Tarmac CLI (tarmac-cli)

Tarmac is a declarative, idempotent Python CLI for bootstrapping and governing enterprise multi-account AWS Organizations.

Define your organizational units, account factory, SCP guardrails, baseline security (CloudTrail, GuardDuty, Security Hub), and IAM Identity Center (SSO) in simple YAML manifests—then reconcile your entire AWS estate with a single command.

Quickstart

1. Installation & Setup

Install globally from PyPI with uv:

uv tool install tarmac-cli
tarmac --help

Or install from source for local development:

git clone https://github.com/nickdchristian/tarmac-cli.git
cd tarmac-cli
uv sync --locked
uv run tarmac --help

2. Initialize Configuration

Scaffold a production-ready starter project with baseline YAML manifests and CloudFormation templates:

tarmac init --with-templates

This generates:

• config/organization.yaml: Organization settings, alternate contacts, OU hierarchy, and SCP guardrails.
• config/accounts.yaml: Account Factory catalog, OU placement, monthly budgets, and security baselines.
• config/identity.yaml: IAM Identity Center permission sets, users, groups, and multi-account assignments.

3. Validate & Inspect Readiness

Verify your configurations and audit your AWS Management Account before deploying:

# Validate YAML syntax, Pydantic schemas, and cross-references
tarmac validate

# Run automated pre-flight readiness checks (MFA, quotas, contacts, hygiene, existing trails)
tarmac preflight

4. Deploy Your Landing Zone

Preview the changes with a dry run, then deploy:

# Compare the configuration with AWS, read-only, and list what each phase would change
tarmac deploy --dry-run

# Run all six phases (the pre-flight checks run first)
tarmac deploy

Starting from a brand-new AWS account

Tarmac runs as the management account of the organization it creates, and it cannot run as the root user (root cannot assume the roles tarmac uses to configure member accounts). A fresh account therefore needs an IAM Identity Center (SSO) login first. Only the steps AWS does not allow to be automated are manual:

1. Sign in as the root user, enable MFA, and add a payment method.
2. In the console, enable IAM Identity Center (choose its Region deliberately; on an account without an organization this also creates the organization).
3. Open AWS CloudShell (it uses your console session, so no root access keys are created), install tarmac (pip install --user tarmac-cli), run tarmac init, put your own email in admin_user in identity.yaml, and run tarmac bootstrap. It creates that user with tarmac's least-privilege permission set on the management account, and a tarmac profile in ~/.aws/config (see Bootstrap access).
4. Sign in as admin_user for the first time. tarmac bootstrap prints these steps:
   1. As root, in IAM Identity Center → Users → your user, register an MFA device.
   2. aws sso login --use-device-code --profile tarmac (see Using AWS CloudShell for why the device-code flag is needed), and open the URL it prints in a private window.
   3. In the console, Reset password → Generate a one-time password for your user. Sign in with it in that window, set your own password, then enter the code.
5. export AWS_PROFILE=tarmac, then confirm with tarmac info --strict.
6. tarmac preflight, then tarmac deploy --dry-run, then tarmac deploy. Phases are idempotent, so a failed run can simply be repeated.

tarmac preflight and tarmac deploy refuse to run as root and print these steps. Each member account needs a unique email address; AWS never lets a closed account's email be reused, and tarmac validate warns about the sample addresses written by tarmac init. Gmail users can create unique addresses with the + trick (you+prod@gmail.com, you+log@gmail.com); all of them arrive in one inbox.

Using AWS CloudShell

CloudShell is the easiest place to run tarmac the first time, with a few quirks:

• Sign in with the device-code flow. aws sso login normally redirects to http://127.0.0.1:..., which is the CloudShell machine and not your browser, so the callback fails. Use aws sso login --use-device-code: it prints a URL and a short code to enter in your own browser.
• Set the profile and Region in every new shell. export AWS_PROFILE=<profile> AWS_REGION=<primary_region>. Without AWS_PROFILE, tarmac uses CloudShell's own (root) credentials and refuses to continue. CloudShell uses the Region you opened it in, which may differ from primary_region.
• Environment variables do not survive a restart; your home folder does. Re-export AWS_PROFILE, AWS_REGION and TARMAC_LOG_FILE after CloudShell restarts. Your config directory and log file persist.
• Use --log-file. export TARMAC_LOG_FILE=$HOME/run.jsonl records every run, which makes a failed phase easy to diagnose with tarmac logs. Download it with the Actions menu.

