> ## Documentation Index
> Fetch the complete documentation index at: https://docs.forge.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Cloud artifact scanning

> Discover and scan cloud images and opt-in Linux filesystem snapshots.

Cloud artifact scanning finds AI software, MCP configuration, skills, models,
packages, and workload evidence inside immutable container images and opt-in
cloud filesystem snapshots.

## Supported sources

| Provider     | Registry images          | Filesystem snapshots                                                  |
| ------------ | ------------------------ | --------------------------------------------------------------------- |
| AWS          | Amazon ECR               | Opt-in EC2 Linux volumes through temporary scanner resources          |
| Azure        | Azure Container Registry | Opt-in Azure Linux disks through a dedicated scanner resource group   |
| Google Cloud | Artifact Registry        | Opt-in Compute Engine Linux disks through configured scanner networks |

OCI images are acquired by registry, repository, platform, and digest. Multiple
tags or workloads that resolve to the same immutable artifact share one current
scan result. Snapshot scanning currently supports Linux `amd64` and `arm64`
filesystems.

## What Forge inspects

The scanner reads bounded package databases, dependency manifests, OCI config,
executables, MCP declarations, `SKILL.md` files, and recognized model files. It
emits paths, hashes, safe metadata, confidence, parser revision, and coverage
gaps. It does not return secret values or unrestricted file contents.

An MCP SDK dependency is evidence, not proof of a running MCP server. A
parseable skill manifest produces a skill identity linked to the image or
workload occurrence.

## Safety and cleanup

Image scans are digest-pinned. Snapshot scans create only deterministically
named, Forge-tagged temporary resources in the configured provider scope. The
one-off scanner receives no Forge database credential or saved cloud-connection
credential, and result upload uses a single-use job-and-attempt capability.

After scanning, Forge detaches and deletes temporary resources and verifies
their absence. A separate cleanup worker reconciles expired tagged resources
after worker or container replacement. Forge does not delete untagged customer
resources.

## Limits and partial coverage

Artifact discovery and scanning are bounded by page, item, byte, layer,
file-count, archive-depth, duration, evidence-size, attempt, and concurrency
limits. Discovery cursors resume later syncs instead of requiring one unbounded
provider walk.

When a scan reaches a bound, Forge keeps the evidence already collected and
marks the result partial with the exact coverage gap. Partial coverage never
claims that an artifact is clean or that a component is absent.

## Setup and verification

1. Open the AWS, Azure, or Google Cloud integration and select **Guided setup**.
2. Apply the current inventory artifact for registry discovery and pull access.
3. Enable filesystem snapshot scanning only for approved source scopes and
   apply its additional setup artifact.
4. Run **Test**, then **Sync**.
5. Confirm discovered digests, scan coverage, workload links, and cleanup state.

Registry scanning requires queue-capable control-plane workers in the Forge
deployment. Snapshot setup also requires exact scanner AMIs and regions on AWS,
scanner networks on Google Cloud, or source subscription and resource group on
Azure.