User guide
1. Requirements
Node.js 22 or newer. Nothing else: no API key, no network, no GPU.
2. Your first audit (2 minutes)
npx edgesutra-ai audit --out ./edge-audit
Open edge-audit/audit-report.html. The sample inputs are bundled synthetic fixtures; the report says so (ES-SYNTHETIC-INPUT).
3. Audit your own model
Create three small JSON manifests (see Technical reference), then:
npx edgesutra-ai audit --out ./edge-audit --model model.json --device device.json --runtime runtime.json
Or inspect a real ONNX file and audit its operators:
npx edgesutra-ai audit --out ./edge-audit --onnx ./model.onnx
4. Read the report
| Section | How to use it |
|---|---|
| Verdict | BLOCKED = stop; REVIEW_REQUIRED = continue, but a human must collect on-device evidence |
| Findings | Fix blockers first; each review item is a missing piece of evidence |
| Evidence tiers | Tells you which artifact is static, host-only or planned |
| Not tested | Your to-do list before release |
| Artifact hashes | Compare with SHA256SUMS.txt to prove the files were not altered |
5. Use in CI
- run: npx edgesutra-ai audit --out audit --model m.json --device d.json --runtime r.json
- uses: actions/upload-artifact@v4
with: {name: audit, path: audit}
Exit code 1 fails the job when a static check fails.
6. Verify a report you received
npx edgesutra-ai verify --input audit-report.json # exit 2 if altered
7. Install the agent skills
npx edgesutra-ai install --ide claude --project .
The command never overwrites. Then ask your agent: "use EdgeSutra Prime to review this model's edge readiness". Skill discovery depends on your IDE; it is not guaranteed by file copy alone.
8. Common questions
- Why REVIEW_REQUIRED and never "OK"? Nothing here executes on your device.
- Can I trust the benchmark numbers? They are real timings of a toy model on this host, useful for instrumentation, not for device decisions.
- Does it upload anything? No.