Adopting an existing organization

The quickest way to start is to let tarmac write the configuration for you:

tarmac import --config-dir my-lz          # read-only: nothing in AWS is changed
cd my-lz && tarmac validate && tarmac status --detail
tarmac deploy --dry-run

tarmac import reads the OU tree, member accounts (name, email, OU, tags), customer-managed SCPs and where they are attached, alternate contacts, trusted services, any existing organization trail, and IAM Identity Center (users, groups, permission sets with their inline policies, and group assignments). It writes organization.yaml, accounts.yaml, identity.yaml and the policy documents under policies/, sets delete_default_vpc: false everywhere, and prints a list of things to review: accounts sitting directly under the root (skipped), SCPs attached only to accounts or nowhere (not imported, because tarmac would attach them to the root), SCP attachments tarmac cannot express, a pre-existing organization trail, and placeholder contacts. organization.name comes from existing tarmac stacks when there are any (so a re-import never orphans them), otherwise from --name or the management email domain. Use --no-identity to skip Identity Center and --force to overwrite existing files.

Run tarmac deploy --dry-run before the first real run and read the Phase 4 table. The rules that keep adoption safe:

• Declare the accounts you already have, with their real emails. An existing account is matched by email first and then by name. If the name matches but the email differs, tarmac refuses to adopt or move it rather than guess.
• Accounts you do not declare are left alone, unless tarmac itself tagged them as managed earlier and you then removed them from accounts.yaml. Those are moved to the Suspended OU, and the Phase 4 plan shows them as SUSPEND. Set quarantine_removed_accounts: false in accounts.yaml to turn that off.
• Do not change organization.name after the first deploy. Every CloudFormation stack is named after it (<name>-organization-scps, ...). Renaming makes tarmac look for stacks that do not exist and try to import resources an older stack still owns, which fails. tarmac preflight warns when it sees stacks under a different name; the fix is to set the name back.
• An organization trail that already exists is not replaced. Tarmac adds its own, so you briefly pay for two copies of management events ($2.00 per 100,000 events beyond the first). Once the new trail is delivering logs (aws cloudtrail get-trail-status --name <trail> shows a delivery time), delete the old one. A multi-region trail can only be read or deleted in its home Region (describe-trails --include-shadow-trails shows it).
• Default VPC purges are opt-out per account. Set delete_default_vpc: false for accounts that already host workloads, and for the management account in organization.yaml.
• What a full run changes in an existing organization: the OU tree, the service control policy attached to the root (it affects every account, including the existing ones), account tags, a monthly budget stack in each member account, alternate contacts on the management account, trusted service access, and new Identity Center permission sets, users and group assignments. Existing permission sets and users that you did not declare are not touched.

Bootstrap access (why tarmac cannot run as root)

Tarmac configures member accounts by assuming OrganizationAccountAccessRole in each of them, and the AWS root user is not allowed to call sts:AssumeRole. (Root access keys are also discouraged.) So tarmac runs as an IAM Identity Center user, and tarmac preflight and tarmac deploy refuse to continue as root.

tarmac bootstrap, run once as root from AWS CloudShell after enabling IAM Identity Center (see Starting from a brand-new AWS account), sets that up:

• the admin_user from identity.yaml (you) in Identity Center,
• two permission sets assigned to that user on the management account, both with 4-hour sessions: TarmacDeployment, holding tarmac's policy, which tarmac runs as, and ManagementAdministrator (AdministratorAccess) for your own work in the console, so root is no longer needed,
• a tarmac profile in ~/.aws/config that signs in as TarmacDeployment.

Choose the Identity Center Region deliberately: an organization can have it in one Region only, and moving it later means deleting the instance and creating another (AWS docs). tarmac info --strict confirms the credentials are non-root and can read the organization; it exits 1 and says why when they cannot. If you already have non-root administrator credentials for the management account (an existing Identity Center login, or a CI role), you can skip tarmac bootstrap and use those.

