AI assistants do not understand our API and web components out of the box. With curated context material they become useful co-developers for your museum IT. We build that material and help with the integration.
Claude Skill, coming soon
In preparation: a Git repository with the Claude Skill for go~mus integration. The advantage over a one-off file download is that updates flow into your AI tools through git pull whenever we adjust the API, web components or common patterns. No more stale context in your skills folder.
Once the repository is public, you will find here:
- The clone command for Claude Desktop and Claude Code
- A short guide for setting the skill up locally
- The content scope: web components, REST API, common integration patterns
If you want to be part of the beta as soon as the repo is ready, drop a note to support@giantmonkey.de.
Other AI tools
The material is not tied to Claude. Markdown context is read by ChatGPT, Cursor, Aider, Continue and other coding assistants just as well. Once the repository is in place, it works in any tool that takes a documentation path into its prompt context.
We help with the integration
An AI toolchain does not replace an architecture conversation. For the actual connection to your museum IT, from ticket sales through POS integration to backoffice connections, we support you, with or without AI tooling. Reach us at support@giantmonkey.de or book a slot via contact.
WebMCP
We are running WebMCP in production on the booking page for our webinars. WebMCP is an experimental browser API that lets a page expose its functions directly to AI agents instead of having them infer behaviour from the DOM. A page registers a tool with name, description and input schema via navigator.modelContext.registerTool(). Browser-internal agents and external agents that drive Chrome over the DevTools protocol then call the tool.
Status May 2026: behind chrome://flags/#enable-webmcp-testing in Chrome 146 Canary. No origin trial so far. The registration call is feature-detected and stays silent elsewhere.
On /en/webinars/ the page registers a book_webinar tool. The input is a termin ID an agent can read from the ?webinar=ID deep-links on the termin cards. The execute opens the booking overlay for the selected termin and returns a status message. The tool deliberately does not prefill the form or submit the booking: submit stays a human action, which matches the fee for third parties.
Lighthouse 13.3.0 picks up the tool, validates the input schema and shows the source location in the bundled JS. The code lives in WebinarePage.svelte, around 50 lines including schema and execute handler.