Many people now understand what an AI keyholder can do — but not what an actual exchange looks like. What an AI keyholder is and how the MCP connection works is covered in the AI keyholder basics article; this post is about the practice. It shows three short example chats between a human keyholder and an AI assistant that the tracker is connected to over MCP. They make visible how a single sentence in a chat becomes a traceable directive — and what of that reaches the tracked user, the sub.
What to watch for in these examples
The AI assistant reads the current state from the tracker, compares it against the free-text rules the human keyholder has stored in the "AI keyholder rules (MCP)" field, and then decides on a directive within that frame. The AI does not invent rules of its own. Every action lands in the action log. The sub does not see the AI's internal reasoning — only the directive itself and its effect in the app.
Example 1: Check the status and set a lock period
Keyholder: What's the current status?
AI: Currently locked, wear time since the last opening a little over six days. No open inspection, no open offense. This week's training goal has been met.
Keyholder: Good. Set a lock period until Sunday evening — no opening before that.
AI: Done. Lock period set until Sunday, 8:00 PM. No opening is scheduled before then.
What happens in the tracker: The AI has set the "lock period" directive. In the app the sub sees a countdown to the earliest possible opening and receives a push and email notification. The action appears in the action log with a timestamp.
Example 2: Handling a late inspection according to the rules
Stored free-text rule (excerpt): "If an inspection photo arrives more than two hours late, log a minor offense and give a factual reason."
Keyholder: The last inspection came in late, didn't it?
AI: Yes. The deadline passed yesterday at 9:00 PM; the photo with the handwritten code arrived at 11:40 PM — a good two and a half hours late. Under your rule that's a minor offense.
Keyholder: Then log it.
AI: Logged as a minor offense with the note "inspection proof 2 h 40 min after deadline." I left the lock period unchanged; your rule doesn't cover that.
What happens in the tracker: The AI used the "judge offense" directive. A new entry appears in the sub's offense record and he is notified. The factual reason is part of the log — the AI's further internal notes stay hidden from the sub.
Example 3: A request that fails at a boundary
Keyholder: Just to be safe, go ahead and request a second inspection for tonight as well.
AI: That's not possible right now. There's already one inspection open with its deadline still running. Only one inspection can be open per sub at a time, so I won't create a second. Once the current one is completed or has expired, I can request a new one. Should I queue it for after that?
Keyholder: Ah, right. Yes, queue it for tomorrow morning.
AI: Noted. I'll request the next inspection tomorrow morning, as soon as the running one is done.
What happens in the tracker: The AI set no directive; it reported back a boundary — the overlap protection. For the sub nothing changes at first; he still sees only his one open inspection. Tomorrow's inspection is a scheduled directive and stays invisible to him until it actually triggers.
What this means for the sub
For the tracked user the experience stays calm and clear. In the app he sees the same things as with a purely human keyholder: a running lock period with a countdown, an open inspection with a deadline, a new entry in the offense record. He gets push and email notifications when something changes. An inspection still asks for the same thing: within the deadline, upload a photo of the handwritten five-digit code.
What the sub deliberately does not see are the AI's internal reasoning and interim judgments, along with scheduled directives that haven't triggered yet. And he can rely on the technical boundaries holding: never more than one open inspection at a time, and safety failsafes such as a health hold always take precedence over any AI directive.
Responsibility stays human
These examples show no autonomous system, but a tool with clear guardrails. The human keyholder writes the rules, can log in and intervene at any time, and every AI action is traceable in the action log. The free-text rules are soft guardrails, not a hard enforcement mechanism — the decision of how strict or how relaxed the setup is stays with the human. The MCP connection makes most sense for self-hosting setups where you run your own tracker; it is opt-in and has to be enabled deliberately.
How to activate the AI keyholder and connect an AI client is laid out step by step in the keyholder manual. The AI takes over the routine — the relationship, the trust, and the boundaries stay human.