<?xml version="1.0" encoding="utf-8"?><feed xmlns="http://www.w3.org/2005/Atom" xml:lang="en"><generator uri="https://jekyllrb.com/" version="4.4.1">Jekyll</generator><link href="https://icecreek.com/feed.xml" rel="self" type="application/atom+xml" /><link href="https://icecreek.com/" rel="alternate" type="text/html" hreflang="en" /><updated>2026-11-11T22:34:33+01:00</updated><id>https://icecreek.com/feed.xml</id><title type="html">icecreek</title><subtitle>One clean interface for hospitality tech vendors, the PMS connector long tail handled underneath. By Christian Ey.</subtitle><author><name>Christian Ey</name><email>hello@icecreek.com</email></author><entry xml:lang="en"><title type="html">Build diary #1: three weeks on the first connector</title><link href="https://icecreek.com/blog/2026/11/three-weeks-on-the-first-connector/" rel="alternate" type="text/html" title="Build diary #1: three weeks on the first connector" /><published>2026-11-10T12:00:00+01:00</published><updated>2026-11-10T12:00:00+01:00</updated><id>https://icecreek.com/blog/2026/11/three-weeks-on-the-first-connector</id><content type="html" xml:base="https://icecreek.com/blog/2026/11/three-weeks-on-the-first-connector/"><![CDATA[<p>Three weeks ago I started the first connector of the integration layer, against <a href="https://apaleo.com">apaleo</a>. The read path now runs end to end against apaleo’s sandbox. The event feed is merged, the write path is in this week’s night runs, and the target is a running demo by the end of November. This is the first entry of a weekly build diary: what shipped, where the API pushed back, what I learned. It is honestly unfinished, on purpose.</p>

<h2 id="why-apaleo-first">Why apaleo first</h2>

<p>apaleo sits at Tier 1 on our <a href="/pms-tier-map/">PMS Tier Map</a>: a free developer account you open yourself, a sandbox you can reset, and OAuth access that the hotel grants to an app. The connector writes data, so a free, resettable sandbox was a hard requirement.</p>

<p>The honest flip side: an open PMS is where an integration layer adds the least. A good team can build against apaleo alone. Connector one is there to harden the interface on the friendliest door. The real test is connector two, at a system that works very differently.</p>

<h2 id="what-apaleo-does-on-its-own-and-how-i-know">What apaleo does on its own, and how I know</h2>

<p>Everything below comes from one of two places, and I label which: apaleo’s <strong><a href="https://apaleo.dev">public documentation</a></strong> (developer guides and Swagger), or my own experiments in the <strong>sandbox</strong>. Sandbox results are observations on one test hotel, not documented behavior.</p>

<p><strong>Check-in makes the room dirty</strong> (sandbox). Rooms turn dirty at checkout. In four out of four check-ins, a clean room also turned dirty within seconds. I found no public document describing this. If it is documented somewhere, tell me and I will correct this. Not a flaw, a behavior to plan for. It matters for any housekeeping product: a room that just went dirty has not necessarily been used, and your product needs to know that the PMS did this on its own.</p>

<p><strong>The night audit dirties occupied rooms</strong> (public documentation, consistent with the sandbox). The Swagger description says so, and a room I set clean in the afternoon was dirty the next morning, with nobody touching it. When the audit runs is part of the hotel’s own automation settings, readable through a separate Automation API that needs its own scope. Until we use it, the platform says the audit time is unknown instead of guessing.</p>

<p><strong>There is no business date to read</strong> (public documentation). I could not find one in the public API, so the platform computes the calendar date in the property’s time zone and labels it as computed. Between midnight and the night audit it can be a day ahead of the hotel’s own date, and the label says so.</p>

<p><strong>Rate limits belong to the app and the account, not the hotel</strong> (public documentation). One apaleo account can hold many hotels, and they all share one budget per app: 5 requests per second and 200 per minute on development and trial accounts. So one connection on our side is one apaleo account. At one point a full connect spent exactly five calls, the entire per-second budget of a trial account. The docs also say that tokens must be cached; doing that cut a connect from six calls to five and a live read from two calls to one.</p>

<p><strong>A throttled request is not processed</strong> (public documentation). That sounds minor and is not: a write that comes back throttled was not applied, so it is safe to retry. Precise statements like this one save whole design discussions.</p>

<p><strong>Small things that shape a design:</strong></p>

<ul>
  <li>The unit list carries the current maintenance without its dates; only the single unit read has them (public documentation). Read naively, alternating list and single reads would report a change on every room under maintenance where none happened.</li>
  <li>The unit list leaves out the unit group’s name unless you ask for it (public documentation).</li>
  <li>Rooms carry no version or modification time (sandbox), so the platform keeps its own version to detect conflicting writes.</li>
  <li>Every transition between clean, to be inspected and dirty is accepted, in any order (sandbox). There is no enforced workflow in the PMS. The workflow belongs to the product on top.</li>
</ul>

<p>None of this is a complaint. Every answer above came from public documentation or a free sandbox, without a call, an NDA or a fee. That is exactly what makes apaleo the right first door: the behaviors are findable.</p>

