An Agentic Layer Over the Enterprise Data Stack, Built on MCP
A self-hosted platform that exposes 21 data connectors as governed MCP servers and runs a fleet of AI agents over a compiled knowledge graph, answering questions, monitoring freshness and quality, and taking audited write actions.
The challenge
Knowledge about a company's data stack is scattered across warehouses, orchestrators, and wikis. Answering a simple question, such as which pipeline produces a table, whether it is fresh, or how many customers there are, means digging across tools, and LLM assistants hallucinate without grounded metadata. The client needed an AI data assistant that runs inside their own network, with real access controls and an audit trail.
Architecture
AI agents
- On-demand chat
- Background cadence agents
MCP servers
- One per connector
- ~268 scoped tools
Governance
- Per-agent grants
- SQL validation
- Approvals, audit
Data stack
- 21 connectors
- Warehouses, ETL, wikis
What we built
We exposed each of 21 connectors (data warehouses, ETL and orchestration systems, and knowledge bases) as its own MCP server built on FastMCP, offering roughly 268 read and write tools, each prefixed and scoped by capability.
Access is governed per agent: an agent can only call the tools it has been explicitly granted. Read SQL is validated as read-only; destructive statements are blocked and instead raise an approval-required alert. Every write is recorded in a full audit trail.
A knowledge-graph compiler turns wiki pages, warehouse metadata, and pipeline state into nodes, edges, and column-level lineage, so agents cite grounded facts instead of guessing.
A unified agent model covers both on-demand chat and background agents that run on cadences (freshness, schema-drift, query-cost, and alerting), with a run-budget enforcer bounding token and tool spend. The whole system runs self-hosted and supports cloud or fully-local LLMs.