specialized capacity, on demand, composed by you, paid according to use.
Delivered through toast, composable AI for the command line.
There is no framework or harness to learn. Unix solved how people and programs work together fifty years ago, and it works for AI too.
Slices: toast reviewer, toast editor, toast kimi. A name for a job or a mind, with the model behind it chosen and tested by us and replaced when a better one appears.
One call, a pipe, a loop, or a cron line. Nothing runs until something asks, and you decide how much the machine does before you read the result.
You write the pipeline in the shell you already use. A file lists what the model may run; every command is printed before it runs; every change can be undone.
Local models and your own keys cost nothing. Our AIs and every major vendor are metered by the token, on one balance. macli is $29 once; a shell on our Mac, or your Mac at our place, by the month.
toast finds a local model before it looks for the cloud: Apple's on-device model (appled), toasted on MLX, or a DwarfStar server running DeepSeek V4 Flash or GLM 5.3 Flash. Start ds4-server and toast uses it; tools, files and images work the same way, and none of it needs credits. Local is that, packaged: built, downloaded, kept current, jam and ito included. Slices always run on our side, so a pipeline can mix the two: grep ERROR *.log | sort | uniq -c | toast "group by fault" | toast paranoid does the reading locally and hands the slice only the lines that need a specialist. The logs never leave the machine; credits go to the judgement.
toast is free, with Apple's on-device model to show you how it works. Add credits and our slices answer. Local puts DeepSeek V4 Flash or GLM 5.3 Flash on your Mac, all six steps, no credits, ever. macli installs 28 command line tools for Mac, easy to use by people and AI. Or a shell on our Mac, your Mac at our place, or all of it on your premises.
Credits. Models through linuxtoaster — our AIs, the slices (toast reviewer, toast editor, toast kimi), and every major vendor: Anthropic, OpenAI, Google, xAI, Moonshot, DeepSeek, Alibaba, Zhipu, Mistral — are paid for with credits, from $20, metered by the token on one balance. Hosted and Colocation include credits every month. Local models and your own keys never need any. All the slices →
At every step, three things hold: a file in your directory lists the commands the model may run; every command is printed before it runs; and every change is recorded with a reason and can be undone.
Chat makes you work. Agents do too much. With toast you select the level of delegation required by the task, with one or more specialized models working for you, alone or together.
$ log stream | head -200 | toast "narrate like a nature documentary" | say Here, in the quiet of a Saturday afternoon, the launch daemon stirs. It has noticed the Wi-Fi. It does not approve. $ arp -an | toast "introduce each device like you are in a bar" .1 is the router. Runs the place. Doesn't drink. .4 is "Toaster". Small, but persistent. And .2 and .6? Strangers.
The tutorial, with every command → Steps 4 to 6 use jam and ito, which come with Local, a hosted shell or a colocated Mac — or build from source.
What is Unix? Press ` to open a Unix terminal. To close, press ` again. Type help to start.
