Create your 24/7 AI travel agent

This is how to build an AI travel agent without writing any code. By the end of this page you will have two things. An assistant, Claude or ChatGPT, that searches live Google Flights fares and Booking.com hotel rates mid-conversation, in plain language. And a scheduled task, an agent that tracks flight prices on your favourite routes every morning and only bothers you when a fare is genuinely cheap.

An AI travel agent is a chat assistant connected to live travel data, so its answers come from current fares rather than from training-data memory. Building one takes two pieces: a data connection the assistant can search, and a schedule so the checking happens without being asked. The agent in this guide connects Claude or ChatGPT to live flight prices through one URL, judges every fare against Google's own price history, and runs a daily scan that flags the cheap ones: a flight price alert with judgment attached.

The conversational part takes about ten minutes. The schedule takes five more. Every step below was run against the live products before it was written down; the UI paths were checked against the vendors' own documentation on 2026-08-27, and the example responses are real captures from the same day.

Why no code, and why no price-history database

A fare alert normally needs two hard things: a place to run every day, and an answer to "is this price actually good?". The second one usually means logging prices for months before your first useful alert.

Both problems are already solved here. Claude and ChatGPT can each run work on a schedule now, on their own servers, with your laptop closed. And every flight result from the API carries Google's own judgment of the fare, computed against the historical price band Google keeps for that route and date:

{
  "price_range_in_relation_to_other_periods": "low",
  "price_insights_low": 115,
  "price_insights_high": 220,
  "total_price": "$136",
  "total_price_as_number": 136,
  "buy_link": "https://www.google.com/travel/flights?tfs=..."
}

That is a real capture (London to Barcelona round trip, October 9 to 12, fetched 2026-08-27). The verdict field is one of low, typical, or high. So the entire alerting brain of your agent is one instruction: tell me only when the verdict says low. Google's standard, not a threshold you guessed.

What you need

The prerequisites, in full:

• A RapidAPI key. Subscribe to the Google Flights Live API (and, if you want hotels too, the Booking Live API; they are separate listings, one key covers both once subscribed). The free BASIC tier is 10 requests per month, hard-capped: enough to verify the setup works, not enough to run it. The honest cost math is in its own section below; the short version is that a real daily scan lands on the $10 PRO plan.
• A Claude or ChatGPT account. On Claude, the conversational agent works even on the free plan (one connector) and the scheduled part comes with any paid plan. On ChatGPT the gate sits at the connector, not the schedule: adding a custom connector needs the Pro plan or a Business workspace, details in Path B.

No server, no cron box, no code. If you would rather run the schedule in your own infrastructure instead of the assistant's, that version of this build is the n8n guide.

Try it with zero setup first (optional)

A free, ad-supported MCP server exists at google-flights-lulu.flightpowers.com/mcp that needs no key at all. Add it as a connector (same steps as below, nothing to paste into the URL) and you can feel the whole loop before creating any account. Two trades, stated plainly: results carry a sponsored line, and the server runs on a capped daily budget shared by everyone, so it can simply stop mid-afternoon. Point your actual agent, and anything scheduled, at the keyed servers.

Path A: Claude

A1. Add the connectors

Claude talks to the API through two hosted MCP servers, one for flights and one for hotels. Your RapidAPI key rides on the connector URL, so the key lives in your settings, never in a chat.

1. On claude.ai in a browser, open Customize > Connectors. (On Team or Enterprise plans an Owner adds it under Organization settings > Connectors.)
2. Click +, then Add custom connector.
3. Name it FlightPowers flights and paste the server URL with your key in place of the placeholder:

https://flights.flightpowers.com/mcp?rapidapi_key=YOUR_RAPIDAPI_KEY

4. Click Add. Ignore the OAuth fields under Advanced settings; this server authenticates with the key on the URL and nothing else.
5. Repeat for hotels if you subscribed to the Booking listing:

https://hotels.flightpowers.com/mcp?rapidapi_key=YOUR_RAPIDAPI_KEY

Three things worth knowing before you move on. Connectors cannot be edited, only removed and re-added, so if you ever rotate your key, that is the move. The free Claude plan allows exactly one custom connector; if that is you, add flights. And treat the URL itself as a secret: anyone who has it can spend your quota.

