Everything the desktop app does, the terminal does too. Generation runs as a local process you can script, schedule and pipe — no API key, no per-call billing, no rate limit to design around.
What it enables
Point a loop at a folder and leave it. One command generates one asset, exits with a status code, and writes where you told it to — so a thousand lines of text or a directory of recordings is a shell loop, not a queue you have to babysit or pay per item.
The command line is a thin client over a local HTTP API, so the same run fits into a Makefile, a build step, a cron job, or whatever language your pipeline is already written in. Nothing leaves the machine, so there is no upload step to wait on and no credentials to rotate.
An MCP server exposes your installed models as callable tools, so an AI assistant can transcribe, translate or generate speech through the same models you use by hand — locally, with the files staying on disk.
A model is a package: a manifest that declares its inputs and controls, its dependencies, and the weights to fetch. Studio renders the form from that declaration, so a model you package yourself arrives with an interface you did not have to write. The format is documented publicly.
The command line
The hutash command manages models and runs generation. It is an HTTP client for the engine on localhost:47990 — the same relationship as docker to dockerd. Output is JSON by default when piped, which is what makes it scriptable.
Full CLI reference — every command, flag, exit code and environment variable.
Exit codes: 0 success, 1 error, 2 not found, 3 engine unreachable — which is what a batch script branches on.
Automation surfaces
The engine is the only thing that talks to models. Everything else — the desktop app, the terminal, an assistant — is a client of the same local API.
Packaging
A pipeline package declares what it accepts and what it returns; the engine builds its environment, fetches its weights, assigns it a port and serves it. The walkthrough builds one by hand and ends with a running model answering an inference request.
