What Can an Agent Actually Do with MCP?

I have been building a travel AI agent for the past few months, which meant spending a lot of time evaluating the Travel MCP. I have tested what actually works — so you do not have to repeat the same experiments.

Last week I was debugging a trip-planning agent. Before connecting any MCP, it could only chat — the user would say "book me a hotel," and it would respond with a wall of suggestions but could not actually book.

After connecting to MCP, I threw the same prompt at the agent. It came back with a real table:

Agent response after MCP connection

`+--------------------------------------+----------+--------+------+----------+
| Hotel                                | Location | Price  | Star | Metro    |
+--------------------------------------+----------+--------+------+----------+
| Hotel Monterey Grasmere Osaka        | Yodoyabashi| $687/n| 4★  | 3 min    |
| Smile Hotel Namba                    | Shinsaibashi| $542/n| 3★  | 7 min    |
+--------------------------------------+----------+--------+------+----------+`

The agent did not just search — it figured out the prompt context and translated it into structured query parameters on its own and then give the link that could be booked.

That was the moment I realized: the way you connect tools to an agent does not just add capability — it changes the level at which the agent operates.

The traditional approach for connecting to hotel data is to call a search API, get a pile of results, and have the LLM read through them to make recommendations. The problem is that a general-purpose LLM has no idea about real-time prices or actual availability.

After connecting to MCP, I sent the agent this prompt:

"Osaka, 5 nights, budget 500–700, near a subway station, free cancellation policy"

The agent broke the request into three search parameters on its own:

• Price range: 500–700
• Location: within 500m of a subway station
• Cancellation: free cancellation only

Then it queried real available hotels in Osaka via MCP, filtered out anything that did not match, and returned exactly three results — all bookable, all meeting every criterion.

Price Comparison — The Agent Picks the Best Option

This is where I felt the MCP connection paid for itself.

I asked the agent to search "Paris, June 20–22, five-star, within 3 km of Eiffel Tower." It called the MCP and got back a batch of hotels. But that was not the end — it then did four things:

1. Sorted all results by total price including tax
2. Filtered for free-cancellation options (reducing user decision risk)
3. Annotated each hotel's cancellation deadline
4. Provided a recommendation rationale: "This one is 23% cheaper than other five-star hotels in the same area, and has the highest review count"

This is not simple price comparison. The agent needs to understand what actually blocks a user's decision — free cancellation, review volume, distance. If the MCP does not return these fields, the agent can only guess. Without real inventory and pricing data, "the agent helps you decide" is a castle in the air.

Multi-Tool Chain — Agent Strings Together Flight + Hotel

This was the most complex scenario I ran, and the closest to what I would call a "real agent."

I sent the agent this message:

"Business trip to Beijing next Wednesday through Friday. Check flights and hotels near CBD, keep total budget under 2000."

It executed in three steps:

Step 1 — Called Flight MCP to search Hangzhou-to-Beijing flights for June 17, filtering out anything over budget.

Step 2 — Called Hotel MCP to search Beijing CBD hotels, checking which ones still fit within the remaining budget.

Step 3 — Combined both results and generated an optimal flight-plus-hotel package.

Notably, the agent never asked me for confirmation during this process. It judged on its own: for a business trip, location matters more than star rating, so it prioritized proximity to CBD over five-star luxury.

Can a general-purpose LLM do this? Not really — because it does not know real-time flight prices or hotel availability. Those are exactly the gaps that MCP fills. But the pity is RollingGo Flight MCP is not working now, and it needs time to repair. Hotel MCP works well.

Why MCP Makes Agents Go From "Chatting" to "Doing"

At this point you might ask: I understand the principle, but why does it have to be MCP specifically?

My take: the LLM is a brain. It is good at reasoning and decision-making, but it does not know what is happening in the world right now. MCP gives that brain eyes and hands.

• Eyes — RollingGo Hotel MCP real-time visibility into 2M+ hotels globally, with actual selling prices and availability status
• Hands — the ability to execute search, filter, compare, and lock actions directly

Without MCP, the agent is a daydreamer — beautiful plans, zero execution. With MCP, it becomes an operator that knows the current state, understands the request, produces a plan, and acts on it.

Configuration

Configuration was not entirely smooth. A few things tripped me up:

Gotcha 1: MCP tools do not appear after configuration. I spent way too long debugging before realizing the client simply needed a restart. After restarting, the tool list refreshed and everything showed up.

Gotcha 2: 401 Unauthorized. There was an extra space after "Bearer" in the authorization header. The correct format is Bearer mcp_xxx — no spaces between "Bearer" and the key.

Gotcha 3: Search returns empty. Most likely the dates were set in the past. The MCP validates strictly — date parameters must be future dates.

Who Should Use This?

Honestly, not everyone needs it. This combination is a good fit if:

• You are building a travel-direction agent and need a real data source
• You want AI to directly help users complete flight/hotel queries and recommendations, not just give suggestions
• You are a technical team — the integration barrier is low, API keys are self-service, no commercial onboarding needed

If your agent only needs general knowledge Q&A without real-time data, MCP will not add much value for you.

GitHub: https://github.com/RollingGo-AI/RollingGo-Hotel-MCP-Global

Final Thoughts

After a few months of testing, my biggest takeaway is that MCP opens a new possibility for agent development — moving beyond "question answering" toward agents that genuinely do things. RollingGo Hotel MCP with 200M+ hotels and 500+ suppliers, the inventory is real and bookable — not a test endpoint returning fake data. For travel agent development, it is more than sufficient.