One uses: line puts the Annex VII and Annex V gap checklist in your CI run: a verdict and a posture table in the job summary, report.json and summary.md as a build artifact, and an exit code you can gate a merge on once you are ready for it. The scan is free and runs offline. This page is about the CI mechanics; the regulation itself is covered in what the CRA requires in your technical documentation and the 21 items.
It is a composite action: a thin wrapper that npx-runs a pinned release of the published ceproof CLI, so the thing CI runs is byte-for-byte the release you can run on your own laptop. No bundled node_modules, nothing to trust that you cannot reproduce.
# .github/workflows/cra.yml
name: cra
on: [push, pull_request]
permissions:
contents: read
jobs:
cra:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
- uses: actions/setup-node@v4
with: { node-version: 22 }
- run: npm ci
- uses: thrillhouse34/ceproof-action@v1
with:
fail-on: never # start here: report, don't block
The CRA scan reads what is already on disk: package manifests, lockfiles, installed dependency metadata and git facts. It needs an npm manifest (package.json and a lockfile if you have one) or a Python one (pyproject.toml, requirements.txt, poetry.lock, uv.lock). C or C++ firmware has no lockfile to read and is not supported by the CRA scan today. The only network call in the whole job is npx fetching the pinned package from npm.
Begin with fail-on: never and mean it. A repository with no ceproof.config.json has every human-required item open, so the default threshold fails on day one — which teaches your team to ignore the check instead of answering it. Get the report landing green first, answer the items, then tighten.
fail-on | The job goes red when… | Use it |
|---|---|---|
never |
never — report only, always exit 0 | Day one, and on any branch where a compliance gap should not block a build |
human-required |
anything still needs an answer from a person | Once ceproof.config.json is committed and filled in. This is the CRA default |
partial |
anything is only partially evidenced, as well | Strictest. Sensible on a release branch, where partial evidence in the technical file is a real problem |
Those three are the whole set. Any other value is rejected with exit 2 rather than quietly defaulted. Left empty, the input follows the CLI's own per-profile default: human-required for the CRA scan, never for the FDA profile below.
The human-required items are answered in a ceproof.config.json that lives in your repository. Generate it once, locally, and commit it — then the same answers resolve in CI:
npx ceproof init
Answered items move from NEEDS-HUMAN-INPUT to DERIVED on the next run. Which items a scan can derive, and which seventeen need a person →
An explicitly-given config: path that does not exist is exit 2, not a silent fallback to “no config present”. A typo in that path would otherwise discard every answer you wrote and then fail the build on the items the config was answering.
Two things, both on by default.
The job summary gets a verdict line, a posture table (items evaluated, derived, partial, needs-your-input, config status) and one row per item that is not yet fully derived, with the requirement text and the evidence found so far. Set summary: false to skip it. Long requirement text is clipped in the table; nothing is clipped in report.json.
The artifact, named ceproof-report, holds both files:
report.json — the full report (schemaVersion: 1): every checklist item with its status, requirement summary, the exact CRA citation and the evidence behind it, plus the detected repository facts, the config status with the list of unanswered item ids, and any warnings.summary.md — the same Markdown as the job summary, for pasting into a ticket.The upload is wrapped in always(), so the artifact still lands when the scan failed the job. That is exactly the run you want it for. Rename it with artifact-name: if one workflow runs the action more than once, and turn it off with upload-artifact: false.
The outputs are there to gate on or print: outcome, total-items, derived-count, partial-count and needs-human-input-count, plus config-present, config-valid, report-path and summary-path.
- id: cra
uses: thrillhouse34/ceproof-action@v1
with:
path: products/gateway
config: products/gateway/ceproof.config.json
fail-on: never
- run: echo "${{ steps.cra.outputs.needs-human-input-count }} CRA items still need an answer"
Exit codes, which the job status follows: 0 nothing crossed the threshold, 1 something did, 2 usage or tool error (no project at path, a bad fail-on value, a missing config path), 3 a licence key was supplied and rejected.
