The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Booking.com Hotel Search (Live Prices) listing page.
One URL, either way in:
In a client that supports MCP authorization, a Sign in button appears: you sign in with
Google, and your first 10 searches each day are free and ad-free on our key, with nothing to
paste. Paste your own RapidAPI key once on the /connect page when you want that cap gone, and
nothing goes in your client config either way. In a client that shows no Sign in button, bring
the key yourself:
Same URL, both times. A request carrying a credential of any kind is served; a request carrying
nothing at all is answered 401 with the OAuth metadata, which is what makes the Sign in button
appear. https://flights.flightpowers.com/mcp/oauth is still live and demands the sign-in on the
first request, for clients whose auth mode is fixed when a server is added.
Hosted. Nothing to clone, nothing to build. Listed in the official MCP Registry as
com.flightpowers/google-flights-mcp. Health check:
/health.
No key yet? You do not need one to start. Add the URL above, sign in with Google, and the first 10 searches each UTC day run on our key, ad-free, with nothing to paste. The exact unit counted is in Try it with no key.
Need a key? Subscribe to the Google Flights Live API on RapidAPI, free tier available,
and copy your x-rapidapi-key: https://rapidapi.com/mtnrabi/api/google-flights-live-api
Want more free searches than that, and do not mind ads? There is a separate free,
ad-supported server:
claude mcp add --transport http google-flights-free https://free-trial.flightpowers.com/mcp
(50 searches a day and 250 a month per signed-in account, one disclosed sponsored card per
result, fan-out capped at 15, and clients that cannot render the sponsored card may be capped
further.) Come back here when the ads, the 15-search cap, or those client restrictions get in
your way.
Two tools that answer a fare question, not a date lookup.
nights value instead of a fixed return date, and expand them internally.price_insights_low, price_insights_high, and a
price_range_in_relation_to_other_periods verdict of low / typical / high. That is what
lets an agent answer "$209 is typical here, don't rush" instead of just quoting a number.buy_link to Google Flights.api_usage: requests used by this call, and
what is left on the caller's plan. See Spend reporting.search_coverage, so the model can say
honestly which dates and destinations the answer is based on.Results are live fares. They go stale within minutes: never cache a fare or reuse an earlier result; search again and state when the data was fetched.
A signed-in account gets 10 free searches per UTC day on our key, ad-free, with nothing to
paste into your client. Sign in the way your client offers it, or at
https://flights.flightpowers.com/connect, and ask for a fare. Ten is what the hosted
deployment runs today; /health reports the live values as trial_enabled and trial_day_cap
and is the number to trust. (The allowance is off by default in this code and needs both
PAID_TRIAL_DAY_CAP and PAID_TRIAL_RAPIDAPI_KEY set on purpose, so a deployment you run
yourself has none until you configure one.)
A search is one date x destination combination, so one call with a three-day range spends three.
The allowance is held for the whole plan before the first request goes out and whatever the
fan-out did not send is given straight back, so two calls at once cannot both spend it. Every
result carries a trial block with the count and the cap. Past the cap the tools answer with
search_status: "trial_exhausted" and no results -- never an error, and retrying does not help;
the allowance renews at 00:00 UTC.
Comparing several sources in one call (compare_hotel_rates, or search_hotels with more than
one providers entry) is not part of the allowance and asks for your own key.
Connecting your own RapidAPI key removes the cap entirely. That is the next section.
The server holds no upstream credential of its own. Every search beyond the free allowance is billed to your RapidAPI subscription, which is why the key travels with the request.
x-rapidapi-key.If a key is missing and the allowance is not available to you, the tools do not fail silently and
do not spend anything. They return needs_api_key: true with the signup URL and these
instructions, phrased for the model to read back to you.
| Way | How | When to use it |
|---|---|---|
| Header (preferred) | --header "x-rapidapi-key: YOUR_RAPIDAPI_KEY" | Anything that lets you set headers. Keys stay out of URLs, and therefore out of proxy and access logs. |
| Query parameter | https://google-flights-mcp.flightpowers.com/mcp?rapidapi_key=YOUR_RAPIDAPI_KEY | Hosts that only let you paste a URL: claude.ai's custom-connector dialog is the case that matters. |
| Client API-key field | Paste the key into the client's own "API key" box | Hosts that send authorization: Bearer <key> or x-api-key. Smithery's saved-config form (config.rapidApiKey=) is also accepted. |
First non-empty source wins, in that order. The key is never logged, never echoed into an error message, and never returned in a tool response.
