~4,700 stars · Apache-2.0, read from /blob/main/LICENSE on 2026-09-28; standard unmodified text, copyright "2026 Magnitude AI Inc.", and the sidebar agrees · @magnitudedev/cli@0.1.5 (2026-09-23), read from /releases/latest and confirmed by ungh; npm's latest dist-tag still serves the DEPRECATED 0.0.15 · Track this in Scout
It measures the machine it is installed on, recommends a local AI model that will run well there, then downloads, tunes and serves it.
▶Repo detailsthe review · specs · pros & cons · install
What it is
A local inference server plus a desktop application. The unusual part is the profiling step: rather than asking which model you want, it measures the hardware first and recommends from there. It runs on Apple Silicon, NVIDIA, AMD or a plain processor, and other programs talk to it over a local connection.
What it is good for. Anyone who has tried to run an AI model locally and given up because the choice of model, size and settings is a maze. It also suits anyone whose per-request bill at a cloud provider has grown uncomfortable, and anyone who cannot send text to a third party at all.
- No API key (a password that lets one program use another over the internet), no rate limit and no per-request cost.
- The hardware-profiling step is genuinely useful and nothing else in this category does it.
- Apache-2.0, read from the LICENSE file and completely plain, with no branding clause and no paid tier.
- The same repository has been three different products, and the install command in its README currently fetches the wrong one. The npm package
magnitude-testdescribes a browser-testing client. The package@magnitudedev/cliat version 0.0.15 describes an AI coding agent and is marked deprecated, with a notice saying the project has moved to a desktop application. The current README describes neither. Anyone running the README'snpm i -g @magnitudedev/clitoday installs the deprecated 0.0.15, not the 0.1.5 that GitHub lists as newest. - It is very new. The whole 0.1.x line began on 17 September 2026, eleven days before this was written, and the first release that day was an alpha.
- No hardware requirements are published at all. The project says plainly that there is no fixed minimum because it profiles the machine. That is a reasonable design choice, but it means there is no number to check before starting, and the models themselves are multi-gigabyte downloads with their own separate licences.
- ollama/ollama
181,752 stars, code on 25 September 2026; overwhelmingly the standard way to run local models, and it does not profile the machine or recommend a model, so you choose the size yourself.
Track this in Scout - janhq/jan
44,308 stars, code on 3 September 2026; a fully offline desktop application that leads with a chat window rather than a server for other programs to use.
Track this in Scout 
mozilla-ai/llamafile25,611 stars, code on 16 August 2026, covered in Edition 26; it ships the model and the runtime as one downloadable file, with no catalogue and no tuning step.
Track this in Scout
npm i -g @magnitudedev/cli magnitude setup # then confirm you did not get 0.0.15