Both SBOM formats are free — CycloneDX 1.6 and SPDX 2.3, no key, no account, no network call — but the action does not produce them without a key. Its report-formats input is only read when license-key is set, so a free SBOM in CI is a plain npx step next to the action:
- run: npx --yes ceproof@0.4.0 report . --format sbom-cyclonedx,sbom-spdx --out sbom
- uses: actions/upload-artifact@v4
with:
name: ceproof-sbom
path: sbom/
Worth knowing which CRA provision that satisfies and which it does not: the duty to draw up an SBOM is Annex I Part II(1), not the Annex VII documentation item, and the most specific public guidance in circulation sets a floor of CycloneDX 1.6 or SPDX 3.0.1 — the SPDX output is 2.3, which does not meet it. Where the CRA's SBOM obligation actually sits, quoted verbatim →
The free scan tells you where you stand. The paid report is the paperwork: the Annex VII technical documentation pack and the Annex V Declaration of Conformity draft, as HTML and PDF, plus both SBOM formats — compiled from the same checklist the free step just computed, stamped with the release they describe, regenerated on every run. Put your key in a repository secret and pass it in:
- uses: thrillhouse34/ceproof-action@v1
with:
license-key: ${{ secrets.CEPROOF_LICENSE_KEY }}
config: ceproof.config.json
fail-on: never
# report-formats: annex7-pdf,doc-pdf # narrow it if you only want the PDFs
Never paste the key as a literal — always a secret, and a secret is not readable by workflows from forked pull requests, which is the behaviour you want here. The key reaches the action only through an env: block the runner does not log, is masked before any other output, and is never written to the job summary, to an action output or to an artifact: the generated files are searched for it and the upload is refused outright if it is found.
It is one extra input on the same action, not a second action, so the pack can never disagree with the step above it about your posture. The pack lands as its own artifact, ceproof-full-report, which downloads independently of the free one. Defaults to all six formats: annex7, annex7-pdf, doc, doc-pdf, sbom-cyclonedx, sbom-spdx.
Two behaviours worth knowing before you wire it up. A rejected key fails the job and writes nothing — it does not fall back to the free scan, because a release going out without the pack you thought you had generated is worse than a red build; put continue-on-error: true on the step if you would rather it were advisory. And the report runs even when the scan failed the build, for the same reason the artifact upload does: with the default threshold, a repository with open items is red, and that is most repositories most of the time.
The obvious thing to do with it is keep the pack that matches each release:
on:
release:
types: [published]
One perpetual purchase, no subscription: $99 individual, $299 agency. Look at a finished pack →
Selling the same device in the United States? fda: true swaps which gap view the free scan computes: instead of the 21 EU CRA items it resolves the 13 SBOM content items in FDA's February 3, 2026 premarket cybersecurity guidance. Free, no key, no network call. Device firmware has no lockfile, so fda-sbom: is how the components get in — CycloneDX JSON, spec 1.4 to 1.6, one path per line or comma-separated.
- uses: thrillhouse34/ceproof-action@v1
with:
fda: true
fda-sbom: |
build/firmware.cdx.json
build/third-party.cdx.json
This profile is report-only: leave fail-on alone and the job exits 0 whatever the scan finds. That is deliberate. Three of the thirteen items — the safety and security risk assessment, the risk controls and the cyber-device determination — can only ever be answered by a human, so a threshold of human-required would fail every run forever no matter how complete your SBOM, and an exit code that is red on a flawless run is not a signal. Gate a build on it deliberately, with fail-on: partial.
The action does not generate the paid FDA evidence pack. license-key and report-formats drive the CRA Annex VII and Annex V report only, and an FDA-tier key passed against the CRA defaults is refused with exit 3 naming the tier mismatch. Run the pack locally for now: npx ceproof report --format fda-pack --sbom firmware.cdx.json.
The action is only a wrapper around the published CLI, so the same checks run anywhere Node 22 runs — GitLab CI, Jenkins, a local hook, whatever you have. Gate on the exit code and keep the --out directory as an artifact.
npx ceproof scan --fail-on never
npx ceproof report --format sbom-cyclonedx,sbom-spdx --out sbom
Pin the action by tag (@v1) or by commit SHA, and pin the CLI with the version: input if you want the checklist itself frozen across releases. Expect the derived/partial split to shift between a shallow CI checkout and your local clone: a real release tag is evidence for one of the Annex VII items, and a shallow clone may not have one.
What this is not. The action produces compliance-posture information. It is not a conformity assessment, not legal advice, and not a Declaration of Conformity — the Annex V draft is a draft until a responsible person reviews, completes and signs it. A green build means nothing crossed the threshold you set; it does not mean your product conforms. ceproof is source-available under the Elastic License 2.0.
Free to scan and to see the gap checklist, in CI or on your laptop — npx ceproof scan. $99, one-time to export the Annex VII pack and the Annex V DoC draft.