Sandbox As A Service
In DevelopmentSecure containment for autonomous AI agents.
Running AI agents and automations requires an environment that cannot leak data or reach the internet unchecked. We productised the reference architecture that runs our own Nerve Centre to give organisations a white-labelled, isolated workspace provisioned entirely as code.
Containment Architecture
Workload (Agent CLI / API)
Verified agent installs • Default-deny egress allowlists
Docker Container
Nested container • NVIDIA Container Toolkit (no-cgroups) • No internet access
LXD System Container
Unprivileged LXD • GPU Profile • OpenTofu Managed
Host OS & Hardware
NVIDIA Drivers • ZFS Storage • TPM Hardware
What you get
Isolated Workspaces
Provisioned as code via OpenTofu. Docker containers run nested within unprivileged LXD containers to strictly isolate workloads from the host.
GPU Passthrough
Hardware GPU passthrough enables local, low-latency AI model inference directly within the secure sandbox, bypassing the need for remote cloud dependencies.
Egress Lock-down
Default-deny egress networks limit connections. Workloads and agent users are strictly fenced using per-workload proxy allowlists.
Model Gateway
A dedicated gateway manages keys, budgets, and call telemetry for local or cloud AI models. Prompts and responses are never stored.
Hardware-backed Secrets
No plaintext configuration. Secrets are stored in a TPM-sealed vault and hardware-key backed storage.
Governance & Rollbacks
Human approval is required for privileged changes. Workflows take system snapshots before execution and automatically rollback on failure.