Sign InOpen Brain
GitHubGitHub RepoNeeds Review

magnitudedev/magnitude

Magnitude profiles local hardware, recommends compatible models, and configures existing coding-agent harnesses to run them privately and offline.

GitHub
Open Source Open MarkdownOpen JSON
Source Summary

Magnitude is an **Apache 2.0** inference server that profiles a machine, recommends fitting local models with estimated speed, then downloads and tunes the selected model. It supports Codex, Claude Code, Cline, OpenCode, Pi, and other harnesses.

Practical Implication

Use it when local privacy, offline operation, or avoiding API keys and token charges matters. Its CLI can write harness configuration, load models on demand, unload them under memory pressure, and accept compatible GGUF models outside its catalog.

Agent-Ready Context
Magnitude is an **Apache 2.0** inference server that profiles a machine, recommends fitting local models with estimated speed, then downloads and tunes the selected model. It supports Codex, Claude Code, Cline, OpenCode, Pi, and other harnesses.

Use it when local privacy, offline operation, or avoiding API keys and token charges matters. Its CLI can write harness configuration, load models on demand, unload them under memory pressure, and accept compatible GGUF models outside its catalog.

Support is limited to **macOS and Linux**, with Windows available through **WSL**. There is no fixed hardware minimum, so practical model size and throughput still depend on the machine.
Connected Context · Feed7 Judgment

Magnitude turns local-model use from manual serving and guesswork into a machine-aware setup path: profile hardware, recommend a fitting model, tune it, and configure existing coding agents. It strengthens the privacy and offline case while narrowing expectations to hardware-dependent estimates; it supplies no comparative quality evidence and does not remove the need for workload testing.

Context Map
toolscoding#open-models#model-selection#coding-agents
Uncertainty
Support is limited to **macOS and Linux**, with Windows available through **WSL**. There is no fixed hardware minimum, so practical model size and throughput still depend on the machine.