Agents your risk teamcan sign off on.

100%

Incidents resolved end to end on the public SREGym-Lite leaderboard, on GPT-6 Astra, checked October 11, 2026

736k

Tokens per run on GPT-6 Astra, the fewest of any agent at 100% on that model

~97%

Review precision on live merge requests at FPT Cloud

SOC 2

Type II report on the Security criteria. ISO/IEC 42001 audit in progress

Scoped connections

Agents reach your systems only through connections a workspace Admin adds, with the credentials and permissions you grant.

Read-only to start

Reads always run. A read-only credential can't write, and each connection has its own Allow writes switch.

Short-lived cloud credentials

AWS connections assume an IAM role through STS, so no long-lived key can leak. Stored secrets are encrypted at rest.

Always blocked

A built-in list of calls never runs, such as stopping CloudTrail logging or deleting kube-system. Nobody can override it.

Approvals

Writes go through Approvals. You decide what runs on its own and what waits for approval, per workspace and per run type.

Two-person rule

Turn it on and nobody approves their own run. Viewers can never approve.

Data masking

Secrets, emails, phone and card numbers are replaced with placeholders before the model sees them. Lock the rules for every workspace.

Audit log

One timeline of user, agent and system actions across every workspace, kept 90 days and exported as CSV.

Isolated workspaces

Each workspace keeps its own connections, credentials, agents and approvals. Commands run in CloudThinker's Sandbox.

1. 01

   Always blocked

   A built-in list of calls that never run. No Rule and no person can change it.

   For example

   Stopping CloudTrail logging, closing an AWS account, deleting kube-system

2. 02

   Rules

   Allow, ask first or deny a named kind of write. Rule enforcement is in preview, and we turn it on per workspace.

   For example

   Ask first on database deletes, deny IAM policy changes

3. 03

   Approval mode

   You decide what runs on its own and what waits for approval, per workspace and per run type.

   For example

   Production waits for a person on every write, while staging restarts a pod on its own

4. 04

   Approver

   A named person approves or declines. Viewers never can, and the two-person rule stops self-approval.

   For example

   Default approvers for most asks, a DBA group for database deletes

Reads always run. Each decision, by a person or by AI judgment, is recorded in the audit log.

CloudThinker Cloud

The hosted service, with nothing to install. With Environments (beta), file and shell work runs on a machine you control.

Early access

Dedicated

A single-tenant instance hosted by CloudThinker, in a region you choose.

Early access

Self-hosted

Inside your own cloud account or data center, with reference architectures for AWS, Azure, GCP and FPT Cloud.

Early access

Air-gapped

Self-hosted and offline, for environments where data can't leave your perimeter.

Masked before the model

Secrets and protected identifiers in tool output are replaced with placeholders before the model or the screen sees them.

Your own model account

With BYOK, model calls run through your own AWS Bedrock, Anthropic or FPT Cloud account. If your key fails, calls fall back to CloudThinker's models.

SSO, SCIM and roles

SAML or OIDC sign-in, SCIM group sync into workspaces, and Owner, Admin, Developer and Viewer roles.

Change management

Approval decisions and agent writes sit in the audit log next to your people's actions, ready to export for change records.

Model risk

Replay your own incidents and score each verdict. Resolve tracks root-cause accuracy and false-positive rate from your team's marks.

Third-party risk

Our SOC 2 Type II report, public security docs and questionnaire answers, ready for your vendor review.

Data residency

Dedicated deployments run in a region you choose, and self-hosted ones stay in your perimeter. Both are in early access.