Skip to main content
Local AI

LocalSutra AI

Estimate fit on this machine. Do not invent a benchmark.

Offline local-AI readiness CLI. npm [email protected] published 2026-11-16. The global command alias is localsutra. The README says a fit result is a heuristic, not a promise, and a synthetic CPU loop is not AI inference. On 2026-11-17 the npm downloads API returned no row for last-week or for 2026-10-01 through 2026-11-17, so both counts are 0. GitHub showed 4 stars and 0 open issues. The name LocalSutra has not been trademark-cleared.

  • Offline
  • 25 skills
  • Node 20+
GitHub version
0.2.1
Licence
Apache-2.0
Language
JavaScript
Weekly downloads
295
Overall downloads
295
GitHub stars
4

Live npm weekly downloads, npm overall downloads, and the latest npm version, plus GitHub stars and the package.json version, fetched 2026-11-17 19:07 UTC. Cached 10 minutes. Weekly downloads cover 2026-11-10 to 2026-11-16. Overall downloads cover 2026-10-01 to 2026-11-17. npm does not publish an earlier total. Stars and downloads show public use. They are not a quality score.

Use it when

  • 1

    You want OS, CPU, RAM, free disk, and an NVIDIA GPU only when nvidia-smi exists.

  • 2

    You want a memory estimate that prints its assumptions.

  • 3

    You want loopback checks of Ollama or llama.cpp without downloading a model.

What it does, and what it does not

Read this before the guide. A limit here is a limit of the tool.

Does

  • doctor reports OS, architecture, a WSL flag, CPU, RAM total and free, free disk, tools on PATH, and an NVIDIA GPU name, VRAM, and driver when nvidia-smi exists. AMD, Intel, and Apple GPUs and NPUs stay unknown, not absent.
  • fit estimates weights, KV cache, overhead, RAM fit, and single-GPU VRAM fit. The README formula is weights times 1.35 plus overhead plus KV cache. The 1.35 and the overhead are assumptions printed in the report. Speed and quality stay unknown.
  • probe checks a loopback Ollama or llama.cpp server and lists models. Inference is not tested. Endpoints must be http://127.0.0.1:PORT/ or http://[::1]:PORT/. localhost, public hosts, credentials, paths, queries, and redirects are blocked.
  • infer is opt-in timing of a model already installed. The README says prompt and output text are not stored, only a hash. tokens/s comes from the runtime counter.
  • benchmark is a synthetic CPU loop with median, p95, and standard deviation. The README says it is not AI inference. compare refuses two reports whose workload or platform differ.
  • audit runs 37 adversarial checks. Verdicts are APPROVED_WITH_EVIDENCE, REVIEW_REQUIRED, or BLOCKED_UNSAFE. Exit behavior follows the report command. verify checks a SHA-256 seal and does not say who made the file.
  • install-skills copies 25 skills (a Prime orchestrator, 19 specialists, and 5 capability skills) and supports --dry-run. The README says native IDE discovery is not verified.

Does not

  • It does not download models, install Ollama, change drivers, or open ports.
  • A fit label such as possibly_fits is not a promise the model will load or answer well. The sample catalog in the README is synthetic.
  • privacy checks the endpoint rule and a data inventory. The README says it does not prove an external runtime has no egress.
  • It will not rank models across unlike hardware, context lengths, or quantizations.
  • The README says the recorded proof run had 122 tests with 0 failed and 0 skipped, and 9,500 seeded hostile inputs with 0 unexpected crashes. This page did not re-run that proof.
  • No telemetry. The only network calls are loopback requests you ask for. serve binds 127.0.0.1 and needs a token of 24 characters or more. SHA-256 detects change. It is not a signature. Trademark clearance for the name has not been performed.

Claims above are from the project pages and are attributed as stated there. They are not independently verified here.

How a run moves

From the LocalSutra AI README, npm 0.2.0, published 2026-11-16. It profiles this computer and estimates memory. It does not download a model.

  1. 1. Doctor

    What was measured

    OS, CPU, RAM, free disk, and NVIDIA only when nvidia-smi exists. Other GPUs stay unknown.

  2. 2. Fit

    A printed formula

    weights times 1.35, plus overhead, plus KV cache. The README says speed and quality are unknown.

  3. 3. Probe

    Loopback only

    127.0.0.1 or ::1. A listed model is not a timed answer. infer is a separate opt-in.

  4. 4. Bench

    CPU loop

    median, p95, and standard deviation of a synthetic loop. The README says that is not AI inference.

How to read the result

Result It means It does not mean
possibly_fits The estimate says free RAM is above the required figure. Not a load test and not a quality score.
unknown GPU nvidia-smi was not used to see that device. Not proof the GPU is absent.
APPROVED_WITH_EVIDENCE The 37-check audit passed its own rules. Not third-party review. The README lists live inference and non-NVIDIA GPUs as not tested.

Wording is from the project page or README and is attributed as stated there. It is not independently verified here.

Commands and notes

  • npx [email protected] doctor hardware profile on this machine.
  • npx [email protected] audit --out ./local-audit 37 adversarial checks.
  • localsutra fit --catalog fixtures/model-catalog.json memory estimate. The README example models are synthetic.
  • localsutra probe --runtime ollama loopback health. Inference is not tested.
  • localsutra infer, benchmark, compare, privacy, verify, install-skills, and serve listed in the README command table.

Guide

How to use the result. GitHub is the source repository. npm is the install.

What 0.2 measures

From the LocalSutra AI README, npm 0.2.0, published 2026-11-16. Node.js 20 or newer. 0 runtime dependencies. The command alias is localsutra.

  • npx [email protected] doctor reports OS, CPU, RAM, free disk, and an NVIDIA GPU when nvidia-smi exists. Other GPUs stay unknown.
  • fit uses weights times 1.35 plus overhead plus KV cache. The README says those factors are assumptions and that speed and quality stay unknown.
  • probe talks only to 127.0.0.1 or ::1. infer is opt-in and stores a hash, not the prompt. benchmark is a synthetic CPU loop, not model inference.
  • audit runs 37 adversarial checks. Verdicts are APPROVED_WITH_EVIDENCE, REVIEW_REQUIRED, or BLOCKED_UNSAFE.

What an estimate is not

  • possibly_fits is not a load test. The sample catalog in the README is synthetic. The CLI does not download models or install Ollama.
  • privacy does not prove an external runtime has no egress. compare refuses unlike hardware or workloads. It is not a leaderboard.
  • The README says a recorded proof run had 122 tests, 0 failed, and 9,500 hostile inputs with 0 unexpected crashes. This site did not re-run that proof.
  • SHA-256 detects a changed file and is not a signature. 25 skills copy into documented folders. IDE discovery is not verified. The name LocalSutra has not been trademark-cleared. On 2026-11-17 npm had no download row, so the snapshot count is 0.

Questions

Does LocalSutra tell me a 7B model will run?

It gives an approximate memory answer and prints the assumptions. The README says a real answer is to load the model in Ollama or llama.cpp, then use infer. The CLI does not download the model.

Is the CPU benchmark an LLM speed?

No. The README says benchmark is a synthetic CPU loop and is explicitly not AI inference. compare refuses reports whose workload or platform do not match. It is not a leaderboard.

WhatsApp