Originally published at samiryuja.dev
I connected the official Hospitable MCP server to Claude, pointed it at my Airbnb, and asked
it to take a look. Here's what I found.
A language model is only as good as the context it's given. On its own it can talk about
short-term rentals in the abstract; it knows nothing about mine. The MCP closes that gap —
it gives the model live, authenticated access to my actual reservations, messages, reviews,
and calendar. So instead of guessing, it can look at the real thing. I run a two-room rental
I've managed for over a year, and I figured I had it under control. Two weeks of letting an
agent actually look turned up a surprising amount that was quietly broken.
Setup
- Hospitable Professional plan (the tier that includes MCP access)
- Claude Pro
- Connect them: Settings → Connectors → add custom connector → OAuth
- ~50 tools load: reservations, messages, calendar, reviews, the Knowledge Hub, tasks, smart locks
- Permissions were monitored the entire time — no messages were ever sent to guests. Everything guest-facing stayed read-only or draft-and-paste: the agent proposes, I decide.
What an AI audit of a real business actually surfaces
Here are the highlights of what it found, in four buckets.
1. Liability
The registered guest booked from overseas, but the person staying was his wife, who had no
verified profile of her own. The agent caught it by reading the message thread. That matters
because AirCover and host protections are written around the registered guest, and nothing in
the structured reservation data reveals a substitution. It only surfaces if someone actually
reads what the guest wrote.
2. Revenue opportunities
This bucket changed how I look at my inquiry inbox. Every inquiry is revenue waiting on a
response, and the agent treated it that way: it read the full inquiry history and identified
exactly where a faster or better-matched reply would have won the booking. One guest asked about
renting the whole place; by the time a reply went out hours later, those dates had gone to
another property. Another inquiry came in written in Spanish, and the automated rule answered in
English because no multilingual template existed. Across the history, it flagged six inquiries
where response time or language likely cost the conversion.
None of this is visible on any dashboard. What the agent produced was effectively a conversion
punch list: respond faster to whole-home inquiries, add a Spanish template, and answer the
recurring questions (parking, check-in) before guests have to ask. Questions are demand —
the goal is turning them into bookings.
3. Things quietly broken
The theme in this bucket: the business changed, and the documentation didn't keep up.
The big one was stale door codes. When I first set up the Airbnb, I hardcoded the entry
codes into the Knowledge Hub — the reference Hospitable's AI uses to answer guests. Later I
synced my Schlage lock with Hospitable, which generates a fresh code for every reservation. The
original setup-era codes stayed behind in seven different Knowledge Hub entries, wrong for every
guest since. If anyone had asked the AI "what's my door code?", it would have confidently handed
them a code that opens nothing.
A typo had been sitting in an active message rule for months: a small formatting error in
the review request that fires after every checkout, the final touchpoint of every stay. Easy to
fix, impossible to spot without going line by line through every rule, which no one does.
The agent also caught a time-sensitive review window. A recent guest had smoked inside and
brought unregistered visitors, and my window to review them was about to close. Once that window
shuts, future hosts get no warning, so that was the catch that mattered. It also surfaced a
backlog of guest reviews I'd never responded to — lower stakes, but review responses are public
and responsiveness factors into Airbnb's ranking, so the backlog went on the list rather than in
the trash.
Parking was in this bucket too. It's the single most common question guests ask, and the
Knowledge Hub had no car parking entry at all. The Parking section existed, but it only covered
bicycles.
4. Features I wasn't using but should be
Some of what it surfaced wasn't broken, just sitting there unused. The Knowledge Hub was thin
and partly wrong; now it's corrected and actually filled in. Tasks was completely empty, so I
registered myself as a teammate and set up text alerts for upcoming cleanings. Right now that's
redundant since it's just me. But the structure exists now, and the day I bring on a
housekeeper, I swap in their number and nothing else changes.
How I actually use it
The dramatic catches are one thing; the reason it's still connected is the boring daily stuff.
The first is complaint clustering, because it goes straight at my biggest pain point: a new
guest arrives and starts complaining, and in the moment it's hard to tell a one-off from a
pattern. I had the agent group every review across Airbnb and Booking.com by theme. The runaway
top three:
- Bedroom doors didn't lock from the inside — four guests, both rooms. The clear #1. Guests wanted keyed locks, not the privacy locks the rooms came with. The main entrance already had a Schlage deadbolt; the fix was recently replacing the bedroom privacy locks with keyed door knobs — and the clustering confirmed the complaints dropped off after the change.
- Location and transit — guests underestimating how far things were.
- Surprise at the shared-home setup — a listing-clarity issue, not a property one.
The clustering also corrected an assumption I'd been carrying: I was sure Booking.com guests
complained more than Airbnb guests. Totally wrong. The data showed the gap was a scoring-norm
artifact, not a real satisfaction difference.
The second is the weekly turnover outlook. One question and I get a clean in-chat calendar
of which days have a checkout and a check-in stacked up — the days I can't schedule anything
else. Combined with the task alerts, I stopped double-booking myself — committing to a day in
the office only to find out it was also a turnover day.
False positives and rough edges
If I only told you the good parts you shouldn't believe me. The honest column:
It flagged a problem that wasn't one. Reviewing my message templates, the agent called out
messages where the %smartlock_code% shortcode hadn't resolved. The real explanation: those
guests had already checked out, and their codes had expired with the reservation. The lesson is
that the agent doesn't automatically know the lifecycle of the data it's reading — what looks
broken on a past reservation can be perfectly normal.
Writes can fail silently. Adding an item to a Knowledge Hub that had never been initialized
succeeded with no error at all — but nothing actually appeared until I generated the hub in the
dashboard. A success response that doesn't result in the thing existing is the kind of surprise
that bites you later.
What's next
The obvious next step is automating the loop I ran by hand these two weeks — issue comes in, gets
detected, a fix gets drafted for me to approve. I'm still working out exactly what that should
look like, so I'll leave it there for now.
The part that stuck with me
What surprised me wasn't any one finding — it was the speed of the audit. Catching a
months-old typo without reading every message rule by hand. Clustering a year of reviews into
three themes in seconds. Fewer complaints, less of my time spent firefighting, more of it back.
The most useful thing an agent did here wasn't anything clever — it was simply looking,
systematically, at data I already had, and sorting it into the four things that actually matter:
liabilities, revenue opportunities, the things that had quietly snuck past me, and the
features I should have been using all along.
Built with the official Hospitable MCP server connected to Claude. Real two-room property, real
findings. Guest details kept deliberately vague.
Top comments (0)