Skip to main content

We built an MCP server so AI assistants know who we are

15 December 2025

Why we built this

Ask Claude or ChatGPT "tell me about craftable software" and you'll get a mix of correct facts, outdated details, and confident guesses. That's how LLMs work — they approximate. We wanted a way to give AI assistants the real information, straight from us.

MCP (Model Context Protocol) is an open specification that does exactly that. You run a server, it exposes your content, and AI clients connect to it. We built one for our own website as a working demo — the kind of thing we'd build for a client, applied to ourselves first.

Where MCP shines

A humanoid robot sitting on a bench reading a tablet
MCP connects AI assistants to the systems where real work happens — not just public websites.

Our website is public content. An AI could scrape it. The protocol was built for harder problems than this.

Where MCP matters most is behind the login screen. A support agent that queries your CRM and ticketing system through one interface, without a custom integration per tool. A coding assistant that reads your internal runbooks and architecture specs — the ones that aren't on a public wiki. An operations dashboard where someone can ask "what failed in the last deploy?" and get an answer from the CI/CD pipeline directly.

We chose our website because it's a safe, public-facing starting point. Same protocol, same patterns, lower stakes. We wanted to learn by building, and we wanted something other people could actually connect to and try.

The protocol vs the server

Abstract close-up of a colourful digital interface
MCP defines the interface — you decide what goes behind it.

MCP is a spec, not a product. There's no platform to sign up for. The specification defines three primitives that any server can implement:

  • Resources — content the AI can read. Think of them as structured documents with URIs, like craftable://services or craftable://case-studies/fintech-platform-migration.
  • Tools — actions the AI can trigger. A job search, a form submission, a database query. They accept parameters and return results.
  • Prompts — reusable templates that guide how the AI uses your resources and tools. Ours include a consultancy fit evaluation and a project assessment brief.

You write a server that implements these primitives, expose it over HTTP, and any compatible AI client can connect. The spec handles the handshake, capability negotiation, and transport. Your job is deciding what content and actions to expose.

What we put in ours

We exposed company knowledge as resources — who we are, what we do, how we price engagements, which industries we work in, and seven case studies with real project details. On the tools side: a keyword search across all content, a live job listings feed from Workable, and a way to submit a contact inquiry directly through the AI conversation.

The content stays in sync with the website because both read from the same data files. When we update a case study or change our team size, the MCP server picks it up on the next build. One place to edit, two surfaces that stay aligned.

Getting it into Cursor, VS Code, Claude, and ChatGPT

Once the server is running over HTTP, connecting it to AI clients is configuration. Each client has its own format, but the idea is the same: point it at the URL.

Cursor — add a .cursor/mcp.json file to your project with the server URL. That's it. Cursor connects on startup and the server's tools and resources show up in the agent.

VS Code — similar setup in .vscode/mcp.json. Works through GitHub Copilot's MCP support. Same URL, slightly different JSON shape.

Claude Desktop — doesn't support remote HTTP servers natively yet. You bridge it with mcp-remote, a small npm package that translates between Claude's local stdio transport and your remote HTTP endpoint.

ChatGPT — supports Streamable HTTP directly. You can point the Responses API at the server URL, or submit to OpenAI's Apps Directory for a public listing with a widget UI.

We also registered the server on the MCP Registry under com.craftablesoftware/mcp. The registry is a public directory — once you're listed, clients like VS Code can discover your server automatically through search. Publishing uses mcp-publisher, a CLI tool that authenticates via DNS (you prove you own the domain) or GitHub Actions OIDC for automated releases.

Try it

Connect to https://mcp.craftablesoftware.com/mcp from any of the clients above. Ask it which industries we work in, what our pricing looks like, or whether we have open roles.

If you're thinking about connecting AI to your own systems — internal tools, proprietary data, client-facing workflows — that's the work we do. Let's talk about it.