oriaj-nocrala/llama-crab-projects

llama-crab-projects is a collection of command line tools written in Rust that run AI language models locally on your own computer, using a GPU with CUDA support. It relies on a crate called llama-crab to load and run GGUF format models. The author describes it as a personal experimentation project rather than a polished library meant for other people to build on. The most developed tool is called ask. You can pipe text into it and ask a question about that text, or just ask a standalone question with nothing piped in. It has a simple pipe mode that gives one answer and exits, useful for chaining with other commands, and an interactive mode (started with the -i flag) that supports a full back and forth conversation with live streaming replies and markdown formatting rendered directly in the terminal. Ask also comes with two real tools it can call on: a calculator with its own parser and a currency converter that checks a public exchange rate API, so it does not have to guess at math or currency values from memory. The default model can be set once in a config file, or overridden per run with a flag. A second tool called debate sets up two AI personalities arguing about a topic, each with its own personality description, temperature setting, and terminal color. If you do not specify the personas or topic, the tool invents them itself in a chaotic default mode. Each line of dialogue is generated in multiple candidate versions and a separate judging model picks the best one, and by default the conversation can branch into parallel storylines partway through. A more unusual feature checks how closely each line still relates to the actual topic using text embeddings, and if the conversation drifts too far off topic for two turns in a row, a moderator character steps in to bring it back on track. A third tool, fim, is an interactive code autocomplete that runs in the terminal like a small text editor. As you type and pause, a gray suggested completion appears after your cursor, which you can accept with Tab. There is also a similar predictive keyboard tool for general text. Building requires an optimized release build and a CUDA capable GPU with local model files, since the tool is too slow to be usable without those optimizations. No license is stated in the project.