A2. Say hello

Open a new chat. Click the + button at the lower left of the message box, choose Connectors, and toggle your new connector on. Then ask something real:

Find me a nonstop LHR to JFK flight on October 13 and tell me if the price is any good.

Success looks like Claude announcing a tool call named search_oneway_flights, a few seconds of searching (this is a live scan of Google Flights, not a cache read), and then an answer that quotes a fare with Google's band and verdict. If Claude answers instantly from general knowledge instead, the connector is not toggled on in this chat; that is the single most common miss.

The four tools it now has: search_oneway_flights, search_roundtrip_flights, search_hotels, find_hotel_by_name. Both flight tools accept a date range and a list of destinations in one call, which matters for the schedule below. More first prompts on the Claude integration page.

A3. Schedule the morning scan (the flight price alert)

This is the 24/7 part: an agent that finds cheap flights while you are not looking. Claude's scheduled tasks live in Cowork mode, are available on every paid plan (Pro, Max, Team, Enterprise; Cowork itself is on the desktop app, with web and mobile in beta), and run on Anthropic's servers: the schedule fires with your computer asleep and the app closed. Scheduled tasks can use your connectors, which is exactly what we need.

1. Click Scheduled in the left sidebar.
2. Click New task (upper right), then Set up manually. (Create with Claude also works: describe the scan in plain language and confirm.)
3. Name it, set the frequency to daily at a time you will actually read it, and paste a task prompt like this one:

Run one search_roundtrip_flights call: from LHR to the destinations
BCN, LIS and ATH, departing the coming Friday, nights=3, sort by
price. For each destination report the cheapest fare, Google's band
(price_insights_low to price_insights_high) and the verdict
(price_range_in_relation_to_other_periods).

If any destination's verdict is "low", lead with it and include its
buy_link. If none is low, reply with one line: "Nothing cheap today",
plus the single cheapest fare for context.

If search_status is "partial", say which destinations actually
returned results and treat the missing ones as unknown, not as
empty.

4. Save. The task now appears on the Scheduled page, where you can edit its instructions or cadence, pause it, or run it on demand right now to test it.

Each run lands as its own session on that Scheduled page. Note the last paragraph of the prompt: it is not decoration. A live scan can complete for some destinations and not others, and the response says so honestly (search_status, plus a search_coverage object listing exactly what was searched). An agent that reads those fields will never present a half-finished scan as "no deals".

Path B: ChatGPT

The honest version first, because OpenAI's side of this moved a lot in 2026 and most write-ups online describe a product that no longer exists. Custom GPTs with Actions, the classic way to do this, can no longer be created on personal accounts (Free, Go, Plus, or Pro); GPT creation is now a Business and Enterprise workspace feature, and scheduled tasks explicitly do not support GPTs anyway. The live route today is a developer-mode MCP connector, and it has a plan gate. Everything in this section is from OpenAI's help center as of 2026-08-27; if you are reading this much later, expect drift.

B1. Add the connector (developer mode)

Custom MCP connectors in ChatGPT require developer mode, which exists on the Pro plan and on Business / Enterprise / Edu workspaces. On Free and Plus there is currently no way to add one; if that is your account, the Claude path above works on Claude's free tier for the conversational part. Pro's developer mode allows read-level tools, which is all this server has: all four tools are read-only searches.

1. On chatgpt.com in a browser (connectors are web-only, not mobile), open Settings > Apps > Advanced settings and enable Developer mode. On Business this is per-user under User settings; on Enterprise an admin has to allow it first under Connected Data permissions.
2. From the Apps/Connectors section, create a new connector. Name it, paste the flights server URL with your key on it (same URL as Path A), and set authentication to none: the key on the URL is the whole handshake.
3. Add the hotels URL as a second connector the same way if you want both.

B2. Say hello

Open a new chat, pick the connector from the tools menu (or just mention it by name in your prompt), and ask:

Is $480 for LHR to JFK in mid-October a good price, or should I wait?

Success looks the same as on Claude: a visible call to search_oneway_flights, a few seconds of live searching, then an answer that quotes Google's band and verdict rather than a vibe. Deep research can also use the connector, read-only.