What the permission set allows. In the management account it gets only the API calls tarmac makes: Organizations, CloudFormation, the organization trail, IAM Identity Center, GuardDuty/Security Hub delegation, Cost Explorer and a few read-only checks. For member accounts it may only assume OrganizationAccountAccessRole. It has no iam:PassRole and cannot create IAM roles or users. A test checks the policy against every AWS API call in tarmac's source, so adding a new call without updating the policy fails CI, and another test keeps it under Identity Center's inline policy limit with headroom. tarmac preflight warns when the permission set's policy differs from your installed release: re-run tarmac bootstrap after an upgrade to update it. The policy has not yet been exercised end-to-end against a real organization; if a run reports AccessDenied, report the missing action (as a stopgap, attach AdministratorAccess to the permission set in the Identity Center console). It guards against mistakes, not against its own user: anything that can manage permission sets and assume OrganizationAccountAccessRole can grant itself more.

Upgrading from 0.2 or earlier. Tarmac no longer uses the tarmac-access stack, TarmacDeploymentRole or the tarmac-bootstrap IAM user. Remove any [profile tarmac] block those left in ~/.aws/config before running tarmac bootstrap. Once you can sign in with the new profile, turn off the stack's termination protection and delete it (aws cloudformation update-termination-protection --no-enable-termination-protection --stack-name tarmac-access, then aws cloudformation delete-stack --stack-name tarmac-access); delete the tarmac-bootstrap user's access keys and MFA device first if you created them.

Bucket names now end in part of their stack's ID instead of the account ID, so a stack that is deleted and recreated never waits for S3 to free a name. Updating an existing stack replaces its buckets: CloudFormation creates new ones and keeps the old ones with their contents. Point Terraform at the new state bucket and run terraform init -migrate-state. Your existing ou_structure is unchanged; the new OU layout only applies to files that tarmac init creates.

Member accounts need nothing extra: Organizations creates OrganizationAccountAccessRole in every account tarmac creates, trusting the management account, and tarmac reaches it from there. Accounts that were invited (not created) need that role added by hand.

Deployment Pipeline

Tarmac executes an idempotent multi-phase deployment. A full tarmac deploy first runs the same checks as tarmac preflight and stops if any of them fails (--skip-preflight skips them):

[Phase 1: Org & Root]  ──► Bootstrap AWS Organizations, contacts, & root access
[Phase 2: OUs & SCPs]  ──► Idempotently sync OU tree & attach guardrail SCPs
[Phase 3: Central Log] ──► Deploy Log-Archive bucket & Org-wide CloudTrail
[Phase 4: Accounts]    ──► Provision member accounts, purge default VPCs, set budgets
[Phase 5: Hardening]   ──► Delegate Security Hub / GuardDuty & anomaly monitors
[Phase 6: Identity]    ──► Sync IAM Identity Center permission sets & assignments

You can target a specific phase at any time:

tarmac deploy --phase 4                     # Run only Account Factory provisioning
tarmac deploy --phase 4 --account dev-app   # Only this account: create or move it, tag it, apply its baseline

Phase 4 writes the account IDs to outputs/accounts.json and outputs/accounts.env (export NAME_ACCOUNT_ID=... lines you can source).

Dry runs

tarmac deploy --dry-run (optionally with --phase N and --json) changes nothing. It reads the current state with read-only calls and lists, per phase, every item and what a live run would do with it: CREATE, UPDATE, REPLACE, IMPORT (adopt an existing SCP or permission set), MOVE, SUSPEND, APPLY or NO CHANGE. Stacks are compared on their exact template text, the parameters tarmac passes and their tags, and an UPDATE names what differs. An item whose state cannot be read shows UNKNOWN with the error. A dry run needs the same credentials as a deploy; to check configuration files without AWS, use tarmac validate.

What tarmac deploys

Everything tarmac creates in AWS comes from the CloudFormation templates and SCP policies below, which ship with the package. <org> is organization.name from organization.yaml.

AWS recommends using the management account only for work that has to happen there, and running services from member accounts registered as their delegated administrators. Tarmac follows that guidance:

• Management account: Organizations (OUs, accounts, SCPs), Identity Center, delegated administrator registration, the root activity alert for the management account itself, the organization trail's stack, and the AWS Config StackSet. The StackSet is there because service-managed StackSets live in the management account even when a delegated administrator creates them.
• Audit (security tooling) account: the delegated administrator for CloudTrail, AWS Config, GuardDuty, Security Hub and IAM root access management, so your security team can manage the trail from there. The trail's stack stays in the management account: the trail belongs to that account, and CloudFormation cannot manage it from any other.
• Log archive account: the CloudTrail and AWS Config log buckets.

