Magnitude

Magnitude is an inference server for people who want their coding agent to stop calling a cloud API. It profiles the machine it is installed on, recommends models that will actually fit the available memory, downloads them, and rewrites the agent's configuration to point at a local endpoint — turning what is normally an afternoon of weight-hunting and config editing into a single setup command.

The pitch is cost and privacy rather than raw capability. There is no per-token bill and no rate limit, prompts and source code never leave the device, and once the weights are on disk the whole thing keeps working offline. Magnitude loads and unloads models on demand, so a machine with limited RAM can serve several without thrashing.

It is deliberately agent-agnostic: the documented integrations include Claude Code, Cline, OpenCode, Codex, Hermes and Pi, and any other tool that accepts a custom base URL can be pointed at the same local endpoint.

Installation is a global npm package — npm i -g @magnitudedev/cli, then magnitude setup — and it runs on macOS, Linux, and Windows through WSL. The project is open source under Apache 2.0, with documentation at docs.magnitude.dev.