Set up your own 24/7 travel deal agent in 5 minutes

Published September 9, 2026

It scans flight deals for you around the clock, on the routes you actually care about, and tells you when a fare is cheap by Google's own standard rather than by a number you guessed. It costs less than any human travel agent, and the setup below is one URL and one sign-in.

You do not need an API key you have hand-copied anywhere, and you do not need to write code. If you already use Claude, Claude Code, Cursor or ChatGPT, you have everything else.

What you are actually connecting

One MCP server, four tools: search_oneway_flights, search_roundtrip_flights, search_hotels, find_hotel_by_name. Every flight result comes back with the fare, a buy_link that opens that exact itinerary on Google Flights, and Google's own price history for that route and date as price_insights_low, price_insights_high and a price_range_in_relation_to_other_periods verdict of low, typical or high.

That last part is why this takes five minutes instead of five months. You do not have to log prices for a season before your first useful alert. The band ships with the fare.

Step 1: pick free or keyed

Two servers, and the choice takes ten seconds.

Free, https://google-flights-lulu.flightpowers.com/mcp. No RapidAPI key and nothing to subscribe to: you sign in with Google and that is it. Every successful result carries one labelled sponsored card, that is how it pays for itself, and it allows 150 searches a day and 2,000 a calendar month under your own account. One call is one search per date and destination combination, so a wide scan spends more than one. It also samples any single call down to 15 combinations and tells you which ones it actually ran in search_coverage. Great for feeling the whole loop before you commit to anything. Not the thing to point a 30-destination daily scan at.

Keyed, https://flights.flightpowers.com/mcp. Same shape: you sign in with Google, paste a RapidAPI key once on the page it opens, and after that your client just works. No ads, no daily ceiling, and the only limit is your own plan. Subscribe on the Google Flights Live API listing: BASIC is free with 10 requests a month, PRO is $10 a month with 2,500. Hotels are a separate listing, the Booking Live API, on the same key. The honest read: BASIC proves your key works and nothing more, a daily scan of 30 destinations is roughly 900 requests a month, so the agent you actually keep runs on PRO.

Everything below uses the keyed URL. Swap in the free URL if you just want to poke at it: the steps are identical, you just have no key to paste at the end.

Step 2: add it to your client

Pick your one. The flights URL is https://flights.flightpowers.com/mcp and the hotels URL is https://hotels.flightpowers.com/mcp. Same sign-in covers both.

Claude (claude.ai and the desktop app)

Connectors live on your account, so adding one in the browser puts it in the desktop app too.

1. Open Customize > Connectors, then click Add custom connector. On Team and Enterprise an owner adds it under Organization settings > Connectors first, and everyone else then presses Connect on it.
2. Name it FlightPowers flights and paste the URL:

https://flights.flightpowers.com/mcp

3. Leave Authentication on what Claude detects. This server answers an unauthenticated request with a 401 and a WWW-Authenticate header pointing at its own metadata, which is exactly the handshake Claude looks for, so it fills the auth settings in by itself.
4. For OAuth client, either hosted client metadata or "register one automatically" works. The server supports both, so there is nothing for you to paste in and no client secret to find.
5. Click Add, then Connect. Google asks for your account id and email, nothing else.
6. In a chat, click + > Connectors and toggle it on. If Claude answers a fare question instantly without showing a tool call, that toggle is off. It is the most common miss by a distance.

Claude Code

One line, and you can run it right now:

claude mcp add --transport http flights https://flights.flightpowers.com/mcp

Then open /mcp in a session and sign in. claude mcp list prints ✔ Connected when it worked and ! Needs authentication when the sign-in has not happened yet. Add hotels the same way with https://hotels.flightpowers.com/mcp.

If you would rather not sign in at all, send your key in a header on the same URL. This is the form to use for anything non-interactive, because a headless run cannot open a browser to complete a sign-in:

claude mcp add --transport http flights https://flights.flightpowers.com/mcp \
  -H "x-rapidapi-key: YOUR_RAPIDAPI_KEY"

Cursor

Add a block to ~/.cursor/mcp.json for every project, or .cursor/mcp.json for one:

{
  "mcpServers": {
    "flights": { "url": "https://flights.flightpowers.com/mcp" },
    "hotels": { "url": "https://hotels.flightpowers.com/mcp" }
  }
}

Restart Cursor. It hits the 401, registers itself, and offers you the sign-in. The key-in-a-header form works here too, same shape as the Claude Code example, on the same URL:

{
  "mcpServers": {
    "flights": {
      "url": "https://flights.flightpowers.com/mcp",
      "headers": { "x-rapidapi-key": "YOUR_RAPIDAPI_KEY" }
    }
  }
}

ChatGPT

Custom connectors in ChatGPT need developer mode, which OpenAI puts on the Pro plan and on Business, Enterprise and Edu workspaces, and it is web only. On Free and Plus the option is genuinely not there. Turn it on under Settings > Apps > Advanced settings, create a connector, paste the same https://flights.flightpowers.com/mcp URL, set authentication to OAuth, and leave the client id and secret empty: the server registers ChatGPT for you.