The starter ou_structure follows the OUs in AWS's Organizing Your AWS Environment Using Multiple Accounts: Security holds only the log archive and audit accounts; Infrastructure holds shared services and networking; Workloads splits into Prod and Test; Deployments holds CI/CD accounts such as the Terraform backend; Policy Staging is where you try SCP changes before they reach other OUs; Sandbox and Suspended are as named.

Two stacks are generated from your configuration rather than a template file: <org>-organization-scps (the SCPs declared in organization.yaml) and <org>-sso-permission-sets (the permission sets in identity.yaml). tarmac bootstrap creates the TarmacDeployment permission set with tarmac-deployment.json as its policy.

To review or change any of this before it reaches AWS:

• tarmac init --with-templates copies the templates and SCPs into your project (cloudformation/, policies/scp/). Tarmac uses your copies in preference to the bundled ones, so you can keep them in version control and review changes.
• tarmac deploy --dry-run compares every stack with what is deployed (template, parameters and tags) and lists what each phase would change, without changing anything.

Command Reference

After tarmac bootstrap, everything that changes AWS goes through tarmac deploy, one phase at a time if you like:

Example: restricting regions

Tarmac does not choose which regions you may use, so no region SCP is attached by default. examples/policies/scp/deny-unused-regions.json is a starting point you can copy into your project's policies/scp/ and reference from organization.yaml:

1. Replace the aws:RequestedRegion list (us-east-1, us-west-2 in the example) with the regions you use. Include your primary_region and the region of your IAM Identity Center instance.
2. Keep the NotAction list of global services (IAM, Organizations, Route 53, CloudFront, Support, STS, ...), or those calls are denied everywhere.
3. Keep the OrganizationAccountAccessRole exemption. Tarmac removes default VPCs in every enabled region of each member account through that role, and Phase 2 (SCPs) runs before the Phase 4 baseline, so without the exemption that cleanup fails in the regions you blocked. Drop it only if you set delete_default_vpc: false (or have already cleaned those VPCs).

SCPs do not apply to the management account. Test it on one OU before attaching it to the root.

Example: protecting GuardDuty and Security Hub

Nothing stops an administrator of a member account from turning off GuardDuty or Security Hub there, so tarmac does not attach a protecting SCP by default (an SCP can lock you out; see above). examples/policies/scp/protect-security-services.json denies only the actions that switch them off or detach the account from the audit account: deleting the detector, disassociating from the administrator, stopping member monitoring, disabling Security Hub and its standards. Reading and configuring stay allowed. The action names are taken from the AWS Service Authorization Reference (GuardDuty, Security Hub).

The policy exempts OrganizationAccountAccessRole, which is how tarmac reaches member accounts, so tarmac destroy --disable-security-services keeps working. Add the role your own administrators use (for example an AWSReservedSSO_* role) to the aws:PrincipalARN list, or they cannot turn the services off either. It is not run against a live organization; attach it to one test OU first.

AWS Config is left out on purpose. AWS publishes an example SCP for Config (config:DeleteConfigurationRecorder, StopConfigurationRecorder, DeleteDeliveryChannel, DeleteConfigRule), but tarmac deploys recorders with StackSets, and I could not confirm which role name StackSets uses in member accounts, so a Config SCP could block its own updates.

Logging and Audit Trail

Each line of the log file is one JSON object with a timestamp, level, a run_id shared by the whole invocation, and (where known) the command, phase, account and Region. The first record (run_start) holds the version and arguments, and the last (run_end) holds the exit code and duration. Records with "event": "audit" describe a single action: what was done, to which resource, and whether it succeeded, was skipped, was only simulated by --dry-run, or failed. Secrets such as access keys and tokens are redacted before anything is written, and the file is created readable by its owner only.

tarmac --log-file run.jsonl run --yes
jq -r 'select(.event=="audit") | "\(.ts) \(.data.outcome) \(.message)"' run.jsonl

Outcomes mean what they say: succeeded is a change tarmac made, unchanged means AWS already held the requested state (tarmac reads before it writes, so a second run of the same command should be almost all unchanged), skipped was deliberately not done, simulated is what --dry-run would have done, and failed is an error. To prove a dry run changed nothing, check that the log has no dry_run_violation records: tarmac logs run.jsonl --run latest --all --json | grep dry_run_violation should print nothing. tarmac deploy --dry-run records every item of the plan, with the action a live run would take, as simulated.

