atmos ai skill install atmos-componentsAtmos Component Architecture
Components are the building blocks of infrastructure in Atmos. Each component is an opinionated, reusable unit of infrastructure-as-code -- typically a Terraform root module -- that solves a specific problem. Atmos separates the component implementation (code) from its configuration (stack manifests), enabling one implementation to be deployed many times with different settings.
What Components Are
In Atmos, a component consists of two parts:
- Implementation -- The infrastructure code itself (a Terraform root module, Helmfile, or Packer template) stored in the
components/directory. - Configuration -- The settings that customize how the component is deployed, defined in stack manifests under the
componentssection.
This separation is fundamental: you write the Terraform module once, then configure it differently for each environment, region, and account through stack YAML files.
Component Types
Atmos natively supports three component types:
| Type | Implementation Location | Purpose |
|---|---|---|
| Terraform / OpenTofu | components/terraform/<name>/ | Provision cloud infrastructure resources |
| Helmfile | components/helmfile/<name>/ | Deploy Helm charts to Kubernetes clusters |
| Packer | components/packer/<name>/ | Build machine images (AMIs, VM images) |
Terraform is by far the most common type. Custom commands can extend Atmos to support any tooling.
Directory Structure
Components are stored in your project's components/ directory, organized by type:
components/terraform/vpc/main.tfvariables.tfoutputs.tfversions.tfeks/cluster/main.tfvariables.tfoutputs.tfs3-bucket/main.tfvariables.tfoutputs.tfiam-role/main.tfvariables.tfoutputs.tfhelmfile/nginx-ingress/helmfile.yamlcert-manager/helmfile.yamlpacker/ubuntu-base/template.pkr.hcl
The base path for Terraform components is configured in atmos.yaml:
components:terraform:base_path: "components/terraform"
Nested directories are supported. A component at components/terraform/eks/cluster/ is referenced as eks/cluster in stack configurations.
Component Configuration in Stacks
Components are configured in the components section of stack manifests:
components:terraform:vpc:metadata:component: vpc # Points to components/terraform/vpc/vars:cidr_block: "10.0.0.0/16"availability_zones:- us-east-1a- us-east-1bsettings:validation:check-cidr:schema_type: jsonschemaschema_path: schemas/vpc.jsoneks-cluster:metadata:component: eks/cluster # Points to components/terraform/eks/cluster/vars:cluster_name: prod-ekskubernetes_version: "1.28"
Each component configuration can include these sections:
| Section | Purpose |
|---|---|
metadata | Component location, inheritance, type, and Atmos behavior |
vars | Input variables passed to Terraform/Helmfile/Packer |
env | Environment variables set during execution |
settings | Integration metadata such as Atlantis, validation, and custom settings |
dependencies | Component, file, folder, and tool dependencies |
hooks | Lifecycle event handlers |
backend / backend_type | Terraform state backend configuration |
providers | Terraform provider configuration |
command | Override the executable (e.g., tofu instead of terraform) |
auth | Authentication identity reference |
Dependencies
Use dependencies.components for component ordering and affected/dependent analysis:
components:terraform:app:dependencies:components:- component: vpc- component: databasestack: plat-ue2-prod- kind: filepath: configs/app.yaml- kind: folderpath: src/lambda
Use dependencies.tools for runtime CLIs at global, component-type, or per-component scope; Atmos auto-installs and injects them, so do not add a separate install step.
Do not add new settings.depends_on examples. If a repository already uses that legacy field,
recommend migrating it to dependencies.components.
Source Provisioning and Workdirs
Components can declare source in stack config for just-in-time provisioning from Git, OCI, S3,
HTTP, or local paths. Treat source provisioning as part of component configuration: it controls
how the component implementation is fetched, while vars, env, backend, and other sections
control how that implementation is run.
Use component source provisioning when different stacks need different remote component versions,
or when the repository should not commit the fetched implementation. When source provisioning is
used, prefer provision.workdir.enabled: true so each component-stack instance gets an isolated
execution directory.
components:terraform:vpc:source: "github.com/cloudposse-terraform-components/aws-vpc.git?ref=1.450.0"provision:workdir:enabled: true
Atmos provisions sources automatically before Terraform commands when needed. Use explicit source commands when an agent needs to inspect, refresh, or clean the fetched implementation:
atmos terraform source pull vpc --stack devatmos terraform source describe vpc --stack devatmos terraform source list --stack devatmos terraform source delete vpc --stack devatmos list sources
For protected sources, route identity and provider details to atmos-auth. For checked-in copies
of remote components, route to atmos-vendoring; vendoring and source provisioning are alternatives
for obtaining component implementation code.
Abstract vs Real Components
Abstract Components
Mark a component as metadata.type: abstract to create a blueprint that cannot be deployed directly:
components:terraform:vpc/defaults:metadata:type: abstractcomponent: vpcvars:enabled: truenat_gateway_enabled: truemax_subnet_count: 3vpc_flow_logs_enabled: true
Abstract components:
- Cannot be provisioned with
atmos terraform apply(Atmos returns an error). - Do not appear in
atmos describe stacksoutput by default. - Serve as base configurations for real components to inherit from.
Real Components (Default)
If metadata.type is not specified, the component is real and can be deployed:
components:terraform:vpc:metadata:inherits:- vpc/defaultsvars:vpc_cidr: "10.0.0.0/16"
Component Inheritance
Single Inheritance
Use metadata.inherits to inherit configuration from a base component:
components:terraform:vpc/defaults:metadata:type: abstractcomponent: vpcvars:enabled: truenat_gateway_enabled: truevpc:metadata:inherits:- vpc/defaultsvars:nat_gateway_enabled: false # Override inherited value
The derived component receives all vars, env, settings, hooks, backend, providers, and command from the base, then its own values are deep-merged on top.
