Elzi documentation

Elzi feature guides

Everything the app can do, one short guide per feature.

Elzi is LeanZero's Mac app for AI agents. These guides explain each part of it in a few minutes: the three kinds of work you can run, how nodes and providers decide which model does the job, the MLX engine that serves models on your own Macs, and the speeds we measured. Every number names its model, its Mac and the date we measured it.

Using Elzi

What Elzi isA Mac app where AI agents do real work, on local models, cloud models or both.Agent sessionsChat with an agent inside a project folder, or set one up to work on a schedule.Software build sessionsOne brief becomes a planned, multi-agent build across your nodes.Benchmark runsGive models the same frozen task and compare their scores before you trust one.Nodes and strategiesThe models Elzi hands work to, and the rules for which one does what.ProvidersWhere models come from: the MLX engine on your Macs, or a cloud provider with your key.

The MLX engine

The LeanZero MLX engineThe local model server inside Elzi, built on our fork of Rapid-MLX.Serving a model on one MacElzi checks the memory, mounts the model and keeps long chats inside what fits.Distributed serving across two MacsSplit one model across two Macs over Thunderbolt, to run models too big for one.N-gram tables and prompt lookupDrafts copied from your own context make code edits much faster, with the same output.Speed: measured numbersDecode, prefill and time to first token from our own runs, with model, Mac and date.Memory lockKeep memory for your local model while other apps compete for it.

Media and devices

ImagesElzi makes pictures on your Mac with open image models and keeps every one.VoicesVoice takes for scripted lines, kept in a library. Off by default and limited for now.LeanZero LinkUse your other Macs: run a model there, copy models over, or split one across two.