Muster: the evening check-in for construction

A construction schedule records who is booked for tomorrow. It never records who is actually coming. Superintendents close that gap by phone, fifteen calls in the evening after a ten hour day, and the calls that matter are the ones nobody had time to make. The short crew turns up at seven in the morning, with the concrete already ordered. Muster is a voice agent that has those calls. It reaches each foreman the evening before, asks how many are coming, what time, whether each prerequisite is ready, and whether anything is in the way. It speaks English or Spanish depending on the contact. The conversation runs on the AssemblyAI Voice Agent API over a single WebSocket, with listening, turn taking, reasoning and speech on one connection. Muster supplies the tools on the other end. Every answer is written into a validated check-in record the moment it is given, through tools such as confirm_attendance, confirm_requirement, log_delivery and report_blocker. Nothing is recorded that was not said, and what the agent cannot confirm it leaves alone rather than guessing. The ranking is deliberately not the model's job. Plain code compares the check-in against the plan and raises short crews, late arrivals, unconfirmed prerequisites, deliveries landing after work starts, blockers, and clashes where two trades are booked into the same area at the same hour. That last one comes from no conversation at all. Every problem on the morning board carries the foreman's own words as evidence. A foreman can be reached two ways. Muster dials through Twilio, or the superintendent sends a check-in link that opens the same agent in a browser with no account and no app. Outbound dialling, the evening round, retries and voicemail are built and tested, but the project's Twilio account is a trial, and a trial silently refuses to open the Media Stream that carries live call audio. The demo therefore shows the link path, which is the same agent, the same tools and the same record.