airom — Python SDK

Python SDK for AIROM, the open-source AI Bill of Materials (AIBOM) scanner. Discover AI assets — models, prompts, datasets, embeddings, vector databases, frameworks, serving infrastructure — across code, containers, and Kubernetes, and get them back as typed Python objects.

pip install airom

This installs both the airom command and the Python library — the wheel ships the scanner binary itself into your environment's bin/, so there is nothing else to install:

$ airom --version
airom 0.4.7

$ airom fs ./my-app -o table          # the CLI, globally

Quick start

import airom

inv = airom.fs("./my-app", min_confidence=0.8)

for c in inv.by_kind(airom.ComponentKind.HOSTED_LLM):
    print(c.name, c.provider.or_default("-"), c.confidence)
    for occ in c.evidence.occurrences:
        print(f"   {occ.location.path}:{occ.location.line}  [{occ.detector_id}]")

gpt-4.1 openai 0.87
   src/rag.py:88  [rules/openai/model-literal]
   src/agent.py:12  [rules/openai/sdk-import]

Every component carries the evidence that justifies it — that is the point of AIROM, and the SDK hands you all of it.

Scanning

airom.scan("./app")                       # auto-detect: path, git URL, or image ref
airom.fs("./app")                         # a directory tree
airom.repo("https://github.com/o/r")      # remote (shallow clone) or a local worktree
airom.image(input="img.tar")              # docker save -o img.tar <ref>
airom.k8s(manifests="./deploy")           # offline: enumerate workload images
airom.version()                           # the underlying binary's ToolInfo

Common keyword args mirror the CLI flags: select, rules, ignore, min_confidence, max_file_size, io_budget, parallel, no_cache, cache_dir, offline, stats, plus binary, timeout, and cwd. None leaves the tool's own default in place — the SDK never invents defaults.

min_confidence=0.8 is the practical high-signal filter: on general-purpose directories, extension-only dataset detection and keyword-only generation-param detection emit low-confidence (0.5–0.6) noise. Note the application root always survives the filter — it is the scan target, not a finding.

select tokens are detector IDs or tags, not languages — "-dataset/file", not "python". Run airom detectors list (or airom.raw(["detectors", "list"])) to see them.

Not wired yet: pulling an image from a live registry/daemon, and live-cluster Kubernetes scanning. Both fail with a clear error. Use image(input=...) / an OCI layout and k8s(manifests=...) today.

Tri-state fields

version, provider, download_location and release_time are tri-state, and the SDK preserves the distinction rather than collapsing it into None:

c.version.known           # bool — only True when a real value is present
c.version.or_none()       # value, or None (collapses absent and unknown)
c.version.or_default("-") # value, or your fallback
c.version.presence        # the full distinction, when you need it

Navigating the graph

inv.components                  # sorted, deterministic
inv.by_kind("vector-db", "framework")
inv.get("airom:1f3a9b2c4d5e6f70")
inv.application                 # the scan-root component
inv.edges_from(c.id)            # typed, evidenced relationships
inv.unknowns                    # "looked relevant, could not process" — honesty channel
inv.stats.files_walked          # requires stats=True
len(inv); [c for c in inv]      # Inventory is sized and iterable

CI gating

A fail_on match is a verdict, not an error — the scan succeeded and the AIBOM is complete, so it is reported rather than raised:

res = airom.execute(
    ["fs", "./app"],
    options=airom.ScanOptions(fail_on="hosted-llm&confidence>=0.9", exit_code=7),
)
if res.policy_matched:
    raise SystemExit(res.exit_code)

Often you don't need fail_on at all — you have the whole graph, so gate in Python:

risky = [c for c in inv if c.model and c.model.pickle_risk]
if risky:
    raise SystemExit(f"unsafe pickle globals in: {[c.name for c in risky]}")

Errors

Detector errors are not exceptions: they degrade to inv.unknowns records and the scan still succeeds. That is AIROM's degrade-by-default contract, and the SDK preserves it.

The binary

The SDK shells out to the airom binary and decodes its native JSON — the lossless superset every other format (CycloneDX, SARIF, YAML, table) projects from. Resolution order:

1. the binary= argument
2. a copy bundled in the wheel (airom/_bin/airom)
3. $AIROM_BINARY
4. airom on PATH

Platform wheels bundle the binary as a script, so pip installs it into the environment's bin/ (Scripts\ on Windows): pip install airom gives you a working airom command on PATH and the importable library, with no Python shim in between. In a virtualenv it is on PATH as soon as the venv is active; pipx install airom or pip install --user airom puts it on PATH globally.

The library resolves the binary without relying on PATH at all (it consults the environment's scripts dir directly), so import airom works even from an unactivated venv's interpreter.

Platforms with no wheel

Wheels are published for macOS (Intel + Apple Silicon), Linux (x86-64 + arm64, glibc + musl), and Windows (x86-64 + arm64). Every one is installed and executed on the platform it targets before it is published; see the smoke jobs in release-pypi.yml.

Anywhere else, pip falls back to the sdist, which ships the Python library and no binary — and cannot build one, since it does not contain the Go module. Rather than install successfully and leave you without a scanner, that build now fails and says so. To proceed, either install the binary for your platform from the releases page and put it on your PATH, or take the library alone:

AIROM_SKIP_BUNDLE=1 pip install airom

The second form is the right one when you already run the standalone binary and only want import airom. Resolution then falls through to $AIROM_BINARY or airom on PATH.

Development

cd sdk/python
pip install -e ".[dev]"
pytest            # builds the binary from the checkout and tests against it
mypy && ruff check .

The suite runs against the real binary, not mocks: a wrapper tested only against mocks proves nothing about the contract it wraps.

Building a wheel needs the Go toolchain (the build hook compiles the binary with CGO_ENABLED=0) and fails without it, because a wheel with no binary installs no airom command. Set AIROM_SKIP_BUNDLE=1 for a pure-Python wheel on purpose.

Publishing

Releases are published by .github/workflows/release-pypi.yml using PyPI Trusted Publishing (OIDC): GitHub Actions authenticates to PyPI directly, so there is no API token stored as a secret, and none to leak or rotate.

One-time setup

On pypi.org/manage/account/publishing, add a pending publisher:

That is all — no secret is added to GitHub.

Cutting a release

The workflow runs on every push to main that touches the SDK or the scanner, but publishes only when the version is new: PyPI permanently refuses to re-upload a version (even a deleted one), so the workflow compares __version__ against the index and skips cleanly if it is already there.

So a release is exactly one deliberate act:

# sdk/python/src/airom/__init__.py
__version__ = "0.1.0"     # bump, commit, merge to main -> published

Publishing is irreversible: a version number is burned forever once used, and yanking does not free it. Test the whole path first with the manual workflow_dispatch run against TestPyPI (which needs its own pending publisher at test.pypi.org).

License

Apache-2.0, same as AIROM.