productivity|
ai

Introducing: magic-cli

Guy Waldman

July 16, 2024 (4 months ago)

Time to read:

4 minutes

#Problem statement

The command line is a superpower every developer should harness.
For me, have always been the "source of truth", even when alternative GUI applications exist for various tasks.

My main issues with command line interfaces are:

Non-intuitive: CLI tools are often not intuitive to use, often due to historical reasons (prevalent in many )
Context switches: I switch around often between operating systems and even shells in the same device, and so remembering the right commands in each system is a challenge. It's something that aliases and functions I share around between devices with dotfiles helps with, but the mental overhead of remembering the right command for each system is not easy
Arguments: Many CLI tools have a lot of arguments, in which sometimes even the order of the arguments matters (case in point: ffmpeg which is notoriously difficult to use)

This is a problem that , despite some of their limitations, can help solve.

Fig terminal (now part of Amazon Q) and GitHub Copilot for CLI have been helpful in accelerating my workflow on the CLI, despite not being perfect (inherently so due to the nature of LLMs), and I wanted to take this idea further.

The main things I wanted from these types of tools are:

Cross-platform: Should run on most platforms and shells
Execution: Should be able to execute commands it suggests in the context of my current shell session (with a confirmation before doing so, to avoid accidentally running commands I didn't mean to)
Local inference: Should support text completions with a local LLM such as ollama or llama.cpp
Great UX: Should be easy to use and have a great UX

Introducing magic-cli which aims to solve these gripes.

#What's Magic CLI?

magic-cli is a command line utility to make you a magician in the terminal.
It started as a hobby project I hacked on during a vacation I took a little while ago, and it's something I think could be useful to others.

It is built in Rust (see the Why Rust? section below) and supports local and remote LLM providers (at the time of this writing, ollama and OpenAI).

#How does it work?

The normal suggest command accepts a prompt in natural language and prompts the LLM (local or remote, depending on the user configuration) to suggest a command to execute.
The user can then revise the command until they are content with the answer, and then either copy it to the clipboard or execute it (depending on the user configuration).

The more advanced ask command (experimental) accepts a "task" in natural language and the LLM will attempt to answer it, asking the user to run various commands if more context is required. The loop continues until the user is satisfied with the answer.

In addition, there is a search command which accepts an expression that should be similar to a previous command in your shell history, and performs a semantic search across your history to find the closest match.

#Why Rust?

A lot of projects use "powered by Rust" as a PR move that emphasizes strong performance and memory safety guarantees, but truth be told, I just really like Rust, especially for building CLI tools.

There are major drawbacks for choosing Rust for this, sure:

Higher entry barrier for contributions, as it's a less popular language with a steep learning curve
Rust's complexity compared to higher-level languages is usually positive ROI for software with significant performance and security requirements, but this may be considered somewhat overkill
Iteration and development time can be slower, especially for newer Rustaceans

An obvious choice here would be Python, which is a popular choice for working with LLMs, and has a great ecosystem for that.
Especially since most tasks are IO-bound rather than CPU-bound (the code does not deal with the inference, but rather delegate to the configured LLM), and so the choice for a lower-level performant language may have diminishing returns.

However, Rust's solid ecosystem and language robustness empower me much more than other languages do.
"If it compiles, it works" generally proves correct in my workflows with Rust.

, but I strongly recommend to watch this talk by Jon Gjengset:

It's a bit old by now, but still relevant.

I have also developed a project for LLM orchestration in Rust, which hasn't really seen much love recently: guywaldman/orch

#Try it now!

You can install magic-cli in a variety of ways (even with homebrew). Check out the installation instructions in the repo: guywaldman/magic-cli

Please try it try out and let me know what you think!