Executive Summary

LocalSutra AI 0.2.1 is an open-source (Apache-2.0), offline-first command-line tool that tells people whether their own computer can run a large language model, whether their local Ollama or llama.cpp server really is local, and how fast an installed model runs. Its promise: Know your hardware. Choose your model. Keep AI local.

The problem

People who want to run AI locally for privacy, cost or learning face four guesses: will this model fit, what do I need to install, is my setup really private, and how fast will it be. Existing answers are scattered forum posts and calculators that present estimates as facts.

The product

Why trust it

EvidenceValue (recorded run)
Automated tests126 run, 126 passed, 0 failed, 0 skipped
Hostile fuzz inputs9500 cases, 0 unexpected crashes
Runtime dependencies0
Network useOnly requests you ask for, to 127.0.0.1
Recorded onwin32/x64, Node v24.15.0, 2026-11-17T17:19:18.302Z

Every number is regenerated by npm run proof and linked to raw logs. See Assurance and Evidence.

What it deliberately does not do

It does not install runtimes, download models, rank models across unlike hardware, claim a setup is "private" or "air-gapped", or run live inference unless you ask. Live inference on a real model was not run for this release; the report says so.

Risks and limits

Memory-fit numbers are heuristics. GPU detection covers NVIDIA only. IDE skill activation is not verified. No third-party security review has happened yet.

Roadmap in one line

Real-runtime integration tests, more GPU vendors, consent-based install planning, signed audit capsules, external review. See Roadmap and FAQ.

Created by Vatsal Shah. Apache-2.0.