Start here by role¶
agent-bom inventories the AI agents and MCP servers you run, the packages and models they load, and what each one can reach: credentials, tools, code and cloud. Every finding sits on one graph, so a vulnerable package is ranked by what the agent that loads it can touch.
Pick your role. Each path takes about five minutes on a laptop with uv installed. Scans run locally and need no account; online scans send package names and versions to public advisory sources.
| Role | First question answered |
|---|---|
| AppSec | Which vulnerabilities can an agent actually reach, and what do I fix first? |
| Platform and DevOps | How do I gate pull requests and run this for the team? |
| AI and ML engineer | What can my agents, MCP servers and skills reach? |
| GRC | What evidence maps to the AI frameworks I report on? |
| Leadership | What is our AI supply-chain posture, and what are the top risks? |
AppSec¶
You get findings ranked by contextual risk: severity, known or likely exploitation, and reach to agents, credentials and tools. Each comes with a fix version and a command to verify the fix.
uvx agent-bom scan .
uvx agent-bom scan . -f sarif -o agent-bom.sarif
uvx agent-bom check requests@2.28.0 --ecosystem pypi
Look at the Blast Radius column (agent → MCP server → credential) and
PROTECT | Fix First, where a KEV-listed package is priority P1. A partial
scan is not a clean one: read the coverage warnings before you trust a zero.
check gates a single dependency before it is installed.
Next: gate pull requests with the GitHub Action and see the same chain in the Blast Radius guide.
Platform and DevOps¶
You get a pull-request gate with SARIF and a findings comment, the same gate on your laptop, and a local control plane that retains scans and serves the graph.
Look at the exit code: 0 passes, 1 is a security verdict or an
incomplete assessment (exit codes). up serves
the API and dashboard on http://127.0.0.1:8422; open Connections to add a
read-only cloud source or New Scan to scan a repository.
Next: choose a deployment: the one-machine Docker pilot, then the authenticated self-hosted instance or the Helm chart.
AI and ML engineer¶
You get every agent and MCP server configured in the project and on this machine, the packages behind them, the credential names they are given (never the values).
Look at DISCOVER | Agents (servers, packages and credentials per
agent) and the AI agent can reach a credential or privileged tool
findings. The demo uses labelled sample data and exits 1 on purpose: the
sample trips the security gate. skills scan reads instruction and skill
files such as CLAUDE.md and .cursor/rules for risky behavior, for example a
remote script piped to a shell; it exits 2 when the project has none.
Next: use agent-bom as a tool inside your coding agent with the MCP server, and see which AI platforms and GPU clouds are covered in the AI infrastructure coverage matrix.
GRC¶
You get a machine-readable report, an auditor-facing narrative per framework (OWASP LLM, OWASP MCP, OWASP Agentic, MITRE ATLAS, EU AI Act, NIST AI RMF and more) and an AI-BOM in CycloneDX.
uvx agent-bom scan . -f json -o agent-bom-report.json
uvx agent-bom report compliance-narrative agent-bom-report.json --framework eu-ai-act -o eu-ai-act-narrative.md
uvx agent-bom scan . -f cyclonedx -o ai-bom.cdx.json
Look at the controls marked not evaluated as well as the failures: an unevaluated control is missing evidence, not a pass. Mappings are review evidence, not audit certification.
Next: the compliance mappings, and the signed compliance pack from a control plane described in the executive brief.
Leadership¶
You get one page: a posture grade, the top remediation priorities with owner and SLA, the ranked exposure paths and technical control evidence.
This scans a bundled sample AI stack, starts the local dashboard and opens its security graph. From Overview, open Executive brief. To export the brief as Markdown from your own control plane, start the server with a key you choose, then request the brief from a second terminal with the same key:
export AGENT_BOM_API_KEY="$(python3 -c 'import secrets; print(secrets.token_urlsafe(32))')"
uvx --from 'agent-bom[ui]' agent-bom serve --persist ~/.agent-bom/control-plane.db
uvx agent-bom report brief --api-url http://127.0.0.1:8422 > executive-brief.md
Look at the grade and the open issue groups, then the ranked exposure paths: each one starts at an agent and ends at what it can reach. Counts are correlated issue groups, not raw finding occurrences.
Next: have the team run the authenticated self-hosted instance so the brief covers your real estate, then read the executive brief guide.