15:I[981,["658","static/chunks/658-a698693f56d0116a.js","183","static/chunks/183-6beeeb801c274aaa.js","474","static/chunks/app/%5Btype%5D/layout-e8e83705b5c8d3b7.js"],"Posts"] 16:T6f5,Most CLI frameworks (Click, argparse) make you manually register subcommands with decorators or parser objects. Xclif tries something different: drop a file in your routes/ folder and it's automatically a subcommand.

myapp/ ├── routes/ │ ├── greet.py → myapp greet │ └── server/ │ ├── start.py → myapp server start │ └── stop.py → myapp server stop

Here's an example of what defining a command looks like in one of these files. Note that it's very similar to Click.

@command() def _(name: str, template: str = "Hello, {}!") -> None: """Greet someone.""" print(template.format(name))

The custom parser is ~25x faster than Click and ~170x faster than Typer for dispatch (2.7 µs vs 72 µs vs 496 µs for a simple command). Also includes Rich-formatted help, WithConfig[T] for config/env/CLI flag layering, agent-optimized help output (auto-detects non-TTY → compact format for LLM tool calls), and an optional manifest compiler to skip route-walking in large codebases.

This started as an exploration of whether file-based routing (inspired by Next.js) could work in the CLI world. It works surprisingly well, especially for projects with lots of subcommands where finding the right add_subcommand call becomes a chore.

I've been using this for many of my other projects like https://codeberg.org/thatxliner/codeberg-cli and https://github.com/ThatXliner/git-worm17:T4fa,Here once again A Token Usage Meter for 12+ AI Providers Anthropic, OpenAI,

AI Agents That Builds Themselves