Skip to main content
Edge AI

EdgeSutra AI

Separate host measurements from claims about a device.

Offline edge-AI compatibility and benchmark-audit CLI. npm [email protected] published 2026-11-15. The README says it does not run models on edge hardware. Host timing is not device evidence.

  • Offline
  • 20 skills
  • Node 20+
GitHub version
0.2.1
Licence
Apache-2.0
Language
JavaScript
Weekly downloads
337
Overall downloads
337
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 have an ONNX file or declared model, runtime, and device manifests.

  • 2

    You want host CPU timing labelled as host-only, not as device latency.

  • 3

    You do not want energy or carbon filled in from CPU time.

What it does, and what it does not

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

Does

  • Reads ONNX operators and opsets from bytes, up to 32 MiB, without executing the model.
  • check compares declared model, runtime, and device manifests. plan writes a conversion plan and does not run it.
  • benchmark and compare time a toy FP32 run against a simulated INT8 run on this host. The README labels that MEASURED_HOST_ONLY.
  • profile reports host OS, CPU, and RAM. GPU, NPU, and storage are not probed.
  • Writes audit-report.html, audit-report.md, audit-report.json, and SHA256SUMS.txt. Exit 0 is ok. Exit 1 is blocked or incompatible. Exit 2 is invalid input or a tool error.
  • 20 skills copied with install --ide for Claude Code, Cursor, Codex, Gemini CLI, Copilot, OpenCode, and Antigravity. The README says the files are instructions, not autonomous services.

Does not

  • It is not an inference framework. ONNX Runtime, LiteRT, ExecuTorch, and llama.cpp inference are not implemented.
  • Physical or virtual edge-device execution is not implemented and is reported as NOT_RUN.
  • Host timing is not device latency. Energy, carbon, and peak memory stay NOT_MEASURED.
  • GO_CANDIDATE is not emitted without verified on-device evidence and human sign-off. The default verdict is BLOCKED or REVIEW_REQUIRED.
  • SHA-256 detects a changed file. The README says it is not a signature and proves nothing about origin.
  • Native skill discovery in each IDE is not independently verified. Trademark clearance for the name EdgeSutra has not been performed. 0 runtime dependencies. No telemetry and no network calls.

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 EdgeSutra AI README, npm 0.2.0, published 2026-11-15. It audits files and this host. It does not run a model on edge hardware.

  1. 1. Input

    Bytes or manifests

    ONNX up to 32 MiB, or declared model, runtime, and device JSON. The model is not executed for metadata.

  2. 2. Check

    Static agreement

    check looks for agreement between those declarations. plan writes a conversion plan and does not run it.

  3. 3. Measure

    This host only

    A toy FP32 versus simulated INT8 timing is MEASURED_HOST_ONLY. Energy, carbon, and peak memory stay NOT_MEASURED.

  4. 4. Gate

    Default deny

    BLOCKED or REVIEW_REQUIRED. GO_CANDIDATE needs verified on-device evidence and a person. That evidence path is not implemented.

How to read the result

Result It means It does not mean
MEASURED_HOST_ONLY A timed run on this machine CPU. Not latency on a phone, NPU, or other device.
NOT_RUN On-device execution was not performed. Not a pass and not a faked device number.
REVIEW_REQUIRED The report is ready for a person. Not a ship approval. GO_CANDIDATE is withheld without on-device evidence.

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] audit --out ./edge-audit report bundle on this machine.
  • edgesutra-ai audit --onnx path/to/model.onnx metadata from the file, no execution.
  • edgesutra-ai inspect, check, plan, benchmark, compare, profile, verify, and render listed in the README command table.
  • edgesutra-ai install --ide claude --project /path/to/project copies 20 skills and does not overwrite.
  • edgesutra-ai serve read-only report server on 127.0.0.1 with a bearer token.

Guide

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

What 0.2 measures

From the EdgeSutra AI README, npm 0.2.0, published 2026-11-15. Node.js 20 or newer. 0 runtime dependencies.

  • npx [email protected] audit --out ./edge-audit writes audit-report.html, audit-report.md, audit-report.json, and SHA256SUMS.txt.
  • inspect reads ONNX operators and opsets from bytes, up to 32 MiB, and does not execute the model.
  • check compares declared model, runtime, and device manifests. plan writes a conversion plan and does not run it.
  • benchmark and compare time a toy FP32 run against a simulated INT8 run on this host. The README labels that MEASURED_HOST_ONLY.
  • profile reports host OS, CPU, and RAM. GPU, NPU, and storage are not probed. Exit 0 is ok. Exit 1 is blocked or incompatible. Exit 2 is invalid input or a tool error.

What a host number is not

  • The README says EdgeSutra is not an inference framework and does not run models on edge hardware. ONNX Runtime, LiteRT, ExecuTorch, and llama.cpp inference are not implemented.
  • Device execution is NOT_RUN. Host timing is not device latency. Energy, carbon, and peak memory stay NOT_MEASURED.
  • The default verdict is BLOCKED or REVIEW_REQUIRED. GO_CANDIDATE needs verified on-device evidence and a person. That path is not implemented.
  • SHA-256 detects a changed file. The README says it is not a signature. 20 skills are copied into documented folders. IDE discovery is not independently verified. No telemetry and no network calls.

Questions

Does EdgeSutra run my model on a device?

No. The README says it is not an inference framework and does not run models on edge hardware. ONNX Runtime, LiteRT, ExecuTorch, and llama.cpp inference are not implemented. Device execution is NOT_RUN.

Does a host benchmark prove the device is faster?

No. The README labels host CPU timing MEASURED_HOST_ONLY and says energy, carbon, and peak memory stay NOT_MEASURED. GO_CANDIDATE is not emitted without verified on-device evidence and human sign-off.

WhatsApp