Capability · Multi-Tenant Auth
Cryptographic tenant isolation from the front door down to the database row.
Zelkor verifies caller JWTs at ingress, isolates conversation threads and runs per tenant, derives tool tenancy solely from the verified user, and enforces that tenant across Postgres, Qdrant, and S3.
Why Prompt-Based Multi-Tenancy Fails
Passing a tenant ID in a system prompt or trusting the LLM to fill in a tenant_id tool argument is not isolation. Any prompt injection can tell the model to ask for another customer’s ID, turning a shared agent deployment into a cross-tenant data leak.
Front-Door JWT Verification
Every request is checked for a valid issuer, audience, and cryptographic signature using your identity provider’s keys. Unsigned tokens or requests missing a tenant claim are rejected immediately—there is no default fallback tenant.
Isolated Threads and Runs
Conversation history, checkpoints, and run state are scoped to the authenticated tenant. One customer cannot list, inspect, or resume another customer’s threads.
Tamper-Proof Tool Identity
When an agent calls an MCP tool, the gateway reads the tenant identity exclusively from the verified user header. If the model passes a tenant_id parameter in the tool call, the request is rejected.
Datastore-Level Enforcement
Postgres sets the verified tenant on every read-only transaction so your Postgres Row-Level Security policies filter rows at the database engine. Qdrant forces a tenant filter onto every vector search and document write. Object Storage prefixes every S3 object path with the verified tenant ID.
How Teams Use It
platform:
tenants:
jwt:
issuer: "https://idp.enterprise.internal"
audiences:
- "zelkor"
jwksConfigMap: "zelkor-tenant-jwks"
tenantClaims:
- tenant_idServe every customer from one shared agent graph without cross-tenant data leaks.