Live demo scenarios5 reproducible security proofs: No provider API keys in agent pods • gVisor Python sandbox • Cryptographic tenant isolation • Gateway request circuit breakerExplore the 5 Proofs
Capability · Code Sandbox

Run untrusted Python in a dedicated gVisor pool, never inside your agent pod.

Your LangGraph, LangChain, or Deep Agents container stays on the standard Kubernetes runtime so your libraries work without friction. Whenever the agent runs generated Python, the code is sent over MCP to a separate pool of gVisor worker pods.

Why Running Code in the Agent Pod Fails

If an agent executes generated Python inside its own container, a single prompt injection gives that script access to the pod’s filesystem, memory, environment variables, and Kubernetes service account. Even moving code to a second standard container still shares the host Linux kernel with neighboring workloads on the node.

Warm gVisor Worker Pool

Three pre-warmed worker replicas run under gVisor, intercepting system calls in a user-space kernel so untrusted Python cannot reach the host kernel or neighboring pods.

Routed Through MCP

Code execution is exposed as a governed MCP tool. Your agent pod sends the code over MCP and gets the output back without running an interpreter locally.

No Direct Agent-to-Worker Access

Agent pods do not know worker IP addresses and are blocked by Kubernetes NetworkPolicy from connecting to sandbox workers directly. Workers require an internal token held only by the MCP gateway.

Simple Node Scheduling

Schedule sandbox workers onto gVisor-enabled nodes using a standard node label, or use automatic detection on GKE Sandbox and Talos Linux clusters.

How Teams Use It

security:
  sandbox:
    enabled: true
    runtimeClass: gvisor
workspace:
  tools:
    sandboxMCP:
      enabled: true
      warmPool:
        replicas: 3

Isolate agent-generated Python without breaking your agent runtime.