Agentic Sandbox is a self-hosted runtime for persistent AI agent processes. It keeps execution, data, and operator control inside infrastructure you manage.
| Axis | Agentic Sandbox |
|---|---|
| Hosting model | Self-host only. No hosted control plane. |
| Isolation boundary | KVM hypervisor (per-agent kernel). Rootless containers as a lighter alternative. |
| Session duration | Designed for sessions measured in hours to days. |
| Data path | All traffic stays on the operator's network. |
| Agent shape | Bring-your-own agent. Claude Code is the primary tested agent; the runtime is agent-agnostic. |
| Orchestration | Single-host today. Multi-host and Kubernetes are on the roadmap. |
Use the dated security status page as the public source of truth for launch claims.
- Claim self-hosted, local-first operation and KVM isolation as runtime capabilities.
- Describe agent transport identity as support for UDS, vsock, and mTLS; do not imply every deployment profile is verified unless the release evidence says so.
- Describe credentials as metadata-first and lease-oriented; do not claim absolute zero credential exposure.
- Describe standards work as alignment; do not claim certification or compliance without a real program and evidence.
- Claim signed artifacts, SBOMs, and image provenance only for releases where those artifacts are attached and verified.
- Agent workloads where source code or data cannot leave the network (regulated industries, on-prem, air-gapped).
- Long-running sessions that need durable terminal and process state.
- Workloads where hypervisor-level isolation is required (untrusted code execution, adversarial workloads).
- Internal platforms running agent workloads for multiple users on shared infrastructure.
- Operators provide and maintain the Linux hosts, KVM/libvirt stack, storage, networking, and release lifecycle.
- Rootless containers provide the lighter runtime tier; KVM-backed virtual machines provide the stronger isolation tier.
- Multi-host orchestration and Kubernetes integration remain roadmap work.
- Public security claims remain bounded by the evidence linked from the security status page.