Reading a log back

tarmac logs turns a log file into something a person can review without jq:

tarmac logs run.jsonl                       # one line per run: status, what changed, what failed
tarmac logs run.jsonl --run latest          # what the last run did, plus any warnings and errors
tarmac logs run.jsonl --run 20261020T1      # a run by id or id prefix
tarmac logs run.jsonl --outcome failed      # only actions that failed
tarmac logs run.jsonl --account dev --phase accounts
tarmac logs run.jsonl --run latest --all    # every record, including AWS calls
tarmac logs run.jsonl --run latest --json   # matching records, one JSON object per line

With TARMAC_LOG_FILE set, the file argument can be left out. Reading a log never writes to it. A run shows as incomplete if the process was killed before it could write its final record.

AWS calls and CloudTrail

Every AWS call is recorded with its service, operation, Region, outcome, duration, request id and a redacted summary of its parameters. Calls that change something (Create*, Put*, Delete*, Enable*, ...) are logged at INFO, so they are in the log file and shown by -v; read-only calls (Describe*, List*, Get*) appear with --debug.

The run id is added to the AWS user agent (app/tarmac-<run-id>) and to the role session name (tarmac-<run-id>), so any CloudTrail event can be matched back to the run that caused it, for example by searching for the run id in userAgent or in the assumed-role session name.

During a --dry-run, any state-changing call is logged as an ERROR with "event": "dry_run_violation": a preview must not change anything.

Keeping CloudTrail logs

By default CloudTrail logs are kept forever, which is the safe choice but means storage cost grows slowly. Set cloudtrail.retention_days to delete log files that many days after they are written:

cloudtrail:
  retention_days: 365   # omit to keep logs indefinitely

Deleted objects are purged for good seven days later and cannot be recovered, so pick a value that meets your audit and compliance requirements (many frameworks ask for a year or more). tarmac deploy --dry-run shows the retention that will be applied. The setting is applied to the log archive stack; changing it re-deploys that stack.

Locking CloudTrail logs (S3 Object Lock)

Recommended, but off unless you set it. Object Lock gives the log bucket a default retention so log files cannot be deleted or overwritten for that period, which is what makes the audit trail trustworthy. tarmac does not turn it on by itself because S3 cannot disable Object Lock on a bucket once it is on. While it is unset, deploy and baseline print a warning and the dry run shows it as off.

cloudtrail:
  object_lock:
    mode: GOVERNANCE      # or COMPLIANCE
    retention_days: 365
  # object_lock: null     # decided against it: hides the warning

GOVERNANCE can be bypassed by principals with s3:BypassGovernanceRetention. COMPLIANCE cannot be shortened or removed by anyone, including the root user, until each object's period ends. The lock applies to objects written after it is enabled, and can be added to an existing log bucket later by setting it and re-running deploy. A lock longer than cloudtrail.retention_days (log expiry) is rejected, because lifecycle cannot delete locked versions.

This changes teardown. tarmac destroy retries locked deletes with the governance bypass, so GOVERNANCE locks do not block it (the audit log records it). COMPLIANCE locks cannot be removed: destroy stops with an error and the log bucket stays until the retention period ends. Use GOVERNANCE in test organizations.

AWS Config recorders (optional)

Off by default: AWS Config bills per configuration item, so the cost grows with accounts x Regions x resource changes. Turn it on to record resource configuration in every account:

security_services:
  config_recorder:
    enabled: true
    include_global_resources: true   # IAM and similar types, recorded in the primary region only
    recording_frequency: CONTINUOUS  # DAILY is $0.012/item vs $0.003: cheaper only above 4 changes a day
  # config_recorder: null            # decided against it: hides the Security Hub warning below

Recorders are created in primary_region plus additional_regions, the same Regions GuardDuty and Security Hub use. They record every supported type except AWS::Config::ResourceCompliance: Security Hub updates that item on every compliance change, AWS says its checks do not need it, and it recommends not recording it when the recorder serves only Security Hub. Set record_resource_compliance: true if you use that history, or if a Region where IAM global types are not supported rejects the recorder.

