Until now, a hotelier who wanted to use AI in the Hotel WD panel first had to get an API key from Google, OpenAI or Anthropic, set up billing, paste the key into the panel and decide which model to pick. Only then could they work in the Hotel OS chat window inside the panel, as far as that key allowed.
That chapter is closed. Hotel WD now runs on MCP (Model Context Protocol). You connect your hotel once to the AI assistant you already use, Claude or ChatGPT, and manage everything else by talking to it in plain language. No API key, no quota to watch, no model to choose.
In this article I explain why we made this decision, what changed and what stayed the same, how we built the security, and how to connect your hotel in five minutes.
Your assistant, the Hotel OS control layer and your hotel
| Topic | Before | Now (MCP) |
|---|---|---|
| Where you talk | The Hotel OS chat inside the panel | The Claude or ChatGPT you already use |
| API key | You got one and entered it in the panel | Not needed |
| AI model | Whatever model your key allowed | Your assistant's latest model |
| External connection | A URL with the secret inside it | Sign-in and approval screen, one-click disconnect |
| Irreversible actions | Approval only in the panel chat | Mandatory preview and approval on the server |
| Records | Limited | Every change is listed in your panel |
| Your hotel's rules | Loose notes | Rules and goals read to the assistant at the start of every chat |
MCP is an open protocol that lets AI assistants talk to outside systems in a safe, standard way. Anthropic released it at the end of 2024, and today many assistants and developer tools support it, Claude and ChatGPT among them.
The easiest analogy: USB-C for AI. Every device used to have its own charging cable; now one port fits them all. Before MCP, every piece of software had to build its own AI chat into itself. With MCP, the software offers a list of tools ("here is what you can do"), and it no longer matters which assistant you use.
At Hotel WD this tool layer is called Hotel OS. When your assistant connects, it does not get access to your database or your password. It sees only the roughly 90 tools we define, all locked to your hotel: "list rooms", "update the rate calendar", "create a guide post", "confirm a booking" and so on. Every tool works only on the hotel you connected, and every call is checked again on our server.
This was not about chasing a tech trend. It answers concrete problems we ran into while working with hoteliers and using the product ourselves. Here are the eight reasons, one by one.
1. API keys were never a hotelier's job
In the old setup, a hotel owner had to open a cloud account, set up billing, generate a key, keep it safe and pick the right model in the panel before the AI could do anything. When the quota ran out, the assistant went silent; when a key was pasted wrong, the error came back in technical language. None of this should take up a hotelier's day. With MCP that whole layer is gone: the assistant you connect already runs in your own account.
2. You already use a good assistant
Many hoteliers use Claude or ChatGPT every day: email drafts, menu translations, social media posts. Rather than putting a second, more limited assistant inside our panel, it makes more sense to connect the assistant you already know to your hotel. When the providers release a new model, your hotel benefits too, and we no longer have to chase model versions.
3. Not a chat window, a working partner
The in-panel assistant could only see the panel. In Claude or ChatGPT you can add an Excel report, a screenshot of a competitor's prices or a CSV exported from Google Ads to the same conversation, then ask the assistant to act on your hotel based on it. A request like "Review our weekday rates against this report and show me your proposal" now starts and finishes in a single conversation.
4. Security: the secret does not live in the address bar
In our previous external connection method, the access key sat inside the connection URL; anyone who saw the URL could reach the hotel. The new system connects with OAuth 2.1, the standard used by banks and large platforms:
5. You stay in control: approval for anything irreversible
This is the most important part of the system. Some actions cannot be undone: deleting a room, cancelling a booking (which starts a real refund if a payment was taken), overwriting every rate in a date range, sending an email to a guest.
For these actions your assistant cannot do anything on the first try. Our server first returns a real preview of the record that would be affected: which room, which booking, which date range. The assistant shows it to you, and only when you approve is the action carried out, with an approval code that is tied to that exact action and valid for a short time. If the assistant changes the arguments (another room, other dates), a new approval is required.
This two-step flow is not an instruction to the assistant; it is enforced on our server: the preview step cannot be skipped, and the approval code works once, only for the action that was shown. It is your assistant that shows you the preview and asks for your OK; in Claude and ChatGPT you can also keep "ask before every tool call" switched on. Actions that go through this step:
Adding and editing are free: your assistant can write posts, update room descriptions and prepare drafts. You can still see and correct everything in the panel.
6. Everything is on record
Every change your assistant makes appears on the Hotel OS page of your panel, in the "What your assistant did" list: when, through which assistant, which action and what the result was. Only owners and managers see the details that were sent and any error messages, because an email action may include a guest's address. Requests that only read data are not listed, so the list stays readable.
