A production MCP knowledge platform
A multi-store retrieval system exposed to LLMs over the Model Context Protocol — built and operated end to end: architecture, data engineering, MCP server, auth, security hardening, and deployment.
A large, evolving body of technical work was accumulating faster than any document store could keep usable: heterogeneous content, dense cross-references, and a need to preserve why conclusions were reached. Plain search returned stale or contradictory answers, because superseded positions looked identical to current ones.
A multi-store pipeline — each store doing what it does best — fronted by a single MCP server, so any LLM can use it as a first-class tool source. An append-only, supersession-aware schema lets a query distinguish a current position from a retired one and walk the decision chain that produced it.
Where the engineering was
- MCP client auth. Diagnosed a discovery-vs-registration mismatch in the live MCP OAuth flow, isolated the failing step from request traces, and implemented a working auth path.
- Secure public exposure. Services bound to localhost, host firewall, database authentication, and IP allow-listing on the public tunnel — so an internet-reachable MCP endpoint did not widen the attack surface.
- Retrieval correctness. Schema and query patterns designed so agent queries resolve to the current state and its justification, not the first lexical match.
What it runs on
ingestion Kafka · documents MongoDB · graph Neo4j · search OpenSearch
objects MinIO · server FastMCP (streamable-HTTP) · runtime Python · Docker
secured host firewall · database auth · IP allow-listing
The same pattern, for your corpus
Most organisations sit on a corpus they cannot fully use — documentation, research, support knowledge, contracts, product data. Vitnis can deliver a secure, LLM-native retrieval layer over it: your data → multi-store retrieval and knowledge graph → a secure MCP server your team queries through Claude.