DAILY is billed at $0.012 per configuration item against $0.003 for CONTINUOUS, and records an item only when a resource differs from its previous one. It therefore saves money only for resources that change more than four times a day; for most accounts CONTINUOUS is the cheaper choice. Firewall Manager needs CONTINUOUS. Member accounts get them from one service-managed StackSet (<org>-config-recorder) targeting the organization root, so accounts created later receive the recorder without re-running tarmac. StackSets cannot target the management account, so it gets the same template as an ordinary stack in each Region, using AWS Config's service-linked role (the deployment role is not allowed to create IAM roles). Suspended accounts are left out.

History and snapshots go to their own bucket in the log archive account (<org>-config-logs-<8-character stack ID>), created by the log archive stack in phase 3 when this is enabled. It is separate from the CloudTrail bucket because AWS Config cannot deliver to a bucket with Object Lock default retention, so Config history is not locked. It uses the same KMS key setting and log expiry (cloudtrail.retention_days) as the CloudTrail bucket. Enabling the setting later means running phase 3 again before phase 5.

Security Hub CSPM standards need a recorder to evaluate most of their controls (unless the newer AWS Security Hub is also enabled, which manages its own). While Security Hub is enabled and config_recorder is not mentioned in your configuration, deploy warns; write config_recorder: null (or enabled: false) to acknowledge it.

cloudtrail.object_lock: null works the same way: both warnings are silenced by the key being present in organization.yaml, which tarmac never writes for you (neither tarmac init nor tarmac import adds it). If you generate that file with another tool, keep the key, or the warning comes back.

Notes:

• An account and Region can have only one customer-managed recorder. If one already exists (for example from Control Tower), that instance fails and the run reports it; tarmac does not replace recorders it did not create, and it never updates a StackSet that lacks its tarmac:managed-by tag. Service-linked recorders (such as Security Hub's) coexist.
• Re-running with fewer Regions removes the recorders from the dropped Regions.
• tarmac destroy removes the StackSet, its instances and the management stacks before it touches the log buckets. It leaves StackSets trusted access enabled.

Security Hub standards for every account

security_services:
  delegated_account: audit
  securityhub:
    enabled: true
    standards: [aws-foundational-security-best-practices]
    central_configuration: true   # default; false = per-Region local configuration

With central_configuration (the default, used when auto_enable_members is true) the audit account switches Security Hub to central configuration and applies one policy, tarmac-security-baseline, to the organization root. Existing and new accounts then get Security Hub with exactly the standards you list, in the primary Region (the aggregation Region) and the Regions linked to the finding aggregator. The old local mode left new accounts with Security Hub on but no standards.

The linked Regions are exactly the Regions you declare (the primary Region aggregates; the rest are linked, or none for a single-Region deployment). A central policy turns Security Hub on in every linked Region of every account, so tarmac never links "all Regions" in this mode. An existing aggregator is narrowed to match.

Switching an existing organization is guarded. tarmac stops, rather than guess, when Security Hub already has member accounts configured one by one, or when the root already uses a configuration policy tarmac did not create, because a policy can replace what those accounts enabled themselves. Review the difference and re-run with --adopt-existing, or set central_configuration: false to keep local configuration. The home Region (organization.primary_region) cannot be an opt-in Region such as af-south-1 or me-south-1, because Security Hub does not allow it; tarmac checks this and stops before changing anything. This behaviour has not been verified against a live organization yet.

Notes: policies can only be managed from the primary Region, which must not be an opt-in Region. In central mode the per-account settings are managed by the policy, so change standards in organization.yaml rather than by hand. tarmac destroy --disable-security-services removes the policy and returns the organization to local configuration before disabling Security Hub.

Encryption of the log archive

CloudTrail encrypts delivered logs with SSE-S3 (AES256) by default, and AWS describes a customer managed KMS key as "an added layer of security". The key costs about $1 a month plus request charges, so Tarmac leaves it off for a new deployment.