B3. Schedule it, with one caveat you should test yourself

ChatGPT's scheduled tasks exist on every plan: ask in a chat ("check this every morning at 7:30 and message me") or manage them at chatgpt.com/schedules. Active task caps by plan, per OpenAI's help center on 2026-08-27: 3 on Free and Go, 5 on Plus, 15 on Pro and Enterprise, 10 on Business and Edu. Results arrive as push or email notifications, configured under Settings > Notifications. Free and Go tasks run in flexible time windows rather than at exact times.

The caveat: OpenAI documents that scheduled tasks can use its built-in connected apps (it names Gmail, Slack, and GitHub) and that tasks do not support GPTs. What the docs do not say, in either direction, is whether a scheduled task will call a custom developer-mode connector during its run. We will not claim it works; check it in your own account. Create the task with the same scan prompt as Path A, then read the first run's output: if you see the tool call and real fares with verdicts, you have your 24/7 agent on ChatGPT. If the run comes back with generic web-search results instead, your plan's tasks are not reaching the connector, and the schedule that works today is Path A on Claude or the self-hosted n8n version.

One more moving part, for completeness: ChatGPT's separate "agent mode" is marked discontinued in OpenAI's own help pages in favour of ChatGPT Work, which is documented to support scheduled and triggered runs. That surface is mid-rollout and aimed at workspaces, so this guide does not build on it.

The whole setup, copyable

Everything above in one block: the connector URLs and the scheduled-task prompt. Add the URLs as custom connectors (Claude: Customize > Connectors; ChatGPT: developer mode under Settings > Apps), then paste the prompt into a daily scheduled task and swap in your own airports.

Flights connector URL:
https://flights.flightpowers.com/mcp?rapidapi_key=YOUR_RAPIDAPI_KEY

Hotels connector URL (optional, separate subscription):
https://hotels.flightpowers.com/mcp?rapidapi_key=YOUR_RAPIDAPI_KEY

Daily scheduled-task prompt:
Run one search_roundtrip_flights call: from LHR to the destinations
BCN, LIS and ATH, departing the coming Friday, nights=3, sort by
price. For each destination report the cheapest fare, Google's band
(price_insights_low to price_insights_high) and the verdict
(price_range_in_relation_to_other_periods).

If any destination's verdict is "low", lead with it and include its
buy_link. If none is low, reply with one line: "Nothing cheap today",
plus the single cheapest fare for context.

If search_status is "partial", say which destinations actually
returned results and treat the missing ones as unknown, not as
empty.

What a run actually looks like

The exact scan from the task prompt above, run live through the server on 2026-08-27 (departing Friday 2026-10-09, three nights):

So that morning the agent's message reads roughly:

Two of your three routes are cheap by Google's own standard today. Barcelona is $176 nonstop for Oct 9 to 12, rated low against its usual $115 to $220 band, here is the booking link. Athens is $304 nonstop on Aegean, also rated low. Lisbon is $304 and merely typical, nothing to act on there.

Every number in that table is a real capture, not an illustration. The buy_link on each result opens that exact itinerary on Google Flights, ready to book. Prices move; if you run the same scan today the fares will differ, which is rather the point of scanning daily.

What this costs, honestly

The API bills per search: each date/destination combination inside a call is one billed request against your RapidAPI plan. One call scanning three destinations for one departure date is 3 requests, not 1. The paid MCP servers return an api_usage block in every response showing exactly what the call consumed and what remains, and the flight tools take a max_searches parameter to cap a wide scan before it spends more than you meant to.

The math for the example scan above:

• 3 destinations, 1 departure date, daily: ~90 requests/month.
• Add a second candidate weekend to each morning: ~180 requests/month.
• The free tier (10/month, hard cap) is exhausted on day four of even the small version. It exists to verify your key, and it is honest to say so.
• The $10 PRO plan on flights is 2,500 requests/month, which fits either version with room for the questions you ask by hand. That plan is the realistic floor for an agent you actually use. Current numbers for every tier are on /pricing.

On top of that sits the assistant's own plan: for scheduling, any paid Claude plan; on ChatGPT, the Pro plan or a Business workspace for the connector itself (the schedule feature exists on every plan). The conversational agent alone works on Claude's free plan (one connector).

