Topic

MCP servers

The Model Context Protocol (MCP) is the open standard through which hosts such as ChatGPT, Claude, VS Code and Gemini CLI let an AI agent use a product's tools. These notes cover when to build a server, what each host requires of it, how to test it, and how websites are being opened to agents.

An MCP server is the program that lists those tools, and an AI host is the application that runs it for a person. The short answers from these notes: build a server now if customers already ask agents about you or agents already scrape you, and wait if nobody is asking yet. Design it to the strictest host rule, because eight hosts add their own rules to the specification and conflict on name length, allowed characters, required annotations and budgets. Expect to assemble the testing yourself: of about eighty tools reviewed on 1 October 2026, every capability exists somewhere, but no open, local tool combines them. And check what your website already exposes, since Wix, Shopify, GitBook and Mintlify sites already have MCP endpoints.

Most of the notes summarise research founder Amir Pournasserian published on pournasserian.com. They are what we build to in our MCP servers and AI integrations service: a two-week design sprint that ends in a working prototype against one host, then a production build secured with OAuth 2.1 and tested against each host's rules.

MCP servers and AI integrations

Working on one of these problems?

The patterns in these notes are the ones we use on client work. Tell us what you are building.

Book an AI consultation