• cloudtrail.enable_kms unset (the default): a new deployment uses SSE-S3. A deployment that already has the key keeps it, so upgrading never lowers its encryption.
• enable_kms: true adds the key. Set it if you follow the CIS benchmark or Security Hub, whose CloudTrail checks (CloudTrail.2, CIS 3.5) ask for SSE-KMS.
• enable_kms: false switches an existing deployment back to SSE-S3 for new logs. The key already created is kept (logs written earlier still need it to be read); remove it yourself once those logs have expired.
• If Tarmac cannot read the existing log archive stack to find out, it stops and says so instead of guessing.

Existing resources with the same name

Tarmac will not overwrite or delete something it did not create. If an SCP or a budget it would manage already exists with the same name:

• SCP that is not tagged tarmac:managed-by=tarmac and whose content differs from your policy file: the run stops and lists the live and declared statement IDs. (An untagged SCP that already matches is taken over quietly, which is what happens after tarmac import.)
• Budget named <account>-monthly-budget that tarmac did not create: the account's budget step fails with its current limit. Tarmac would otherwise delete it and create its own.

After reviewing, re-run with --adopt-existing (on tarmac deploy) to let tarmac take it over. The old budget's limit and period are recorded in the audit log.

GuardDuty features (all off unless you ask)

security_services.guardduty turns on GuardDuty's foundational detection only. The optional protection plans are billed on top of it, so they stay off until you set them:

security_services:
  guardduty:
    enabled: true
    features:
      s3_data_events: true      # S3 Protection
      # eks_audit_logs, malware_protection (EBS), rds_login_events, lambda_network_logs

AWS switches on every optional feature (except Runtime Monitoring) when a detector is created without a feature list (CreateDetector), so tarmac sends the list explicitly, and sets the same choice for new member accounts. On a detector that already exists tarmac only turns requested features on; it never turns one off, and it reports any optional feature that is on but not requested. Runtime Monitoring is not managed by tarmac.

Terraform state locking

Phase 5 creates a versioned, encrypted S3 bucket for Terraform state, named <org>-terraform-state-<8-character stack ID>; the backend stack's BucketName output has the exact name. Locking is S3-only by default, so Terraform needs use_lockfile = true (Terraform 1.10 or later) and there is no DynamoDB table to pay for. Terraform now marks DynamoDB-based locking as deprecated (S3 backend). Set terraform_backend.enable_dynamodb: true to create the lock table anyway. A backend that already has the table keeps it unless you set enable_dynamodb: false, and tarmac warns before removing it, because Terraform configurations that still set dynamodb_table would stop locking.

Safe Teardown

Tarmac includes built-in safeguards to prevent accidental infrastructure or state loss:

• Production Environment Lock: Refuses to destroy any environment whose Environment/Env/Stage tag is production, prod or prd unless --force is explicitly provided.
• Ownership checks: SCPs, the standalone CloudTrail trail and Identity Center permission sets that tarmac did not create (no tarmac:managed-by=tarmac tag) are skipped, not deleted.
• Remote State Preservation: Pass --preserve-backend to destroy infrastructure while protecting remote Terraform state buckets and lock tables.
• Non-Empty Bucket Protection: Halts if unmanaged data exists in S3 buckets unless explicitly confirmed.

# Preview teardown actions
tarmac destroy --dry-run

# Execute teardown preserving Terraform state
tarmac destroy --preserve-backend

Termination protection

The stacks that hold data or access are created with CloudFormation termination protection: the organization trail, the log-archive and audit-account hardening stacks and the Terraform state backend. A stray aws cloudformation delete-stack is refused for them. tarmac destroy is the deliberate way to remove them: it turns protection off on each stack just before deleting it and prints a warning saying so (and --dry-run shows what it would delete). --preserve-backend leaves the state backend, and its protection, untouched. Enabling protection needs cloudformation:UpdateTerminationProtection, so re-run tarmac bootstrap once to refresh the permission set; until then a deploy reports that protection could not be turned on and carries on.

