security
Agent autonomy without unlimited access.
Cloud coding agents can edit files, run commands, and consume model credits. CloudAgent is designed around isolation, explicit permissions, and auditability.
principles
Six rules we design against.
this website, today
Account & billing controls.
These controls are implemented in the account platform you're using right now.
| Control | What it does |
|---|---|
| Invite-only access | Only admins can create accounts. No public registration path exists. |
| Role-based access | User and admin roles re-checked server-side on every request. |
| Credit limits | Prepaid balance, budgets, and an append-only ledger prevent runaway spend. |
| Audit logs | Admin, payment, and credit actions are recorded immutably. |
| Webhook verification | Only cryptographically verified payment events can grant credits. |
| Signed usage events | Token usage is accepted only from HMAC-signed internal requests. |
| Bot protection | Turnstile and rate limiting on waitlist and login. |
| Generic auth errors | Login never reveals whether an email has an account. |
cloud agent runtime controls
The runtime's guardrails.
These describe the separate cloud agent runtime being built for the beta — listed here as design commitments for that service, distinct from the website controls above.
- Per-task sandbox
- Git branch/worktree isolation
- Scoped secrets
- Command approval policies
- Network egress policies
- Credential redaction
- PR-first output
- Runtime audit logs
compliance posture
CloudAgent is in private beta. We do not claim SOC 2, HIPAA, or other formal compliance certifications yet. We are building with auditability, least privilege, and billing transparency from day one.
Security-sensitive team? We offer security review calls for selected pilots. Report vulnerabilities or ask questions: wu@learnwithpal.org
Talk to us about safe agent autonomy.
We are inviting developers and small teams who already use AI coding agents and want an always-on cloud environment for serious work.