AI-Operable Control Planes for Cloud Infrastructure (2026)
The short answer
An AI-operable control plane lets an AI agent change cloud infrastructure through the same governance a human change goes through: policy-as-code, permissions, approvals, state and audit. As of October 12, 2026, the companies offering one are:
- Pulumi (Pulumi Neo) — the most complete agent: an infrastructure agent inside Pulumi Cloud, running on Anthropic’s Claude via Amazon Bedrock, with Plan Mode, approval-based task modes, a read-only mode and scheduled autonomous work. Metered at $3 per million Neo tokens.
- Spacelift (Spacelift Intent) — best bring-your-own-agent control plane: your MCP client (Claude Code, ChatGPT, VS Code, Cursor) calls Spacelift’s
/mcpendpoint, and OPA/Rego policies are evaluated before execution. The Intent engine is open source under Apache-2.0. - HashiCorp (Terraform MCP server) — best for existing Terraform estates: an open-source MCP server for the Terraform Registry and HCP Terraform / Terraform Enterprise APIs, generally available since June 2026.
- AWS (AWS MCP servers) — the open-source
awslabs/mcpcollection gives agents tool access to AWS services; governance comes from IAM, not a separate control plane.
For governing the agents themselves (identity, lifecycle, kill switches) rather than the infrastructure they touch, see the hub page: best AI agent control planes.
The comparison
| Pulumi Neo | Spacelift Intent | Terraform MCP server | AWS MCP servers | |
|---|---|---|---|---|
| What it is | Infrastructure agent in Pulumi Cloud | Natural-language provisioning via MCP, under Spacelift governance | MCP server for Terraform Registry + HCP Terraform | MCP servers for AWS services |
| Which AI | Pulumi’s own agent (Claude via Bedrock); BYOK on Enterprise | Any MCP client | Any MCP client (VS Code, Cursor, Claude Code, Codex CLI, Gemini, Bob) | Any MCP client |
| Guardrails | Pulumi policies, RBAC, Plan Mode, approvals, read-only mode | OPA/Rego before execution, spaces + RBAC, state, audit history | Existing HCP Terraform auth and workspace permissions; tool filtering | IAM permissions of the credentials you give it |
| IaC engines | Pulumi (any language) | OpenTofu/Terraform providers; export to code | Terraform | None (direct API) |
| Price (Oct 12, 2026) | Free (1 user) · Essentials $40/mo · Pro $400/mo · Enterprise $2,000/mo; Neo tokens $3/M | Free plan (2 users); Starter+, Business, Enterprise by subscription/quote | Open source; HCP Terraform billed separately | Open source; AWS usage billed normally |
| Open source | No | Intent engine, Apache-2.0 | Yes | Yes |
Picks by situation
You want an agent that does platform-engineering work, not just writes code: Pulumi Neo. It reads your organization’s live state in Pulumi Cloud and can answer questions, investigate drift, review pull requests and run scheduled tasks. Pulumi’s docs are explicit about autonomy controls: Plan Mode agrees on a plan before acting, task modes decide which steps need approval, and read-only mode removes the ability to change anything. “Neo never has more access than you do, only less.” Scheduled Neo tasks need the Pro plan.
You want your developers’ existing AI tools to provision safely: Spacelift Intent. Instead of writing Terraform first, the client describes the infrastructure; Intent discovers providers from the OpenTofu registry, learns their schemas, plans the operation, and attached policies allow or deny it on resource type, operation or attribute. Approved changes run through cloud integrations with scoped credentials, and state and operation history are recorded automatically. Anything prototyped this way can be promoted into a normal IaC workflow.
You already run HCP Terraform or Terraform Enterprise: the Terraform MCP server. The agent stays inside the workspaces, variables and runs your team already governs, and grounds itself in real provider schemas instead of guessing resource arguments. It supports stdio and StreamableHTTP transports and token passthrough for centralised deployments.
You are all-in on AWS and accept IAM as the guardrail: AWS MCP servers. Fastest to try, but there is no plan-and-approve layer of its own — give the agent read-only roles unless a human reviews every change.
What every setup needs, whichever you choose
- Separate the policy author from the operator. The agent must not be able to edit the policies that constrain it (Spacelift spaces and RBAC exist for exactly this).
- Plan, then approve, for production. Free iteration in sandboxes; human approval on prod.
- Short-lived, scoped cloud credentials per task — see short-lived credentials for AI agents.
- One audit trail that attributes each change to the agent and the person who asked.
- Drift detection so out-of-band agent changes are caught.
Last verified: October 12, 2026.