Security & network
Cloud agents run in isolated environments. This page covers the controls that keep those runs secure, the network access they require, and the domains your corporate firewall must allow.
Isolation
Each run executes in its own sandboxed environment with only the repositories, secrets, and permissions you grant. Runs don't share state with each other or with your local machine.
Isolation is one layer. Combine it with least-privilege repo access, secret injection (not prompts), and human review of PRs before merge.
Network egress
By default, restrict outbound network access to what the task needs. Allowlist specific domains (package registries, your APIs) rather than permitting unrestricted egress.
Common allowlist entries
Add these domains when cloud agents need to install dependencies and reach 4RGE services:
If you use a private npm, PyPI, or Docker registry, add those hostnames explicitly. Test with
a smoke run that only runs npm ci or pip install before delegating larger tasks.
Corporate desktop and CLI allowlist
Users on locked-down networks need the same domains for the desktop app, CLI, and browser apps — not just cloud run sandboxes:
Allow 4RGE service domains
Permit HTTPS to
*.4rged.aiand your Supabase project URL (*.supabase.co).Allow GitHub if using integrations
Permit
github.comandapi.github.comfor OAuth and the GitHub App.Configure proxy trust
If traffic passes through an HTTPS proxy, ensure the proxy presents a trusted certificate chain to the desktop app and CLI.
Verify from each surface
Test sign-in (auth), start a local agent run (desktop), and start a cloud run (agents API) from the same network segment.
Access control
Durable job queue
Cloud execution uses a separate x4rge_agent_jobs queue decoupled from browser streams.
Each run gets one job row with status queued, claimed, running, completed, failed,
cancelled, or dead_letter. Workers claim jobs with time-bounded leases; max_attempts
defaults to 3 before a job moves to dead_letter. This design lets agents-api scale workers
independently of the Agents workspace UI and gives Sentinel visibility into backlog and retry
state alongside run status.
Secrets handling
Use org secrets, not prompts
Configure API keys and tokens in environment settings. Never paste credentials into task descriptions — they may appear in logs or audit exports.
Rotate on departure
When a team member with secret access leaves, rotate affected keys and review recent audit events for cloud runs that used them.
Keep secrets out of git
Ensure
.env, credential files, and key material are in.gitignore. Cloud agents respect the same ignore rules as local agents.
Security review checklist
Before enabling cloud agents for a production repository:
- Repository access limited to required repos only
- Egress allowlist includes registries and APIs the setup commands need
- Secrets injected via platform — none in prompts or committed files
- Budget policy set with headroom alerts
- Automations start in
draftand are manually verified beforeenabled - PRs from cloud agents require human review before merge
If your organization proxies or filters traffic, allowlist the 4RGE service domains listed above so the desktop app, CLI, and cloud runs can authenticate and operate. Share this page with your network admin.