Being straight about the state of this one: we have no paid ChatGPT account, so we have not run this end to end ourselves. The server side is the same code that Claude, Claude Code and Cursor sign into, and it is standards-plain, but if your connector refuses to attach, that is the half we cannot debug for you. The key-on-the-URL form, https://flights.flightpowers.com/mcp?rapidapi_key=YOUR_RAPIDAPI_KEY with authentication set to none, is the fallback that needs no OAuth at all. More detail on that path in the ChatGPT MCP guide.

Step 3: the prompt

This is the whole agent. Swap TLV for your airport and the dates for yours, and paste it in.

Use search_roundtrip_flights ONCE, not once per city.

from_airport: TLV
to_airport: FCO,ATH,CDG,BCN,MAD,LIS,AMS,BRU,BER,MUC,FRA,VIE,ZRH,PRG,BUD,
WAW,OTP,SOF,IST,LCA,MXP,VCE,NAP,CTA,TBS,EVN,GYD,DXB,LHR,SKG
departure_date: 2026-10-17
nights: 4
currency: usd
sort_by: price
limit: 5
max_searches: 30

Do not set max_price. It removes the price band, and the band is the
whole point.

Then read by_destination, which has one entry per city I asked for,
empty ones included. For each city take the cheapest fare and compare
it to that city's own price_insights_low.

Report only the cities whose cheapest fare is BELOW their own
price_insights_low, cheapest first, each with the fare, the
low-to-high band, the verdict, and the buy_link.

If none is below its low end, say so in one line and give me the single
cheapest fare anyway.

Check search_status first. A city that came back empty is unknown, not
"no flights there" - say which ones those were.

Thirty destinations on one departure date is thirty billed requests, and max_searches: 30 is the ceiling you are agreeing to. Want a date range instead of one date? Drop departure_date and use departure_date_from and departure_date_to, and raise max_searches deliberately, because every extra date multiplies the count: 30 cities across 3 days is 90 requests, not 30. Ask for more than max_searches allows and the request is not refused, it is sampled evenly, and search_coverage lists exactly which dates and cities it looked at. Read that before you conclude anything about a city.

nights also takes a list, so nights: [3, 4, 5] compares trip lengths in the same call, at the same multiplier.

What a run looks like

The same scan, run through the REST endpoint on 2026-09-09 at 13:18 UTC out of Tel Aviv, departing Saturday 17 October and back on the 21st:

Every figure there is from that one run. Fares move within minutes, so yours will not match, which is the entire reason for running it daily rather than reading a screenshot.

One thing not to over-read: three fares in that run came back above the high end of their own band and Google still labelled them typical. So the verdict is not simply arithmetic on the band. Alert on the comparison you can do yourself, fare against price_insights_low, and treat the verdict as a second opinion.

Step 4: make it daily

The scan above is the agent. Running it on a schedule is what makes it 24/7.

• Claude scheduled tasks. Any paid Claude plan, they run on Anthropic's servers with your laptop shut, and they can use your connectors. Full click path, the cadence settings, and a worked example are in Create your 24/7 AI travel agent.
• Claude Code Desktop routines. The Code tab has a Routines page: New routine, Local for a task that runs on your machine, or a cloud routine that runs without it. Local tasks only fire while the app is open and the machine is awake, which is the trade.
• A plain cron line, if you would rather own it. claude -p "<the prompt above>" in a crontab entry is the whole thing. One catch worth knowing before you write it: a non-interactive run cannot walk you through an OAuth sign-in, so either sign in once from an interactive session first, or configure that server with the key-in-a-header form from Step 2 and skip sign-in entirely. The header form is the one I would use for cron.

If none of those fit your setup, run it each morning by hand. It is one paste, and it still beats checking six tabs.

What it will not do

• It cannot book. It finds the fare, judges it, and hands you the buy_link. You click.
• The verdict can be null on thin routes where Google publishes no band. That means no signal, not "not cheap". On those, fall back to a plain price threshold.
• A scan can come back partial. search_status says ok or partial and search_coverage lists what actually ran. The prompt above makes the agent surface it, which is why that paragraph is in there.
• Searches take seconds, not milliseconds. It is a live read of Google Flights, not a cache.
• Flights and hotels are separate subscriptions. One key, two listings. A 403 on the hotel tools almost always means the key is subscribed to flights only.

Related

• Create your 24/7 AI travel agent: the long version, with the scheduling paths and the cost math in full
• Cheapest destinations from one airport: the same expansion from your own code, and the limit setting that silently drops a city
• Is this flight price good?: the band and the verdict documented, with three live captures
• Free tools: what the free server does and does not do
• Using the MCP in Claude · in Cursor · in ChatGPT