Terraform Provider Anthropic Claude Managed Agents

terraform-provider-anthropic-claude-managed-agents

⚠️ Not for production use — learning & prototyping only

This is an experimental, community-built project intended solely for learning, experimentation, and prototyping. It is NOT intended, tested, or supported for production use.

It is provided "AS IS", without warranty of any kind, express or implied. You use it entirely at your own risk. The author(s) and contributors accept no liability for any damage, data loss, downtime, cost, security incident, or other harm arising from its use. The "1.0.0" version number denotes API/schema stability under SemVer — it is not a statement of production-readiness or fitness for any purpose.

See the LICENSE (MPL 2.0, sections 6 & 7) for the full warranty disclaimer and limitation of liability.

A Terraform / OpenTofu provider for Anthropic Claude Managed Agents — declaratively manage agents (with tools, MCP servers, skills, and multiagent coordination), environments, vaults, custom skills, and memory stores from Terraform configurations.

Unofficial. This is a community project. It is not maintained by, endorsed by, or affiliated with Anthropic. Upstream API: Claude Managed Agents docs.

Anthropic Terraform provider — what this covers

A community-maintained Terraform / OpenTofu provider for the Anthropic Claude Managed Agents API. Resources and data sources:

• Agents — with nested HCL for tools (agent_toolset_20260401, mcp_toolset, custom), mcp_servers, skills, and multiagent coordination
• Environments — cloud and dedicated, with apt / npm / pip package install
• Vaults and vault credentials — secrets handled via Terraform 1.11 write-only attributes (not persisted to state)
• Custom skills — multipart upload, content-hash drift detection, immutable versions
• Memory stores — persistent agent memory, attached at session time

The Managed Agents API is distinct from the standard Anthropic Messages API. This provider does not manage the Messages API or generic /v1/messages calls.

What this provider does

Claude Managed Agents is Anthropic's hosted runtime for long-running, tool-using agents. The platform exposes a REST API for creating and configuring those agents and their supporting resources (environments, vaults, memory stores).

This provider lets you put that configuration under Terraform so you can:

• Review changes in pull requests instead of via ad-hoc API calls.
• Version-control your agent definitions alongside the rest of your infra.
• Promote agents across workspaces with the same workflow you use for everything else.

Status / scope

For per-release change history (including deprecations and breaking changes), see CHANGELOG.md.

Quickstart

terraform {
  required_version = ">= 1.11"

  required_providers {
    claude-managed-agents = {
      source  = "modus-agendi/anthropic-claude-managed-agents"
      version = "~> 1.0"
    }
  }
}

Don't rename claude-managed-agents to something shorter. That string is the prefix on every resource type (claude-managed-agents_agent, claude-managed-agents_vault, …). If you alias it (e.g. to anthropic), every resource block must add an explicit provider = claude-managed-agents argument — noisier than keeping the literal name.

provider "claude-managed-agents" {
  # api_key defaults to the ANTHROPIC_API_KEY environment variable.
}

resource "claude-managed-agents_agent" "coding_assistant" {
  name        = "Coding Assistant"
  model       = "claude-opus-4-7"
  system      = "You are a helpful coding agent."
  description = "Pairs on Go and Terraform tasks."

  metadata = {
    team = "platform"
  }
}

output "agent_id" {
  value = claude-managed-agents_agent.coding_assistant.id
}

export ANTHROPIC_API_KEY="sk-ant-..."
terraform init
terraform apply

Authentication

The provider reads credentials from, in order of precedence:

1. The api_key argument on the provider block.
2. The ANTHROPIC_API_KEY environment variable.

api_key is marked Sensitive in the schema, so it is redacted from plan/apply output. Provider configuration is not written to Terraform state, so the key never lands in the state file regardless of how you supply it. The risk of hardcoding it in the provider block is that it then lives in your configuration files (and version control), and can surface in TF_LOG debug or crash logs — so prefer the ANTHROPIC_API_KEY environment variable in production and CI.

claude-managed-agents_vault_credential uses Terraform 1.11 write-only attributes for its secret fields (token, access_token, refresh_token, client_secret): these values are never persisted to state. Rotate by incrementing the matching *_wo_version integer.

provider "claude-managed-agents" {
  api_key     = var.anthropic_api_key   # only set this if env var is impractical
  base_url    = "https://api.anthropic.com"
  max_retries = 3
}

Resources and data sources

Guides

• Deployments: scheduling and running agents — scheduled vs manual runs, the desired_status/status split, mounting resources, reading run history, and the caveats to know before relying on deployments.
• State security and remote backends — what is and isn't in state, and encrypted/locked backends (S3+KMS, GCS, Terraform Cloud).
• Secrets management and rotation — write-only attributes and the rotation ceremony for vault credentials and deployment github tokens.
• Import and state recovery — terraform import per resource, the composite vault-credential id, and re-supplying write-only fields after import.
• Drift detection and remediation — plan -refresh-only -detailed-exitcode, the provider's expected-drift patterns, and ignore_changes guidance.
• Upgrading and migration — version constraints, the registry slug/namespace history, and a safe upgrade procedure.

