MCP gateway
I set up the company Model Context Protocol (MCP) gateway on FastMCP, with Okta sign-in, as the one route AI agents use into internal systems. Each new backend runs as a pilot before release. At 92 tools I moved them behind a search tool and a call tool, so an agent loads only what it needs.
Agents load 2 tools instead of 92 and every tool stays callable.
- 92 → 2tools loaded into each agent
- 5backends behind one gateway
Context
Every MCP backend added its tools to the context of every agent. By 23 September 2026 that was 84 tools, about 41 from GitHub and 22 from Atlassian.
What I did
- Ran one FastMCP gateway on the EKS cluster of the shared AI platform, with Okta sign-in. Every request carries the caller and a request ID.
- Set the release path for new backends. A new MCP opens to admins and testers first and goes public after the pilot.
- Brought the data warehouse MCP onto the gateway and set the rule that the company runs one gateway instance.
- At 92 tools, turned on server-side keyword search (BM25) so agents see 2 tools and every tool stays callable.
- Kept 5 tools exposed directly for n8n workflows that call them by name.