Liberra reads your cloud. Nothing changes until you say so. Here is exactly how that works.
Connection
How Liberra connects
Liberra never holds a credential of yours on either cloud. How that is achieved differs, because the two clouds grant access in genuinely different ways, so both are spelled out rather than blurred into one reassuring sentence.
AWS
You deploy a CloudFormation stack. It creates one IAM role. Liberra uses that role and nothing else.
Your account, your role
The IAM role lives in your AWS account. You own it. Liberra never touches credentials directly.
External ID
The role requires a unique ID tied to your Liberra account. Without it, nothing can assume the role, including anyone who compromises Liberra.
1-hour sessions
Every STS session token expires after 60 minutes and rotates automatically. No long-lived credentials stored anywhere.
Revoke in seconds
Delete the CloudFormation stack. Access is gone immediately. No action needed on our end.
Azure
You approve an Azure Lighthouse delegation, scoped to the subscription you pick. No keys of yours are created or stored.
Your tenant, your delegation
Access happens inside your own tenant. Nothing is copied out of it.
Scoped to one subscription
The delegation covers only the subscription you choose, at the role you grant it.
No keys at all
There is no secret to leak, because none is issued. Nothing of yours is held on our side.
Revoke in seconds
Remove the delegation. Access is gone immediately. No action needed on our end.
Access levels
Free vs Pro
Two clear modes. No surprises.
Free
Read-only. Nothing changes.
Inventory, costs, security checks. Liberra reads your account and tells you what it finds. Nothing changes.
Enforced in application code on every request.
Pro
Changes need your approval.
Liberra proposes the action and shows you exactly what will happen. You click Approve or Reject. Nothing runs until you say so.
One yes covers the follow-through. Dangerous ops always re-ask.
Approval flow
Nothing runs until you approve
Before changes start in your account, you see exactly what will happen. Not a summary. The actual resource, the actual action. One approval covers the follow-through steps of the same request; destructive and security-sensitive operations always re-prompt.
Pending Approval
Stop dev-worker-2, idle 21 days
snapshot first · saves ~$77/mo
Approve
Reject
Hard blocks
What is permanently blocked
Any call with delete, terminate or purge in its name is blocked before it reaches your cloud. On Azure the same block catches every HTTP DELETE. It is a keyword check in code rather than a list, so it covers every service on every cloud, including ones that do not exist yet.
ec2.terminate_instances
rds.delete_db_instance
DELETE /virtualMachines/{name}
s3.delete_bucket
DELETE /resourceGroups/{name}
cloudtrail.stop_logging
A few examples. Full list in the security repo.
Six AWS services blocked entirely
organizationsstsaccountssosso-adminidentitystore
These services can remove accounts from your AWS Organization, assume arbitrary roles, or close your entire account. No operation on them ever executes, with one deliberate exception: sts get_caller_identity, a harmless read that answers "which account am I?". Everything else is denied.
Your data
What we store, what we never store
Straight answers to the questions everyone asks.
What we store
Your account details (name, email), a reference to the connection itself (an IAM role ARN and encrypted External ID on AWS, a delegation on Azure), the cloud metadata that makes up your Cloud Index, and your chat history.
What we never store
Cloud access keys, session tokens, secret values. Reading Secrets Manager values is blocked in code, so secrets can't even pass through a conversation.
Where it lives
Encrypted at rest in MongoDB Atlas. Your chat history, Cloud Index, and connection credentials are deleted when you delete your account.
What the AI sees
Your messages and cloud metadata go to Anthropic's Claude API to generate answers. Anthropic does not train on this data and retains it only briefly. Your credentials never go to the AI.
Can the AI be tricked?
Safety is code, not prompt. Blocked operations are refused by the executor before your cloud is ever called, no matter what the conversation says, no matter who's asking.
Audit trail
Full audit trail
Every action Liberra takes is recorded in your own audit log: CloudTrail on AWS, the Activity Log on Azure. Tamper-evident, provider-managed, independent of our systems.
Your log, not ours
Every API call lands in your own account, under your own identity. You have a complete record we cannot touch or alter.
Just ask
Not sure what happened? Ask Liberra directly. It will tell you exactly what it did, in plain English.