Lifecycle gotchas

• Agent destroy maps to archive. The upstream API has no DELETE /v1/agents/{id} endpoint. terraform destroy issues POST /v1/agents/{id}/archive. Archived agents are read-only and cannot be unarchived.
• version is server-managed. The provider passes it along on update for optimistic concurrency. If you see version conflicts in plans, refresh the state with terraform apply -refresh-only.
• metadata is key-level merged. The upstream API uses merge semantics: removing a key from your HCL causes the provider to send JSON null for that key, which the API treats as a delete. Applies to both claude-managed-agents_agent.metadata and claude-managed-agents_vault.metadata.
• Environments are immutable. The API has no environment update endpoint, so every attribute on claude-managed-agents_environment is ForceNew. Any change triggers replacement. Destroy issues DELETE /v1/environments/{id} and falls back to archive if the API returns 409 (active session reference).
• Vault and memory_store destroy archives by default. Set delete_on_destroy = true on the resource to hard-delete instead. Vault archive cascades through credentials; memory_store hard-delete cascades through memories and memory versions.
• Deployment pause/resume is driven by desired_status, not status. Set desired_status to active or paused; the provider reconciles via the pause/resume endpoints. desired_status is your intent and is deliberately decoupled from the observed status: when the API auto-pauses a deployment on error, status becomes paused while desired_status stays active, and Terraform will not fight it (no resume flap). Inspect paused_reason, fix the cause, then re-apply to resume. Editing initial_events forces replacement (the API does not patch events in place). Destroy archives the deployment (one-way; no DELETE endpoint). The github authorization_token in resources is write-only — bump authorization_token_wo_version to re-send it on rotation.
• Vault-credential secrets are write-only. Secret fields are never read back from the API. Rotate by incrementing the corresponding token_wo_version, access_token_wo_version, refresh_token_wo_version, or client_secret_wo_version integer. auth.type, auth.mcp_server_url, auth.refresh.token_endpoint, and auth.refresh.client_id are immutable (RequiresReplace).

OpenTofu compatibility

The provider builds on protocol v6 and uses the Plugin Framework. It requires Terraform 1.11+ or OpenTofu 1.8+ — both engines support TF 1.11 write-only attributes, which the provider uses to handle secrets in claude-managed-agents_vault_credential.

Local development

Pre-requisites: Go 1.23+, Terraform 1.11+ (or OpenTofu 1.8+), golangci-lint.

git clone https://github.com/modus-agendi/terraform-provider-anthropic-claude-managed-agents
cd terraform-provider-anthropic-claude-managed-agents
go mod download

make test            # unit tests
make testacc         # acceptance tests, httptest fixture (free)
make testacc-live    # acceptance tests against api.anthropic.com (requires ANTHROPIC_API_KEY)

# Behavioral scenarios (L5) — opens real sessions and grades them via an
# LLM-as-judge call. Bills real inference tokens; see
# internal/scenarios/README.md for cost notes.
TF_ACC_SCENARIOS=1 go test ./internal/scenarios/... -count=1 -timeout 10m -v

make lint
make docs            # regenerate docs/ via tfplugindocs
make coverage-html   # full coverage profile + HTML report
make coverage-split  # mirrors CI: split unit + acceptance profiles
make pr              # full local PR-check pipeline (mirrors CI)

make install         # build + install into ~/.terraform.d/plugins/...

Once installed locally, pin the dev version in your test main.tf:

terraform {
  required_providers {
    claude-managed-agents = {
      source  = "modus-agendi/anthropic-claude-managed-agents"
      version = "0.0.1-dev"
    }
  }
}

Gotcha — make install shadows registry releases. make install copies the binary into ~/.terraform.d/plugins/registry.terraform.io/modus-agendi/anthropic-claude-managed-agents/, which Terraform treats as the authoritative version source. Any external config that pins a different version (e.g. version = "~> 0.2") will fail terraform init with "no available releases match the given constraints". Run rm -rf ~/.terraform.d/plugins/registry.terraform.io/modus-agendi to drop back to registry-pulled releases.

Testing & quality

Testing runs in five layers so that cheap, deterministic checks gate every PR while the expensive, real-API checks run on a schedule. Full details (and the exact CI triggers) are in CONTRIBUTING.md.

L5 is the unusual one and worth calling out for transparency. It provisions resources via Terraform, opens a real session, captures the full event trajectory, asserts deterministic properties on it, then has a separate judge model grade the final answer against a rubric. It bills real inference tokens — roughly $0.05–$0.20 per scenario (the current catalog of six runs at about $0.68 total). It is a hard gate on releases: a scenario FAIL blocks goreleaser unless explicitly skipped. See internal/scenarios/README.md for the scenario format, the deployment/lifecycle scenario kinds, and a note on the undocumented manual-run endpoint the harness relies on.

Coverage target is 85% (codecov gates at 70% project / 75% patch). Coverage is treated as a gap-finder, not a goal — see CLAUDE.md.

Versioning

The provider follows Semantic Versioning. As of v1.0.0 the resource and data-source schema is stable: breaking changes happen only in a new major version (2.0.0), never in a minor or patch release. Pin version = "~> 1.0" to receive non-breaking 1.x updates automatically. Breaking changes, when a 2.0 eventually happens, will be called out in CHANGELOG.md with a migration note.

A breaking change is one of:

• Removing or renaming a schema attribute.
• Changing the type of a schema attribute.
• Changing the required/optional/computed status of an attribute in a way that affects existing configurations.
• Changing the default value of an attribute.

Roadmap

Under consideration for the next minor: an ephemeral resource for vault-credential validation, optional skill version retention attributes, and broadening the L5 behavioral scenario catalog. Sessions, dreams, memory contents/versions, and webhook endpoints are not currently on the roadmap. Open an issue if you'd like to discuss.

Contributing

See CONTRIBUTING.md.

Community

• SECURITY.md — how to report vulnerabilities privately.
• CODE_OF_CONDUCT.md — expected conduct for contributors and maintainers.
• Issue templates — structured bug reports and feature requests.
• Pull request template — checklist for opening a PR.
• CODEOWNERS — review routing.

License

Distributed under the Mozilla Public License 2.0. See LICENSE for the full text (add the file via the GitHub UI's license picker if it's missing from a fresh checkout).

Disclaimer

This is a community-maintained project. It is not affiliated with Anthropic. "Claude" is a trademark of Anthropic; usage here is descriptive, identifying the API this provider integrates with.