Enterprise AI · Fortune 500 Consulting Firm

A governed MCP gateway for enterprise agents

We extended an open-source MCP gateway and registry into a single governed entry point for the firm's AI agents, with access rights enforced and every call audited.

25% → 74%
Tool routing accuracy
Top-1 on the hard query set, with LLM reranking
177
Tools across 78 servers
In the routing evaluation catalogue

The challenge

The firm had dozens of internal MCP servers and agents that needed to reach them. Without a central gateway, there was no single place to control which callers could use which tools, and agents struggled to route a natural-language request to the right server.

What we built

  • ✦Built on the open-source MCP Gateway & Registry: an nginx data plane, a FastAPI registry, and an OAuth2/OIDC auth server
  • ✦Hybrid BM25 and vector retrieval, followed by LLM reranking to pick the right tool from the top candidates
  • ✦RFC 8693 on-behalf-of token exchange, so agents act with the calling user's identity
  • ✦Staleness and drift detection, failover and SLA monitoring, and a CSV bulk importer for onboarding servers
  • ✦Deployed on AWS ECS with Terraform, and decisions recorded in ADRs

Security & compliance

  • ✦Rights enforcement separates agentic callers from GenAI-only callers. Machine tokens are blocked from human-only servers
  • ✦Unknown callers are denied by default, and every decision writes an audit line
  • ✦SSO through Keycloak and Okta, with scope-based access control
  • ✦Routing measured on held-out queries with significance testing. Self-written test queries had inflated accuracy by about 30 points

Stack

PythonFastAPInginxOAuth2 / OIDCAzure OpenAIAWS ECSTerraform

Building in a regulated space?

Send us an email and we'll walk you through how we'd approach it.

Email us