Limits, and what the agent cannot do

• It cannot book flights. It finds, judges, and hands you the buy_link. You click it. That is the correct division of labour anyway.
• The verdict can be null on routes where Google publishes no historical band, usually thin or unusual city pairs. Treat null as "no signal", not as "not low". If your route never gets a verdict, alert on total_price_as_number under a budget you pick instead.
• Searches take seconds, not milliseconds. This is a live scan of Google Flights. Round trips on hard routes take the longest. When a search fails, the server retries and then escalates to a second, independent data source on its own; you do not need to manage that.
• A scan can be partial. The response never hides this: search_status says ok or partial, and search_coverage lists what was actually searched. The task prompt above makes the agent surface it.
• Rate limits are per minute (150/min on flights PRO), which only matters if you ask for a very wide scan; the servers expand ranges internally and pace themselves, and max_searches is your cap.
• Flights and hotels are separate subscriptions on RapidAPI. One key, two listings; a 403 on hotel tools usually means the key is only subscribed to flights.

Troubleshooting

Claude answers from general knowledge instead of searching. The connector is not enabled in this chat. Click + > Connectors and toggle it on. On the free plan also confirm you have not hit the one-custom-connector limit.

missing_api_key error in the tool result. The key did not make it onto the URL: a typo in rapidapi_key=, or the placeholder was left in. Remove the connector and re-add it with the corrected URL (connectors cannot be edited in place).

403 or "not subscribed". The key is valid but not subscribed to the listing that tool belongs to. Flights and hotels are separate subscriptions; subscribe on the listing's pricing tab and it clears immediately.

429 or quota exceeded. The free tier's 10-request hard cap, most likely. The api_usage block in each paid-server response tells you where you stand before this happens.

Empty results on a route that obviously has flights. Check search_status in the response. partial means the search did not complete and the results are a floor, not a verdict of "no flights". Run it again; the server retries and falls back to a second data source by itself, so a repeat usually completes. The full taxonomy is in Handling empty flight search results.

The scheduled task never fires. On Claude, scheduled tasks need a paid plan and live in the Scheduled page of the sidebar, where every past and upcoming run is listed; check there first, and use "run on demand" to test outside the schedule. On ChatGPT, check chatgpt.com/schedules and your active-task cap, and confirm notifications are on under Settings > Notifications.

ChatGPT has no way to add the connector. Developer mode is Pro and Business/Enterprise only, and web only. On Free or Plus the connector option is genuinely absent, not hidden; use the Claude path.

A ChatGPT scheduled run ignored the connector. Known open question, covered in Path B: custom connectors inside scheduled tasks are undocumented. If your first run does not show the tool call, move the schedule to Claude or n8n.

Quick answers

Can Claude or ChatGPT track flight prices? Yes, with a data connection: on their own, both only guess from training data. Connected to a live flight API, either can quote a current fare with Google's own low, typical, or high verdict. For tracking on a schedule: Claude's scheduled tasks (any paid plan) run connectors on Anthropic's servers; ChatGPT has scheduled tasks on every plan, but adding a custom connector needs Pro or a Business workspace, and whether a scheduled run calls it is undocumented, so test yours once before trusting it.

Do I need to know how to code to build an AI travel agent? No. The connection is one URL pasted into the assistant's settings, and the alert logic is one sentence of the task prompt: report only fares whose verdict is "low". There is no server, no cron job, and no code in this guide.

What does a 24/7 AI travel agent cost? Two subscriptions. The data side: RapidAPI's free tier is 10 requests/month, hard-capped, enough to verify the setup; a real daily scan fits the $10 PRO plan (2,500 requests/month on flights). The assistant side: any paid Claude plan for scheduling, or ChatGPT Pro or Business for the connector.

Does the scan run while my computer is off? On Claude, yes: scheduled tasks run remotely on Anthropic's servers, on their cadence, with the app closed. On ChatGPT, task results arrive as push or email notifications rather than needing an open app, with the connector caveat above.

Related

• Claude integration page: connector reference, all four tools, FAQ
• ChatGPT integration page: connector reference for ChatGPT
• Price Insights API: the verdict field, documented and proven
• Using a flight API in n8n: the same fare watch, self-hosted
• Handling empty flight search results: why partial and empty are different answers