Root user alerts

SCPs do not apply to the management account, so the deny-root SCP cannot stop root being used there. Tarmac deploys a <org>-management-root-alert stack that emails contacts.security.email (and posts to security_services.incident_response_webhook_url, if set) when the root user signs in or makes an API call in the management account. It is on by default and costs next to nothing: EventBridge ingests AWS management events for free, and the only charge is a few SNS emails. AWS records root sign-in events in us-east-1, so the stack is deployed there and in your primary Region; each Region has its own SNS topic, so confirm the subscription email in each. Root use in member accounts is denied by the SCP, and the attempts are in the organization trail; the audit account also alerts on its own root activity.

security_services:
  root_alert:
    enabled: true        # false turns the alert off
    all_regions: false   # true also watches every declared Region for root API calls

Without a security contact email or webhook there is nobody to alert, so tarmac skips the stack and says so. The permission set needs the events: and sns: actions this adds: re-run tarmac bootstrap once.

Starting over

A full tarmac destroy removes everything tarmac deployed (stacks, buckets, budgets, SCPs, permission sets, milestone tags) so tarmac deploy can rebuild it. It deliberately keeps:

• the organization, OUs and member accounts (closing an account is irreversible and AWS reserves its email, so use tarmac account suspend --close per account and declare a new email if you want a truly fresh account);
• GuardDuty / Security Hub / AWS Config delegation and member detectors, which keep billing until disabled. Pass --disable-security-services to turn them off as part of Phase 5 (it turns off auto-enable, deregisters the delegated administrators, deletes the GuardDuty detectors and disables Security Hub in the audit, management and declared member accounts). This deletes finding history, so it is opt-in. Failures per account/region are reported, and a failed step makes the command exit non-zero;
• IAM Identity Center users and groups, OUs, organization settings and the permission sets from tarmac bootstrap (kept unless you pass --everything, below).

The log-archive KMS key is scheduled for deletion (30-day window, cancel with aws kms cancel-key-deletion) once its buckets have been purged and deleted; if any log-archive cleanup failed the key is kept.

Removing everything tarmac did

tarmac destroy --everything --dry-run     # preview
tarmac destroy --everything               # asks you to type the organization name

On top of a normal destroy, --everything removes, in dependency order: the Identity Center users and groups tarmac created (and their assignments), the OUs tarmac created (only when empty), the organization settings tarmac enabled (delegated administrators, root credentials management and root sessions, trusted service access, alternate contacts, cost-allocation-tag activation), the milestone tags, and finally the TarmacDeployment and ManagementAdministrator permission sets from tarmac bootstrap.

What counts as "created by tarmac" is decided by evidence, never by name:

• Stacks, SCPs, permission sets and accounts carry the tarmac:managed-by=tarmac tag.
• OUs, users, groups and organization settings cannot be tagged, so tarmac records each one in <config dir>/.tarmac/inventory.json the moment it changes it. Things that already existed (an event logged as "unchanged") are never recorded and therefore never removed. tarmac inventory shows the record.
• For a deployment that predates the inventory, rebuild it from a run log: tarmac destroy --everything --inventory-from-log "$TARMAC_LOG_FILE" (or tarmac inventory --from-log FILE to inspect it first).

Member accounts are never closed by this command. It lists the accounts tarmac manages with the exact tarmac account suspend NAME --close command for each. An alternate contact that replaced one which already existed is left in place (tarmac cannot restore the old one). Disabling Identity Center trusted access is also left to you, because it can delete the Identity Center instance. The permission set the teardown itself is signed in with is kept (deleting it would end the session), and the command tells you to delete it in the Identity Center console.

If a teardown leaves a CloudFormation stack in DELETE_FAILED (for example a bucket that could not be emptied), fix the cause and re-run teardown or deploy: both now delete and recreate stacks stuck in DELETE_FAILED, CREATE_FAILED or ROLLBACK_COMPLETE.

Templates exported with tarmac init --with-templates take precedence over the ones bundled with the CLI. After upgrading tarmac, delete or diff your local cloudformation/ and policies/ copies against the bundled ones to pick up fixes. (Do not use tarmac init --force for this: it also overwrites your YAML configuration.)

Troubleshooting

Always start with tarmac logs --run latest: it shows the failing action, its AWS error and the phase it happened in.

Development & Testing

# Sync environment dependencies from lockfile
uv sync --locked

# Run the Pytest suite
uv run pytest

# Run static analysis, formatting, and template/config validation
uv run ruff check .
uv run ruff format --check .
uv run basedpyright
uv run cfn-lint src/tarmac/templates/cloudformation/*.yaml
uv run tarmac validate --config-dir examples/config

License

Licensed under the Apache License, Version 2.0.