Skip to main content

How AI TrustBOM works

Scan your repository, have Trust approve what it found, then send the approved pack to your buyer. The approval lives in Git, next to your code.

Three steps

  1. Scan

    trustbom scan reads tracked lockfiles, manifests, allowlisted configuration and YAML detector packs. It combines that evidence with the manual declarations your engineers write in trust/manual.yml into one typed graph of models, tools, MCP servers and declared data flows. Each row cites its file.

  2. Review and approve

    Trust, the person at your company allowed to publish how the product uses AI, approves the exact graph by changing trust/approvals.yml. That file is protected by CODEOWNERS, so the approval is an ordinary reviewed pull request, and a different reviewer must approve it.

  3. Send the pack

    trustbom export writes the approved pack. trustbom answer fills a buyer questionnaire CSV from the approved graph, leaving anything it cannot support for a person to answer. trustbom publish produces an offline HTML and PDF pack.

What detection is, and is not

  • Detection is manifest, configuration and pattern-pack based. It is not a runtime proxy and does not observe live traffic or prove what ran in production.
  • Coverage is bounded to supported signals, such as Anthropic, OpenAI, Amazon Bedrock, Google Vertex AI and Azure OpenAI SDK use, MCP configuration and evaluation workflows. Anything else, you declare manually, with a reason.
  • Scanner output is a draft. Classifications such as risk category are drafts too, and a human decision is recorded separately by Trust.
  • No evidence is never turned into a claim. Unknown stays unknown.
  • AI TrustBOM does not certify compliance with any law, standard or contract, and it does not replace a human decision.

Keeping the pack accurate

trustbom init adds a GitHub Action workflow to your repository. On every pull request it re-scans, compares the result with the approved graph and posts one comment and an AI TrustBOM check. A change in models, tools or MCP servers, or an unresolved finding, fails the gate until Trust approves the new graph or the change is corrected.

Moving from one AI system to several is supported: a portfolio scan keeps a separate approval per system, and trustbom pack combines independently approved systems into one cover sheet and pack.

Everything runs on your machine or in your own CI runner. Repository contents are not sent to us, and the product has no usage telemetry.

Free and paid commands

Free, no time limitNeeds a trial or subscription
init, scan, gate (any number of systems), readiness, request-approval, default-branch stamp, export --draft (watermarked, never approved), the Action's scan and gate check, and license commands Approved export, publish, pack, answer and drift, including workflow steps that run them

See pricing for the trial and subscription, install for the quick start, and the example pack for what the output looks like.