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
The self-hosted agent runtime

Ship agents, not cluster breaches.

Bring the agent you already wrote — it can't break out, reach unauthorized data or networks, its prompts are verified, budget controlled, and it is under observation; tenants stay isolated.

View on GitHub →

helm upgrade --install zelkor-platform charts/zelkor-platform \
  --namespace zelkor \
  -f profiles/values-production.yaml \
  -f profiles/values-gateway-greenfield.yaml \
  --wait
gVisor code sandboxBrokered model keysNeMo GuardrailsGoverned MCP toolsVerified JWT tenancySelf-hosted Langfuse tracesEgress network lockdowngVisor code sandboxBrokered model keysNeMo GuardrailsGoverned MCP toolsVerified JWT tenancySelf-hosted Langfuse tracesEgress network lockdown

Prompt filters are probabilistic.Kernel, network, and credential boundaries are not.

System prompts cannot stop a hijacked tool call from reading environment variables, dialing an external server, or asking a database for another customer’s records. Zelkor enforces those boundaries in the Kubernetes infrastructure around your agent pod, not inside the prompt.

Walkthrough

4-tab architecture walkthrough

Untrusted Python runs outside the agent pod

Your agent pod runs your graph and Python dependencies on the standard container runtime. When the agent needs to execute generated Python, the code is sent over MCP to a separate pool of warm gVisor worker pods. Your agent pod’s memory, files, and service account are never exposed to that code.

  • Dedicated gVisor Pool: Three warm worker replicas run untrusted code inside a user-space kernel boundary.
  • No Direct Network Path: Agent pods cannot dial sandbox workers directly; all code requests pass through the MCP gateway.
  • Authenticated Execution: Sandbox workers require an internal gateway token and reject unauthenticated requests.
helm upgrade --install zelkor-platform charts/zelkor-platform \
  --namespace zelkor \
  -f profiles/values-production.yaml \
  -f profiles/values-gateway-greenfield.yaml \
  --wait
How it ships

3 steps to production

1

Bring your agent

Package the LangGraph, LangChain, or Deep Agents application you already built. You keep your state machine and dependencies as they are.

2

Close the doors

Install the Zelkor Helm chart into a namespace your cluster admin created. Turn on Kubernetes NetworkPolicy, route model calls through the AI gateway, and send untrusted Python to the gVisor worker pool.

3

Run it for teams

Onboard multiple teams into isolated agent namespaces with CPU and memory requests and limits set on every container. Track every run and tool call in self-hosted Langfuse.

Proofs

5 live security proofs

The Vanishing Key — Prompt Injection Dumps Environment Variables

Scenario. A malicious prompt tricks the agent into reading its own environment variables to steal your OpenAI or Anthropic API key.

Control. The agent pod holds no upstream provider keys—only a local token for the cluster AI gateway. Kubernetes NetworkPolicy also blocks the pod from calling external model APIs directly.

Inspecting the pod environment reveals no provider secret to steal, and the local token does not work outside the cluster.

The Flight Recorder — Reconstructing an Incident Without Developer SDKs

Scenario. An agent misbehaves in production, and security needs to know which prompts, guardrails, and tools ran—even though developers never added tracing code.

Control. The runtime and gateways record OpenTelemetry spans automatically into self-hosted Langfuse, while keeping raw prompts and secrets out of default logs.

A single waterfall trace in Langfuse shows the entire run, gateway checks, and tool calls, alongside automated run scores.

The Shell Moved — Generated Python Tries to Escape the Container

Scenario. Agent-generated Python code runs shell commands, reads local files, or probes the kernel to break out of the container.

Control. Untrusted Python never runs in the agent pod. It executes in a separate pool of warm gVisor workers reached only through the MCP gateway.

System calls are trapped by the gVisor user-space kernel, and the agent pod’s filesystem and service account are untouched.

Two Desks, One Graph — Prompt Injection Requests Another Tenant’s Data

Scenario. A user in Tenant A tricks the model into calling a database or vector search tool with Tenant B’s identifier.

Control. Front-door JWT verification locks in the caller’s identity. The MCP gateway takes the tenant solely from that verified user context and rejects model-supplied tenant IDs.

The cross-tenant tool call is rejected before reaching Postgres, Qdrant, or S3, and Tenant B’s records stay hidden.

Loop Dies at the Door — Runaway Agent Loop Burns Model Quota

Scenario. A broken reasoning loop fires dozens of model requests in seconds, threatening upstream rate limits and API bills.

Control. Envoy AI Gateway enforces a global request cap per minute at the cluster boundary, and Langfuse run checks flag excessive tool steps.

Once the per-minute cap is reached, the extra model call is rejected at the gateway before reaching the LLM provider.

Who it is for

Built for 3 core personas

Security & CISO Teams

Stop relying on system prompts as a security control. Enforce gVisor sandboxing for untrusted code, keep provider keys out of workload pods, lock down pod egress with Kubernetes NetworkPolicy, and bind data access to verified JWTs.

  • gVisor sandboxing for untrusted code
  • Provider keys out of workload pods
  • Kubernetes NetworkPolicy egress lockdown
  • Data access bound to verified JWTs

Put real boundaries around the agents you already wrote.

Deploy Zelkor Community Edition on your Kubernetes cluster, or walk through the five live security proofs with our engineering team.