7. An open standard: you are not tied to one company
The old setup was tied to specific providers' keys and models. MCP is an open standard: Claude and ChatGPT today, another assistant tomorrow. Whichever one you prefer connects to your hotel under the same security rules. The choice is yours.
8. No more AI calling another AI
Buttons like "Generate SEO plan" in the old panel called a second AI model in the background, using your key. So while you were talking to one assistant, it was handing work to another model: double the cost, double the wait, inconsistent results. Now the assistant you are talking to writes the plan and Hotel WD simply saves it. The Marketing Hub screens are where they were; plans land there and you tick off the tasks.
What was removed
What stayed the same
Once connected, your assistant can use roughly 90 tools on your hotel:
A few example requests:
In the first example your assistant does not change the rates right away: it shows which room and which nights would be affected and waits for your approval.
Ready-made tasks
For the jobs you do most often, your assistant also gets ready-made workflows: morning briefing, weekend rate suggestion, SEO check-up, guide post, neighborhood guides, Google Ads roadmap, SEO plan, email marketing plan and keyword report review. In claude.ai, pick them from the + menu in the chat box under Hotel OS; in Claude Code, type /. Each one reads your hotel's rules first; marketing plans are saved straight to your Marketing Hub and guide posts are saved as drafts. You can see the full list in the Ready-made tasks section of the Hotel OS page in your panel.
Every hotel has unwritten rules: "the Swim-Up room never goes below €120", "guests are always addressed formally", "the family room sells for at least two nights". In the new system you save these as rules. You add measurable targets such as "October occupancy 85%" or "40% share of direct bookings" as goals.
Your assistant reads these rules and active goals at the start of every new conversation. If a request conflicts with one of your rules, it warns you before doing anything. Your goals steer the plans, posts and rate suggestions it prepares.
You can add rules in two ways: in the Rules and goals section of the Hotel OS page in your panel, or by telling the assistant "save this as a rule" in the chat. Adding a rule and removing one both go through the approval step, so text from outside, a guest review for example, cannot quietly make your assistant save a permanent rule.
First, open the AI Assistant (Hotel OS) page from the left menu of your panel and copy the connection address. It looks like this:
https://www.hotelwd.com/api/mcp/hotel/<your-hotel-id>Each connection belongs to a single hotel. If you have more than one hotel, add a separate connection for each, so one hotel's assistant can never see another hotel's data.
claude.ai (web and desktop)
Claude Code (terminal and VS Code)
Run this command in your terminal:
claude mcp add --transport http hotel-os https://www.hotelwd.com/api/mcp/hotel/<your-hotel-id>Then type /mcp in your Claude Code session, select the hotel-os connection, sign in and approve in the browser.
ChatGPT
In ChatGPT this option may depend on your account's plan. The assistants also rename their menus from time to time; we keep the How to connect guide in the panel up to date.
Can the assistant see another hotel's data?
No. Each connection is bound to the single hotel you approved, and our server checks this again on every call.
Where are my conversations stored?
Your conversations are stored in your Claude or ChatGPT account under that service's privacy terms. Hotel WD does not see your conversations; we only record the changes the assistant makes to your hotel.
How do I disconnect?
On the Hotel OS page of your panel, click Disconnect next to the assistant in the Connected assistants list. Its access ends immediately.
Who on my team can connect?
Owners and team members with the manager role can approve a connection. A connection depends on the authority of the person who approved it: if that person leaves the team or is downgraded, their connection stops by itself. Connections unused for 90 days also expire. Roles with Hotel OS access can see the list of what the assistant did; only owners and managers see the details that were sent.
What if the assistant does something wrong?
Irreversible actions cannot be done in one step; your assistant first shows the preview and asks for your OK. Other changes appear in the "What your assistant did" list and can always be corrected in the panel.
What happens to my old connection address?
Old-style connections with the secret inside the URL keep working during a transition period, and you will see a warning in the panel. Once you have connected with the new address, revoke the old one on the same page.
Is my guest chatbot affected?
No. The chatbot on your website keeps running on your own AI key, as before.
With this switch we did not give up on AI; we changed how you reach it. We removed the hassle of keys, quotas and models. Security moved to a sign-in and approval screen, control to the approval gate, and transparency to the record in your panel. The assistant is yours, the rules are yours, and the final approval is yours.
To get started, open the AI Assistant page in your panel. If you get stuck anywhere, our experts are here to help.
AI does not replace a hotelier; it gives the hotelier their time back.
Tahir Dinç
I have been working in the digital field since 1999. I still hold the position of Digital Marketing Manager at Türk SEM. I have also been involved in tourism-related activities since 2005.