AI supply-chain BOM: agents, MCP servers, packages and recorded reach. Scan and investigate via MCP.
Model Context Protocol (MCP) Server: io.github.koda-ai-studio/agent-bom
This MCP server provides a capability to scan AI/MCP dependencies, identify descriptor injection concerns, evaluate risky tool inputs, and review authentication configuration. Its purpose is to surface security-relevant issues related to dependency usage and configuration.
🛠️ Key Features
Scans AI/MCP dependencies
Detects descriptor injection risks
Identifies risky tool inputs
Assesses authentication configuration
🚀 Use Cases
Auditing AI/MCP projects for dependency-related security concerns
Reviewing agent descriptors for injection-related issues
Checking tool input handling for risk factors
Validating authentication configuration in AI/MCP setups
⚡ Developer Benefits
Centralizes checks for dependencies, tool inputs, and authentication configuration
Highlights configuration and input risk areas during development and review
⚠️ Limitations
Limited description is available: specific scan depth, outputs, and supported formats/tools are not provided in the available data.
AI supply-chain BOM: supported agents, MCP servers and AI packages, and what each one can reach.
For AppSec teams and developers who run Claude Code, Cursor, Codex or any MCP
client: see which agents, MCP servers, tools and packages are installed, which
are vulnerable, and what each one can reach. Local scans need no account.
From the repository you want to inspect, with uv installed:
bash
uvx agent-bom scan . -f html -o agent-bom-report.html
Open agent-bom-report.html for inventory, findings and fix guidance. The scan
stays inside this repository: packages, MCP configuration, secrets and personal
data, skills and IaC; it does not read agent configuration elsewhere on the machine.
Run the same scan again later and agent-bom prints what changed since the last one.
No MCP configuration in the project? Package findings still apply.
No project handy? Run uvx agent-bom scan --demo --offline for a bundled sample;
Exit 1 is expected there because the sample trips the security gate. A non-zero
exit can mean a security gate or incomplete assessment; inspect the report and
coverage (exit codes, what a scan does and
does not establish,
first-run guide).
From laptop to team to org
Where
Start with
You get
On your laptop
uvx agent-bom scan . · claude mcp add agent-bom -- uvx agent-bom mcp server
Findings with fixes, a delta on every rescan, and agent-bom as a tool inside your coding agent (MCP server, plugins)
In your team
uses: koda-ai-studio/agent-bom@v0 on pull requests and a weekly schedule:
A PR gate with a findings comment and SARIF in code scanning, plus a weekly rescan that catches new CVEs in unchanged code (GitHub Action)
Across your org
agent-bom up locally, then Docker Compose or the Helm chart from oci://ghcr.io/koda-ai-studio/charts/agent-bom
Retained scans, one graph of agents, MCP servers, packages and credentials, SSO and audit, in your own infrastructure (self-host)
Why an AI supply-chain BOM
Agents and MCP servers ship without an inventory. A CVE in an AI package, a
poisoned tool description or an over-scoped credential matters by what the
agent that loads it can reach. A package SBOM stops at packages; agent-bom
records the chain and keeps the evidence for every edge.
One graph: agent → MCP server → tool → package → credential → data, each edge with its source
Fix, gate CI, block poisoned tool calls, rescan to verify
Fictional sample estate. Dashed use, credential and access links are configured
or modeled, not observed at runtime; path analysis covers the collected snapshot only.
The AWS badge describes environment membership; these sample agents have no
recorded execution host or model endpoint.Agent, hosting and model context.
Start with the workflow you need. An individual developer can scan, inspect
findings and verify fixes without a shared control plane or a compliance
workflow. Teams can add retained scans and shared inventory; organizations can
add identity, audit review and governance when needed. These are adoption paths,
not separate editions or permissions. See the
deployment chooser.
Offering
What it includes
Cost
Start
CLI, CI and exports
Local and CI scans; SARIF, CycloneDX, SPDX, JSON and HTML exports; check pre-install gate; GitHub Action
Agents, MCP servers, tools and credentials from local configs and agent frameworks, linked to what they can reach.
Models and AI platforms
Supported OpenAI, Hugging Face, Ollama, MLflow and W&B assets, with available model provenance beside dependencies and agent context.
Runtime
An MCP proxy and gateway that alert on poisoned tools and injected results, or block them with --injection-mode enforce.
SBOM, AI-BOM and Agent-BOM
Software SBOMs in CycloneDX or SPDX. CycloneDX AI components carry collected model metadata. Native agent-bom JSON connects agents, MCP servers, tools, credential references and packages.
Dependencies
Lockfiles and manifests across Python, JavaScript, Go, Java, .NET, Rust and more, matched against OSV, GitHub advisories and NVD with KEV and EPSS.
Code
Findings imported from any SARIF producer, plus native analysis of AI code (prompt-injection sinks, guardrails, tool exposure) and whether your code reaches the vulnerable function.
Containers and IaC
Images scanned in bounded memory and time; Dockerfile, Terraform, Kubernetes, Helm and compose risks with exact lines in SARIF.
Cloud and data platforms
Read-only collection from AWS, Azure, GCP, Kubernetes, Databricks and Snowflake, with provider-specific posture checks and graph context. Per-lane rule counts are generated from code in the capability matrix.
GPU and AI infrastructure
AI workloads on CoreWeave, Nebius, Lambda, RunPod, Vast.ai and Crusoe; GPU containers and Kubernetes GPU nodes. Collection depth varies by provider.
Generate a BOM from a project, or inspect an agent inventory with sample data:
Product names and logos are trademarks of their owners, used only to identify supported evidence sources. Their use does not imply endorsement or partnership.
