Your agent never holds an AWS key. When it needs to change something, you see the Terraform plan in Slack and click Approve. Only then does it get a key that works for that one action and expires in 15 minutes.
account prod-payments
plan 2 to add, 1 to change, 1 to destroy
destroy aws_db_instance.ledger
if approved: key valid 15 min,
only for this plan
In April 2026 an AI coding agent found an API token in an unrelated file and deleted a company's production database and its backups in nine seconds (PocketOS). In July 2025 an agent wiped SaaStr's production database during an explicit code freeze (Replit). Same pattern both times: a key that could do far more than the task needed.
Four steps, every time the agent wants to touch AWS.
Via the AWS CLI, an SDK or an MCP tool. It holds no key of its own.
Read, change or destroy, using fixed rules, not an AI guess.
In Slack, with the Terraform plan and the account. Not the person who asked.
A temporary AWS key for that one action, gone in 15 minutes. Logged.
You choose the rules per account: stricter for production, looser for sandboxes.
If your agents only read, and every change goes through a pull request a human reviews, you may not need this. Many teams run exactly that way and it works.
Keyminute is for the next step: agents that need to apply a plan, roll back a deploy or fix an incident, without anyone handing them a long-lived key.
Production OUs get stricter rules than sandboxes, without a config file per account.
The agent never gets more than a human in that role would.
Not a command string you have to run in your head. Destroys are called out.
Installed with Helm or Terraform. No keys and no command output pass through anyone else's servers.
Not final. Flat price per AWS account, paid by card, cancel any time. No sales call, no contract.
Not yet. I'm validating the idea before building it. Join the waitlist and I'll tell you when there is something to try.
The goal is any agent that reaches AWS through the CLI, an SDK or MCP, including Claude Code, Cursor and Codex. It does not depend on one agent vendor.
A proxy only helps if the agent has no other way in. If it finds a long-lived key anywhere, it walks around the proxy. Removing permanent keys closes that gap.
Everything that touches AWS runs in your account. The hosted part handles licensing and settings only, never AWS keys or command output.
Built by a platform engineer who runs multi-account AWS for a living. Questions or war stories: hello@keyminute.dev