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
Editions & Deployment

Every security control ships in Community Edition. Enterprise adds architecture advisory.

We do not hold back security controls for a paid tier. The gVisor sandbox pool, Envoy AI Gateway, NVIDIA NeMo Guardrails, governed MCP tools, verified JWT tenancy, NetworkPolicy lockdown, and self-hosted Langfuse traces are all included in the open-source Helm chart.

Free & Open Source · Apache-2.0

Quickstart (CE)

Local or dev Kubernetes clusters, running the five security proofs, and single-namespace evaluation.

  • Single Helm chart install into an existing namespace
  • Runs LangGraph, LangChain, and Deep Agents without graph rewrites
  • Standalone Envoy Gateway and Envoy AI Gateway setup
  • Warm gVisor worker pool (3 replicas) for untrusted Python
  • Self-hosted Langfuse with built-in run checks
  • Reproducible 5 Live Security Proofs out of the box
Deploy on Kubernetes
Free & Open Source · Self-Hosted Production

Production (CE)

Platform teams running multi-tenant agents under strict Kubernetes quotas, network policies, and identity controls.

  • Everything in Quickstart (CE), configured for production clusters
  • Standalone, layered (behind existing ingress), or shared gateway topologies
  • Kubernetes NetworkPolicy egress lockdown enabled
  • Isolated agent namespaces for run-state database protection
  • Front-door JWT verification and fail-closed NVIDIA NeMo Guardrails enabled
  • CPU and memory requests and limits on every container for quota compliance
Deploy on Kubernetes
Custom Engagement · Direct Engineering Partnership

Enterprise Architecture Advisory

Regulated organizations that want hands-on topology design, identity and database isolation review, and go-live validation with our engineers.

  • Same open-source platform and security controls as Production (CE)
  • Gateway topology and Kubernetes NetworkPolicy design for your cluster
  • gVisor node pool sizing and scheduling review (GKE Sandbox, Talos, or labeled nodes)
  • End-to-end JWT identity, Postgres Row-Level Security, and Qdrant tenancy review
  • Custom NVIDIA NeMo Guardrails policy and external MCP secret custody review
  • Live execution and sign-off of the 5 security proofs on your cluster
Book Architecture Review
Compare

Feature matrix

CapabilityQuickstart (CE)Production (CE)Enterprise Advisory
1. Agent Frameworks (LangGraph, LangChain, Deep Agents)IncludedIncludedIncluded + packaging review
2. Standard Run & Thread EndpointsIncludedIncludedIncluded + multi-graph routing review
3. Gateway TopologiesStandalone defaultStandalone, behind ingress, or shared gatewayIncluded + cluster ingress design
4. Model Key Custody on AI GatewayIncludedIncludedIncluded + secret rotation review
5. Global Request Cap (50/min default)ConfigurableOn by defaultIncluded + capacity sizing
6. NVIDIA NeMo GuardrailsConfigurableOn by default (fail-closed)Included + custom rail workshop
7. gVisor Python Sandbox Pool (3 warm replicas)IncludedIncludedIncluded + node pool sizing
8. Unified MCP Tool GatewayIncludedIncludedIncluded + tool governance review
9. Built-In MCP Tools (Postgres, Qdrant, S3, Sandbox)IncludedIncludedIncluded + Postgres RLS review
10. External MCP Backends (Dedicated Secret, No JWT Leak)IncludedIncludedIncluded + third-party threat review
11. Front-Door JWT VerificationConfigurableOn by default (no default tenant)Included + IdP & JWKS integration
12. Verified Tenant Propagation to Tools & ThreadsIncludedIncludedIncluded + cross-tenant test sign-off
13. Self-Hosted Langfuse Traces (No SDK in Agent Code)IncludedIncluded + Prometheus metricsIncluded + storage sizing review
14. Redacted Default Logs & Built-In Run ChecksOn by defaultOn by defaultIncluded + audit workflow review
15. Kubernetes NetworkPolicy & Container Resource LimitsLimits on by default; NetworkPolicy optionalBoth on by defaultIncluded + CNI & quota audit
FAQ

Questions

Which agent frameworks work with Zelkor?

Zelkor runs applications built with LangGraph, LangChain, and Deep Agents. You bring the graph you already wrote, and Zelkor serves its thread and run endpoints behind Envoy without asking you to rewrite your state machine. You can also bind MCP tools into the graph automatically at startup.

Does my entire agent pod have to run inside gVisor?

No. Your agent pod runs on the standard Kubernetes container runtime so your Python packages and native extensions work normally. Only untrusted Python execution is sent over MCP to a dedicated pool of three warm gVisor worker pods. Agent pods cannot dial those workers directly, and workers require an internal token from the MCP gateway.

How does Zelkor keep one tenant from reading another tenant’s data?

Every request is verified at the front door using a signed JWT—unsigned tokens or requests without a tenant claim are rejected. When the agent calls a tool over MCP, the gateway reads the tenant only from that verified user context and rejects any tenant ID passed by the model. Postgres runs read-only queries with the tenant set on the session for Row-Level Security, Qdrant injects a mandatory tenant filter, and S3 keys are prefixed by tenant.

How does Zelkor stop a looping agent or a leaked model key?

Agent pods never hold upstream API keys for OpenAI, Anthropic, or Bedrock; those keys stay in Kubernetes Secrets on Envoy AI Gateway. The gateway enforces a global request cap (50 requests per minute by default) and rejects extra calls before reaching the provider. You can also enable NVIDIA NeMo Guardrails on the gateway to inspect chat completions and fail closed on policy violations.

Do I need a specific CNI or a tracing SDK in my agent code?

Neither. Network lockdown uses standard Kubernetes NetworkPolicy, so it works with the CNI you already run. Tracing is handled at the runtime and gateway layer and sent to self-hosted Langfuse inside your cluster, with no SDK imports required in your graph and no prompt bodies or keys written to default logs.

Is Enterprise a separate product with locked security features?

No. Every security control in Zelkor—including the gVisor sandbox pool, AI gateway key custody, NeMo Guardrails, MCP tenant enforcement, NetworkPolicy, and self-hosted Langfuse—is included in Community Edition. Enterprise is a direct architecture advisory engagement where our engineers help you integrate Zelkor with your cluster ingress, identity provider, Postgres RLS policies, and compliance sign-off.

Run your agents inside hard Kubernetes boundaries.

Install the open-source Helm chart into your cluster, or book an architecture review to map Zelkor to your network, identity, and database setup.