AI MLS Automation: How I Look Up Showing Instructions and Schedule Tours Hands-Free

My client sent me a list of properties to tour. Normally that means 30 minutes of clicking through the MLS one by one — pulling showing instructions, checking lock box codes, figuring out which ones are owner-occupied versus vacant, removing the ones that don’t qualify, and then planning a route. Instead, I had my AI executive assistant do all of it hands-free. This is what AI MLS automation actually looks like for a practicing real estate agent — not a concept demo, but a real showing day.

This is Episode 2 of the “AI EA for Real Estate” series, where I’m mapping AI to every step of a real estate agent’s workflow using Claude Code. In Episode 1, I showed how my AI assistant Ashley handles my morning briefing — CRM pipeline, YouTube analytics, and daily task list. This time, the task is more hands-on: navigating the actual MLS through browser control, pulling data from live listings, and building an optimized showing route.

If you’ve ever wished you had a transaction coordinator or assistant who could just handle the MLS lookup grunt work, this is what that looks like with AI. Claude Code is one of several AI tools for real estate agents I’ve tested, and for MLS workflows, nothing else comes close.

How AI MLS Automation Works with Claude Code

Claude Code has a feature called computer use — it can control a browser the same way you would. Click buttons, type into fields, scroll through pages, read what’s on screen. That’s what makes this possible. I’m not copy-pasting data into a chatbot. The AI is literally navigating Unlock MLS on its own.

Here’s what the workflow looked like for 7 properties:

1. I gave Claude the MLS listing IDs my client sent me
2. Claude opened Unlock MLS in the browser, entered each listing ID, and navigated to the showing instructions
3. For each property, it pulled: showing instructions, lock box type (Supra vs combo), contact info, owner/tenant occupancy status, and any special notes
4. It compiled everything into a single summary I could review in one place

The whole process took about 15 minutes. Not instant — browser automation is slower than an API call would be. But I wasn’t doing any of it. I stepped away, came back, and had a full breakdown of every property ready to go.

Teaching AI to Search by Address (Not Just MLS Number)

Not every client sends MLS numbers. Half the time I get a screenshot of a Zillow listing or just an address in a text message. So after the initial MLS-number lookup worked, I tested whether Claude could also search properties by street address.

This is where it got interesting. Claude entered the street number and street name into the MLS search fields — and it worked on the first property. But on the second one, it didn’t clear the previous search. It concatenated the old street name with the new one, which obviously returned zero results.

Here’s the part I didn’t expect: Claude recognized the error on its own. It saw the concatenated search terms, understood what went wrong, and looked for a way to fix it. It found a “Clear” button at the bottom left of the MLS search interface — a button I didn’t even know existed. I’ve been using this MLS for years and never noticed it.

That’s one of the underrated benefits of AI agents navigating software. They read every element on the screen without the shortcuts and habits we develop as humans. Sometimes they find features the software developers built that we never learned to use.

Filtering Out Properties That Don’t Qualify

My client’s list included some properties that wouldn’t work. A few were expired or withdrawn. Three were affordable housing listings that require the buyer to meet income thresholds my client doesn’t qualify for. One was already under contract.

Instead of manually going through each one, I told Claude: “Remove the ones no longer available and remove the affordable housing listings.” It filtered the list down to 7 active, available properties — exactly what I needed.

This kind of filtering takes 5 minutes by hand. Not a big deal for one showing trip. But across dozens of buyer tours per year, those small tasks stack up. More importantly, it’s the kind of work that breaks your focus when you’re trying to coordinate a full day of showings.

Route Optimization with Google Maps

Once I had the final list of showable properties, the next step was planning the tour route. I told Claude to optimize the route for the vacant properties starting from East Austin (where I’d be meeting my client).

Claude opened Google Maps, entered all four vacant property addresses, and built a route. The order it chose made geographic sense — it wasn’t just listing them in the order I gave them. It optimized for shortest driving distance between stops.

The route looked solid: Stop 1 to Stop 2 to Stop 3 to Stop 4, moving efficiently across the area. For the occupied properties that need appointments, I’d schedule those separately for the next day once I confirmed times with the listing agents.

Lock Box and Supra Access Readiness

Even for “go showings” where no appointment is required, you still need to confirm the lock box situation. A Supra box means I can access it with my Supra eKEY app — no calls needed. A combo lock means I need to reach out to the listing agent or schedule through ShowingTime to get the code.

Claude checked each property’s lock box type from the showing instructions it had already pulled. Two properties had Supra boxes — good to go. One had a gate code I’d need. Another had remarks that didn’t confirm Supra access, so I flagged it for a manual check.

This readiness check is exactly the kind of detail that gets missed when you’re rushing to set up a showing day. Having AI compile it into a checklist — with action items for which properties need follow-up — means nothing slips through.

Building Reusable Skills with Skill Creator

After the session, I used Claude Code’s Skill Creator to save everything the AI learned into a reusable skill. The MLS lookup skill now includes:

• MLS number search — enter listing IDs and pull showing instructions
• Address-based search — search by street number and name, with proper field clearing
• Route optimization — plan showing routes via Google Maps from a starting direction
• Lock box readiness check — flag which properties need Supra, combo codes, or agent outreach
• Showing readiness action items — a summary of what’s ready to go and what needs follow-up

Next time a client sends me a list of properties, I run one command. The AI already knows how to navigate Unlock MLS, what fields to check, how to handle address searches, and how to build the route. Each session gets faster because the skill improves over time.

Short-Term vs Long-Term Memory: Why It Matters

One thing I noticed during this session: Claude performs best when the full context of what it’s done is still in the conversation. I stepped away for a couple hours mid-session to show a home, and when I came back, the reconnection was slightly slower because it had to reload context.

This is the difference between short-term and long-term memory in AI agents. Short-term memory is everything in the current session — every click, every error, every correction. Long-term memory is the skills and files saved to disk. When I used Skill Creator to save the MLS lookup skill, I was moving short-term learnings into long-term memory so future sessions start with that knowledge baked in.

Think of it like training an employee. The first day, they’re learning everything in real time. At the end of the day, you document the SOPs so they don’t start from scratch tomorrow. That’s exactly what Skill Creator does for AI agents.

Model Routing: When to Use Haiku vs Sonnet vs Opus

I used Opus for this session because it was the first time teaching the AI a complex new workflow. Opus is the most capable model — think of it as your graduate-level employee. For initial skill building where the AI needs to figure out navigation, handle errors, and learn from corrections, Opus is worth the cost.

Once the skill is built and saved, future runs can use Sonnet (college-level) or even Haiku (high school-level) for the routine execution. The heavy thinking already happened. Running a known skill through the MLS doesn’t need PhD-level reasoning — it needs reliable execution of documented steps.

This model routing strategy keeps token costs down. Use the expensive model to teach, then the cheaper models to execute.

What’s Next: CRM Texting Integration

The showing instructions include listing agent phone numbers. The natural next step is connecting Claude to my CRM (Follow Up Boss) so it can send texts to listing agents directly — “This is William Zhang with eXp Realty, I’d like to schedule a showing for tomorrow at 2 PM. Can I get the access code?”

That’s coming in a future episode. The goal is a single command that goes from “client sends property list” to “showings scheduled, route planned, lock boxes confirmed” without me touching the MLS, Google Maps, or my phone.

If you want the MLS Lookup and Route Optimization skills from this video, they’re free to download — grab them from the newsletter signup. Every episode in this series includes the skill files so you can plug them into your own Claude Code setup.

Catch Episode 1 if you haven’t seen how the AI executive assistant is set up from scratch. And subscribe to the newsletter to get notified when the CRM texting integration episode drops.