Help Scout, in plain English

MCP (Model Context Protocol) is the standard that lets AI assistants like Claude or ChatGPT reach into outside systems. Most MCP servers ship as a local install: you run a piece of software on your laptop, and that becomes the bridge between Claude and the service. That’s fine if you’re a developer wiring up your own tools. It doesn’t work for a SaaS, where every customer should be able to point their AI assistant at their workspace without installing anything themselves.

This one is the third attempt. I built the first two while working on the AI-resolution reviewer, thinking I’d need a clean tool layer for the evaluator to talk to Help Scout. Turned out I didn’t. By then the server existed, and I’ve used it on every AI project since.

The current implementation: hosted, multi-tenant, OAuth at the door, around fifty tools across the Help Scout API. A customer adds one URL to Claude (or ChatGPT, or Cursor), signs in to Help Scout once, and their assistant can search conversations, draft replies, manage tags, query the docs, and pull reports. Nothing to install locally. Tokens are encrypted at rest with HMAC-derived lookup keys, so a stolen database dump doesn’t yield a replayable credential. I built it end-to-end, and I maintain it.