<h2 id="building-ai-native-as-a-solo-founder">Building AI-native as a solo founder</h2>

<p>AI agents write most of the code. I design, review and operate. The loop:</p>

<ul>
  <li><strong>One issue, one build session.</strong> A coding agent (Claude Code) works one issue on its own branch.</li>
  <li><strong>A separate review agent</strong> checks every diff against the written architecture decisions, at most two rounds.</li>
  <li><strong>I decide a short list only:</strong> the public interface, architecture decisions, database migrations, security, PMS app registrations, dependencies and money. And I merge.</li>
</ul>

<p><strong>Night runs.</strong> In the evening an orchestrator agent gets a planned list of issues. It delegates each to a builder, sends every result through the reviewer, and merges into a night branch, using its own account, which cannot touch the main branch. Small open questions it decides in my name, but only by taking the recommended, reversible option and logging the alternatives and how to undo it. In the morning I review one merge request. On the first full night, 17 of 18 planned issues landed. The second night produced ten merge requests in about two hours.</p>

<p>What went wrong, in the first nights:</p>

<ul>
  <li>My own safety hook raised five false alarms. Each one stalled the run until I answered.</li>
  <li>One agent read a refused command as a “no” and stopped, instead of retrying once. Its work was finished and green, but never submitted.</li>
  <li>Two branches with green pipelines did not compile together. A trial merge before every merge now catches that.</li>
  <li>Two agents briefly wrote into the main working copy instead of their own.</li>
</ul>

<p>Each of these became a line in the next night’s brief.</p>

<p><strong>Night runs cost money, too.</strong> Agents run the full build and test suite constantly, and on hosted CI minutes a single night run cost about 5 to 10 euros. The builds now run on a small server I already had, for about 4 euros a month. The switch paid off on day one in another way: the tests had passed only because the hosted runners happened to run them in one particular order. On the new machine the order changed and a hidden dependency between tests showed up. Fixed the same morning.</p>

<p><strong>The biggest lesson: agents are tireless at closing gaps nobody has hit.</strong> By my rough count, about a quarter of the read-path time went into tightening our own automated architecture checks against cases no real change had shown. Thorough, and the wrong priority three weeks before a demo.</p>

<p>So the project now has a <strong>decision charter</strong>, one document every session reads first: where we are going (a running end-to-end demo by the end of November), a priority order (invariants, a working path, a clear contract, simplicity, elegance last), what sessions decide on their own, at most five questions per round for me, and stop rules. A new rule only for a concrete defect that a diff actually shows. Tooling frozen until the demo. Merge requests under 400 lines.</p>

<p><strong>The numbers so far.</strong> Agents logged about 89 hours across the three weeks (parallel sessions counted as their sum). My own time, counted conservatively from the timestamps of my inputs, comes to about 27 hours. Agents make the code cheap. What they do not make cheap is getting through the doors, and keeping every connection working once it runs. That is not a productivity claim. The number that matters comes later: how long a housekeeping service takes to build on the layer, compared with building it directly against a PMS. Meanwhile, the public interface has not needed a single new endpoint since the first milestone.</p>

<h2 id="what-one-interface-for-every-pms-honestly-means">What “one interface for every PMS” honestly means</h2>

<p>It does not mean every PMS behaves the same. On a single PMS, behavior already depends on the hotel: automatic check-in on or off, when the night audit runs, what a room’s status is after maintenance. Three hotels on the same system can be three configurations. So the promise has to be more precise:</p>

<ul>
  <li><strong>Same meaning, not same behavior.</strong> “Clean” means the same thing on every connection. But the layer does not suppress what a PMS does on its own, and it does not fake what a PMS cannot do. If apaleo turns a room dirty at check-in, your product sees that, marked as the PMS’s doing.</li>
  <li><strong>Ask, do not assume.</strong> Every connection states what it can do, how we know (documented, tested, or observed on that connection), and what is still unknown. “Unknown” is a valid answer. Capability discovery is not a feature of the layer, it is the product.</li>
  <li><strong>The process stays with you.</strong> Whether a hotel inspects rooms before release, who gets which task, what counts as urgent: that is your product’s job, and it differs per hotel. The layer carries the facts.</li>
  <li><strong>The value grows with the doors.</strong> At apaleo, a strong team can do this alone. The layer earns its keep on the second system, the fifth, the twelfth, and in maintaining all of them once they run.</li>
</ul>

<h2 id="coming-up">Coming up</h2>

<p>The write path (setting a room clean from outside the PMS), recognizing your own writes in the event feed, then change detection, the connect flow, and a reference housekeeping app as the layer’s first client.</p>

<p>If you are a hospitality tech vendor and a PMS integration is stalling a deal right now, I would like to hear which one: <a href="mailto:hello@icecreek.com">hello@icecreek.com</a>. Early access when it runs goes to the people who write first.</p>]]></content><author><name>Christian Ey</name></author><summary type="html"><![CDATA[Three weeks into the first PMS connector (apaleo), built solo with AI agents: what the PMS does on its own, how night runs and a review agent work, and what one interface for every PMS honestly means.]]></summary></entry></feed>