Company Knowledge Base (MCP Server)

The problem

"Ask Bob how the system works" had become a standing bottleneck for support, success, and sales. Every "does X work this way?" question routed through me personally, whether or not the answer was written down anywhere. I designed and built a Model Context Protocol (MCP) server that turns the team’s documented product knowledge into something anyone can query directly through Claude, without waiting on me. This was not a wiki with search bolted on. It is a live service: a small, purpose-built API with three read-only tools, backed by a documentation set organized by product domain, with an authoring pipeline to keep it current and an audit trail for every query.

What it does

• 3-tool MCP server (list_topics, search_docs, get_doc): read-only, bearer-token authenticated, every call audit-logged.
• 5-domain documentation set (payments, onboarding, reporting, loyalty/cashback, controls & risk), each doc scoped by customer segment so an answer never bleeds context across two different product models.
• An authoring pipeline, not manual upkeep: "generate docs" and "check staleness" workflows, both operating through branch and draft pull request so nothing writes to main unreviewed.

Build and ship

Designed, built, and tested locally first, proving the server worked against real docs before touching hosting. Caught and fixed a same-day production-only regression: a newer version of the MCP SDK had silently defaulted to a security setting that rejected every real request with a cryptic error, while every local test passed clean. Traced it, patched it, and added a regression test and a canary self-check so the fix cannot silently regress again.

Stack

Python, FastMCP, Streamable HTTP transport, Render, git branch and PR pipeline for content.