hatchling.build"

[project]
name = "my-tool"
version = "0.1.0"
requires-python = ">=3.11"
dependencies = ["fastapi>=0.115.0", "uvicorn[standard]>=0.30.0"]

[project.scripts]
my-tool = "my_tool.cli:main"

[tool.hatch.build.targets.wheel]
packages = ["my_tool"]
exclude = ["my_tool/tests"]

# my_tool/__main__.py
from my_tool.cli import main
main()

# my_tool/cli.py
import argparse, sys

def _add_serve_flags(parser):
    parser.add_argument("--host", default=None)
    parser.add_argument("--port", type=int, default=None)

def serve(host=None, port=None):
    from my_tool.settings import load_settings
    settings = load_settings()
    host = host if host is not None else settings.get("host", "127.0.0.1")
    port = port if port is not None else settings.get("port", 8080)
    import uvicorn
    from my_tool.app import app
    uvicorn.run(app, host=host, port=port)

def doctor():
    """Run diagnostic checks."""
    print("\033[32m✓\033[0m Python", sys.version.split()[0])
    print("\033[32m✓\033[0m my-tool", __version__)
    # Add your checks here

def main():
    parser = argparse.ArgumentParser(prog="my-tool")
    _add_serve_flags(parser)
    sub = parser.add_subparsers(dest="command")
    serve_p = sub.add_parser("serve")
    _add_serve_flags(serve_p)
    sub.add_parser("doctor")

    args = parser.parse_args()
    if args.command == "doctor":
        doctor()
    else:
        serve(host=args.host, port=args.port)

Gotchas & Lessons Learned

• argparse default=None is load-bearing. If you set default="127.0.0.1" on the --host flag, the 3-tier resolution breaks — you can never tell if the user explicitly passed --host 127.0.0.1 or if argparse filled it in.

• exclude in hatch config is about the wheel, not the sdist. Test files still appear in the source distribution. This is fine — you don't want tests in the installed package, but they should be in the sdist for downstream repackagers.

• uv tool install builds a wheel in an isolated environment. If your package has undeclared dependencies (imports something not in [project.dependencies]), it will fail at install time, not at import time. Explicitly declare transitive deps that may be missing on clean environments.

• Service files need the full PATH. When systemd or launchd runs your tool, PATH is minimal. Capture os.environ.get("PATH") at install time and bake it into the service unit. Without this, subprocesses can't find docker, git, tmux, etc.

• The --force flag on reinstall matters. uv tool install without --force is a no-op if the package is already installed. Upgrade commands must use --force to ensure the latest git HEAD is fetched.

Related skills

• one-line-installer-patterns — For projects that can't use uv tool install: multi-language stacks, raw TS/React apps, tools that need system bootstrapping, or non-technical audiences. Also contains the full community convention survey for post-install commands referenced in §6 above.
• config-state-patterns — Where to store the config and state created by your tool's init / configure / login flow.
• http-service-patterns — If your tool is an HTTP service (FastAPI lifecycle, SPA + API, WebSockets, SSE).