Write @claude in a GitHub issue and get a pull request with a plan, checks and screenshots; you merge, and the site goes live. People you allow can ask without rights in the repo. About an hour to set up: a capped Claude Console workspace, a federation rule instead of an API key, one workflow.
Every agent PR carries a plan and waits for your approval, using only GitHub's own settings: a required plan, a read-only token, a human approval gate, allowlisted MCP tools and hooks that say no.
Know for each AI system in your Swiss company whether the EU AI Act applies, in which role and risk class, and which data may go into US- or EU-hosted AI. Three scope questions, nine bans, one data traffic light and a register entry per system, checked against the official texts. Not legal advice.