Multiple Inheritance
A component can inherit from multiple bases. Entries are processed in order with later entries having higher precedence:
components:terraform:rds:metadata:component: rdsinherits:- base/defaults # Applied first- base/logging # Applied second- base/production # Applied last (highest base precedence)vars:name: my-database # Inline has highest precedence
This enables composing "traits" -- reusable abstract components that represent independent configuration concerns (logging, security, sizing, environment settings).
metadata.component
The metadata.component field maps an Atmos component name to its Terraform root module directory:
components:terraform:vpc-main:metadata:component: vpc # Uses components/terraform/vpc/vars:name: main-vpcvpc-isolated:metadata:component: vpc # Same Terraform modulevars:name: isolated-vpcenable_internet_gateway: false
Both components use the same Terraform code but maintain separate state files and configurations. This is the multiple component instances pattern.
Multiple Component Instances
Deploy the same Terraform module multiple times in the same stack by giving each instance a unique Atmos component name:
components:terraform:vpc/1:metadata:component: vpcinherits:- vpc/defaultsvars:name: vpc-1ipv4_primary_cidr_block: 10.9.0.0/18vpc/2:metadata:component: vpcinherits:- vpc/defaultsvars:name: vpc-2ipv4_primary_cidr_block: 10.10.0.0/18
Each instance has its own Terraform state and is independently deployable:
atmos terraform apply vpc/1 -s plat-ue2-prodatmos terraform apply vpc/2 -s plat-ue2-prod
metadata Section Fields
All fields available in the metadata section:
| Field | Type | Description |
|---|---|---|
component | string | Path to Terraform root module relative to components base path |
inherits | list | List of component names to inherit configuration from |
type | string | abstract (non-deployable) or real (default, deployable) |
name | string | Stable logical identity for workspace key prefix |
enabled | boolean | Enable or disable the component (default: true) |
locked | boolean | Prevent modifications to the component |
terraform_workspace | string | Explicit workspace name override |
terraform_workspace_pattern | string | Workspace name pattern with tokens |
custom | map | User-defined metadata (preserved, not interpreted by Atmos) |
Catalog Patterns
The stacks/catalog/ directory is the conventional location for reusable component configurations:
stacks/catalog/vpc/_defaults.yaml # Abstract base for all VPC instanceseks/_defaults.yamlcluster.yamls3-bucket/_defaults.yamliam-role/_defaults.yaml
Catalog files define abstract components with sensible defaults. Top-level stacks import from the catalog and override only what differs:
# stacks/catalog/vpc/_defaults.yamlcomponents:terraform:vpc/defaults:metadata:type: abstractcomponent: vpcvars:enabled: truenat_gateway_enabled: truemax_subnet_count: 3
# stacks/orgs/acme/plat/prod/us-east-1.yamlimport:- catalog/vpc/_defaultscomponents:terraform:vpc:metadata:inherits:- vpc/defaultsvars:vpc_cidr: "10.0.0.0/16"
Mixins for Reusable Configuration
Mixins are small, focused configuration snippets that alter component behavior. They are typically stored in stacks/mixins/ and imported into stacks:
stacks/mixins/region/us-east-1.yamlus-east-2.yamlus-west-2.yamlstage/dev.yamlstaging.yamlprod.yamltenant/plat.yaml
# stacks/mixins/region/us-east-2.yamlvars:region: us-east-2environment: ue2
# stacks/orgs/acme/plat/prod/us-east-2.yamlimport:- mixins/region/us-east-2- mixins/stage/prod- catalog/vpc/_defaults
Remote State Access Between Components
Components can access outputs from other components using the remote-state module or YAML functions:
# Using YAML functions (state access is the fastest option)components:terraform:eks-cluster:vars:vpc_id: !terraform.state vpc vpc_idsubnet_ids: !terraform.state vpc private_subnet_ids
For a cold terraform plan --all, upstream state may not exist yet even though component
dependencies establish deployment order. Put a provider-valid fallback in the YQ expression; the
real state value supersedes it after the upstream component deploys:
components:terraform:eks-cluster:vars:vpc_id: !terraform.state vpc '.vpc_id // "vpc-mock"'
For the Terraform-side approach, use the remote-state module:
# In components/terraform/eks/cluster/remote-state.tfmodule "vpc" {source = "cloudposse/stack-config/yaml//modules/remote-state"version = "1.5.0"component = "vpc"}
This reads the VPC component's Terraform outputs from the same or a different stack.
Best Practices
- One concern per component: Each component should provision a single logical piece of infrastructure (VPC, EKS cluster, database). Do not combine resources with different lifecycles.
- Use abstract base components: Define catalog defaults as
abstractcomponents and inherit from them. - Keep inheritance chains shallow: Limit to 2-3 levels for readability and debuggability.
- Use metadata.component for instances: When deploying the same module multiple times, use
metadata.componentto share the implementation. - Use metadata.name for versioning: Set
metadata.nameto maintain stable Terraform workspace key prefixes across version upgrades. - Design for reuse: Components should accept configuration through variables, not hard-coded values. Use the catalog pattern to define sensible defaults.
- Use
atmos describe component: Always verify the resolved configuration before applying changes.
atmos describe component vpc -s plat-ue2-prod
References
- references/component-types.md -- Detailed reference on component types, metadata fields, abstract components, inheritance chains, and versioning
- references/examples.md -- Concrete configuration examples for VPC, EKS, S3, IAM patterns