Solutions · Multi-Tenant B2B SaaS
One agent deployment. Thousands of customer tenants. Zero cross-tenant tool calls.
Skip per-customer agent silos. Zelkor lets B2B SaaS teams serve every customer from shared LangGraph, LangChain, or Deep Agents deployments while enforcing cryptographic tenant boundaries across threads, Postgres, Qdrant, and S3.
Risks
Where it fails without a runtime boundary
Prompt-Injected Tenant Spoofing
A malicious user tricks your SaaS copilot into passing another customer’s tenant ID in a tool call. How Zelkor stops it: The MCP gateway takes the tenant solely from the verified user context and rejects model-supplied tenant IDs.
Cross-Tenant Thread Hijacking
Customer A tries to read or resume conversation threads belonging to Customer B. How Zelkor stops it: Thread and run state is scoped strictly to the verified tenant claim on the caller’s JWT.
Unfiltered Vector Search
Shared RAG indexes returning document chunks uploaded by another customer. How Zelkor stops it: The Qdrant MCP tool forces a tenant filter onto every vector search and document write.
Shared Bucket Path Traversal
An agent reading or overwriting files in another tenant’s storage path. How Zelkor stops it: The object storage tool enforces tenant prefixes on every key, caps reads and writes at 256 KB, and disables delete by default.
Noisy-Neighbor Runaway Loops
A buggy customer session enters a loop that burns shared LLM rate limits. How Zelkor stops it: The AI gateway enforces a per-minute request cap, and Langfuse checks flag excessive tool steps.
Unsigned Internal Requests
Microservices calling agent endpoints without a valid customer token. How Zelkor stops it: Front-door JWT verification rejects unsigned tokens or requests missing a tenant claim—there is no default tenant.
Controls
Compliance & Control Mapping
| Requirement | What Zelkor Does | How You Check It |
|---|
| Front-Door Tenant Authentication | Verifies JWT issuer, audience, and signature; rejects unsigned or incomplete tokens. | Test requests with missing or unsigned JWTs. |
| Conversation & Checkpoint Isolation | Isolates threads and runs per authenticated tenant inside the runtime. | Attempt to fetch Tenant B’s thread with Tenant A’s token. |
| Database, Vector & Storage Scoping | Sets Postgres tenant session for RLS, forces Qdrant tenant filters, and prefixes S3 keys by tenant. | Run Two Desks, One Graph cross-tenant proof. |
| Automated Loop & Quality Checks | Scores runs in self-hosted Langfuse for step budgets, repeated tool calls, and tenant attribution. | Inspect automated scores in The Flight Recorder proof. |
Topologies
3 Multi-Tenant B2B SaaS Workflows
Customer-Facing SaaS Copilot
Authenticate customer JWTs at the front door, route to your graph, and scope every Postgres query and Qdrant search to that customer’s verified tenant.
Tenant-Scoped File Processing
Read and write customer text artifacts up to 256 KB in S3-compatible storage under strict per-tenant path prefixes.
Startup Tool Injection
Bind built-in and external MCP tools into LangGraph or Deep Agents workloads at startup without embedding database or SaaS credentials in application repos.
Ship multi-tenant AI features from a shared codebase without cross-tenant risk.