Investigate context and exposure
Which agent uses the vulnerable component? What could it reach? What evidence supports that path?
Correlate retained scans by exact, tenant-scoped identities and trace the
recorded blast radius from each agent through its MCP servers to packages,
findings, credentials and data stores. Each relationship keeps its source and
evidence kind. Agent kind, platform, execution host and model links are
separate facts: an environment badge does not say where an agent runs.
From agent to verified fix: select Document Intake Agent → inspect
CVE-2023-4863 (high, KEV-listed) and the evidence on each arrow → take the
proposed remediation →
rescan the changed pin and compare snapshots.
The comparison verifies the changed input, not a running deployment.
Am I affected by a new CVE? Match it against stored scans, including CVEs
only NVD has published. You get affected assets with their patch target,
possibly affected distro packages to verify, and the blast radius
(walkthrough):
bash
agent-bom db update # refresh OSV, NVD, KEV and EPSS
agent-bom db advisory affected CVE-2026-109618 # most recent saved scan
agent-bom db advisory affected CVE-2026-109618 --report <scan-report.json> --format json # a JSON report written by an earlier scan
From an MCP client, the read-only affected_assets tool (graph and full
profiles) takes the same advisory_id. In CI, run uvx agent-bom db update
before the GitHub Action step (its auto-update-db default only refreshes
feeds that missed the daily freshness target). To explain a finding to your team,
create an impact brief with affected assets,
source receipts, evidence gaps and the next action.
More sample graphs: package evidence, repository and cloud (3)
One package, three parsed inputs. A repository pin, an image SBOM and a
modeled Kubernetes workload resolve to one pillow@9.0.0. The image inventory
and the deployment occurrence stay separate nodes; the advisory receipt is open.
Real parsers, modeled infrastructure; a graph path does not prove exploitation.
Repository. Source imports, deployment configuration and CI traced to a
published container and its vulnerable packages:
Cloud. A public API gateway into a vulnerable workload, its role and a PII bucket:
Try the demo in your browser: a read-only copy of the UI on
labelled sample data, with no install and no backend. Scans, settings and
sign-in are turned off there. To open the same graph locally:
Run the server command it prints, then open its graph URL. Offline quickstart
skips package-CVE lookup, so zero CVEs is not a clean verdict; use the bundled
demo for advisory-backed findings or populate the advisory database.
Self-host in your environment
Your infrastructure, identity, database, and audit boundary. Start a local
pilot from a published release checkout:
bash
git clone --depth 1 --branch v0.114.0 https://github.com/koda-ai-studio/agent-bom.git && cd agent-bom
AGENT_BOM_IMAGE_TAG=0.114.0 docker compose up -d
Open http://localhost:3000, then Connections or New Scan. The pilot
binds to loopback and retains state in a Docker volume. For cloud accounts, add a
scoped read-only connection, verify access, then start a scan. Before sharing
access, move to the authenticated deployment.
Scan with CLI or CI; export findings, SBOMs and reports.
Centralize inventory, graph evidence, remediation and compliance in your own control plane.
Enforce policy on selected live MCP tool calls through the proxy or gateway.
Sources describes recorded evidence: local scans, fleet inventory, cluster
scans, CI/CD, imported evidence, or mixed sources. It does not select a
deployment type or enable runtime enforcement. Demo data marks labelled
sample inventory.
Assets are inventoried resources; Findings are detections on those
resources; Issues group related findings for investigation or remediation.
Affected assets counts distinct resources only when that count is
available. One finding may belong to several security areas, so area counts
must not be summed. Repeated observations across scans are shown separately.
Preview and community deployment paths have separate
support boundaries.
In Codex, run codex plugin marketplace add koda-ai-studio/agent-bom, then
install agent-bom-local from the Plugins Directory. Both plugins run the
pinned agent-bom release locally and send nothing to an agent-bom service.
Ask
Skill
"Is this project safe to ship?"
checkup: one verdict and the three fixes that matter most
"Am I affected by CVE-2025-29927?"
affected: affected or not found in what was scanned, never "safe"
"What have my agents been doing?"
activity: MCP usage, unused servers, risky commands from local session history
"Are my agent settings too permissive?"
audit-mcp and audit-skills
"Give me an AI-BOM"
bom: CycloneDX or SPDX for this machine, a project, an image or a cloud account
Use CLI or GitHub Action, REST API, or MCP; export SARIF, CycloneDX,
SPDX, JSON and HTML. Cloud connectors and fleet sync collect inventory;
proxy and gateway deployments add runtime evidence.
For other MCP clients, run agent-bom mcp server.
Start with eight focused tools in the default profile.
Choose an MCP profile for graph, cloud, runtime or audit work.
Try three investigations: affected agents, exposure receipts and a before/after scan comparison.
Import existing scanner reports without replacing their source evidence.
Check a dependency before installing it with
uvx agent-bom check requests@2.33.0 --ecosystem pypi. Add
pre-commit and CI setup for repeated checks.
Discovery uses read-only access by default. Explicit disk side-scans create
temporary cloud resources. Missing evidence stays unavailable or partial;
control mappings are not audit certification.
Runtime evidence and MITRE mapping boundaries
Persisted runtime, proxy and gateway activity retains source receipts and
bounded time windows in the graph. Exact inventory references can bind records;
unresolved identities and collection gaps remain explicit. Recorded activity
does not prove execution or grant access. See the runtime evidence boundaries.
MITRE technique associations provide investigation context, not proof of an
attack. ATLAS mapping receipts retain the weakness
association and source catalog hashes; modeled graph paths do not establish
observed execution or effective access.