AI infrastructure
Terraform from a prompt.
Describe the change in a sentence. Datis writes the Terraform, prices the change and runs your policy checks — all before a single resource exists. Nothing is applied without an approval.
AI infrastructure here means infrastructure the agent writes and your policy gates — not hosting, not model serving. Your cloud account stays yours.
What comes back is a plan you can read.
You describe the infrastructure in a prompt. What comes back is not a resource — it is the Terraform files, the monthly cost delta and the result of every policy check, written down while nothing has been created yet.
Watch it plan a cache and a queue on the landingThe Terraform Datis wrote
Real files, written from your sentence — not a diagram, not a wizard. A plan is a file and a cost estimate: you review it the way you review any change.
Priced before it runs
Every plan carries its monthly cost delta, on top of what the account already spends — so the cost is part of the review, not a discovery on the invoice.
Checked before it exists
Datis scans what it writes: your policy checks and security findings land on the plan itself, each one under the file it fired on, before a single resource exists.
Nothing is applied without an approval.
Applying is a separate act with its own floor — rules the agent may never cross, not settings to go and find. Your policy can only make them stricter.
An approval the policy names
Applying infrastructure without a named approver is on the policy floor. Apply exists only where an Admin granted it, and only with the approval the policy requires — whether that is a named person or an earned per-class grant is the policy’s own setting.
Bound to one plan
The approval is bound to one plan hash — regenerate the plan and the approval is void. What runs is exactly what was read, or nothing runs.
Two nevers
Never in an environment marked protected. Never where the approver is the account that requested the run. Neither is a setting — they cannot be turned off.
Read by default. Plan by grant. Apply by policy.
Every connection is created read-only — cloud accounts included. Widening a mode is an Admin’s act, done per connection after re-authenticating, and the change lands in the audit log with actor, time and IP.
Read only
The default, and nobody has to turn it on. On a cloud account that means describe and list — no apply, no branch, no commit. The grant is enforced at the credential you create, not in the product, so a bug in Datis cannot widen it.
Plans, never applies
Opt-in, by an Admin, per connection. Everything in Read, plus write a Terraform plan and push a branch carrying a draft pull request. Nothing is applied and nothing is merged — a plan is a file and a cost estimate.
Applies with approval
Opt-in and gated: the policy must name at least one approver before the mode exists. Everything in Plan, plus apply a plan — with the approval the policy requires, only in the environments it lists, never in one marked protected, never self-approved.
The ladder is not a Build-only setting — it is how the whole product earns autonomy: each rung is dialled up as the track record earns it, and down the moment it stops clearing the bar. The run the landing replays went from the first bad signal to the merged fix in 4m 13s, measured end to end, because a rule you can read allowed every step of it.
Describe the change. Read the plan. Approve it — or don’t.
The beta is where this runs against your account: connected read-only, widened only by your Admin, applied only with the approval your policy requires. Every request on the waitlist is read by a person.