/connectThis is the page the sign-in URL at the top of this README sends you to. A client that speaks MCP authorization walks you through it on its own; the steps below are the same thing done by hand.
Where a deployment has it enabled (check connect_enabled on /health), there is a page at
/connect that replaces all of the above with a sign-in:
…/mcp?fp_token=fpk_…, and use that as
the server URL in your MCP client. Clients that let you set headers can send the same token
as Authorization: Bearer fpk_… instead. (It is hidden until you ask for it: that URL is a
90-day bearer credential for your plan, and a page that prints one by default puts it in
every screenshot and screen share.)If your client signed you in itself -- Claude, Cursor, ChatGPT and anything else that speaks
MCP authorization -- there is no URL to copy. /connect says so: it shows the key you
connected and tells you to go back to your assistant. There is a link on it for the case where a
second client cannot sign in and does need a connect URL.
What that buys you: your RapidAPI key is not in your client config, not in a URL, and not in
whatever logs that URL passes through. What it costs: the server stores your key, encrypted, and
knows your Google account id and email address. Disconnect on the same page deletes the record
and kills every connect token for your account, immediately. The full description is
section 2a of the privacy policy.
Some details worth knowing:
x-rapidapi-key header (or any of the
other channels above) and carry a connect token, the request's own key is used. Nothing you
already have set up changes behaviour because you signed in.needs_api_key reply telling you to reconnect. It never falls back to
somebody else's subscription and never spends anything./connect on your own deploymentOff unless all four of these are set. A half-configured deployment registers none of the
routes and serves keyed callers exactly as before; /health reports connect_enabled so that is
visible rather than guessed.
| Variable | What it is |
|---|---|
GOOGLE_OAUTH_CLIENT_ID | Google Cloud Console → Credentials → OAuth client ID, type Web application. Ends .apps.googleusercontent.com. |
GOOGLE_OAUTH_CLIENT_SECRET | The same client's secret (GOCSPX-…). |
MCP_KEY_MASTER | 32 bytes, base64: openssl rand -base64 32. Encrypts stored keys (AES-256-GCM) and derives the cookie and token signing keys. |
DATABASE_URL | Neon Postgres, pooled endpoint (…-pooler…). Schema: migrations/001_mcp_user_keys.sql. |
Optional: MCP_CONNECT_VALIDATE=0 stores a pasted key without checking it first.
Authorised redirect URIs to register on the Google client, one per product origin, exactly:
flights.flightpowers.com needs no entry. /connect and /connect/start bounce an alias to
the canonical origin before the sign-in starts, because cookies are per-host and Google compares
redirect_uri literally: an alias that started its own sign-in would come back to a host with no
state cookie and fail with a message that reads like a Google misconfiguration.
Also on the OAuth consent screen: scopes openid and .../auth/userinfo.email, and nothing else.
Rotating MCP_KEY_MASTER logs everybody out and invalidates every stored key. That is
deliberate: after a rotation nothing is left holding a token that resolves to a key nobody can
read. Users see "connect again", not a failed search. key_version on the table is there so a
staged rotation is possible later without a flag day.
Migration first, once per database:
Then, after deploying:
A tools/list that succeeds proves nothing about any of this: a token is only consulted when a
tool actually runs. Step 4 has to be a real search.
An MCP client only starts a sign-in when a request comes back 401 with a
WWW-Authenticate: Bearer resource_metadata=… header. Nothing else on the wire tells it one is
available. Until 2026-09-09 /mcp never sent that header, so a caller with no key got a 200
whose body said needs_api_key: correct JSON, invisible to every client's auth machinery, and
the user saw "the tool failed" with nowhere to click.
Now /mcp challenges, but only a caller who brought nothing, and only on a call that would spend something:
| URL | Behaviour |
|---|---|
https://flights.flightpowers.com/mcp | The one to use. A RapidAPI key in a header, on the query string or in a Smithery config blob, an fpk_ connect token or an fpo_ access token: all served exactly as before. Nothing at all: initialize, tools/list and the rest of the read-only handshake are still answered, and tools/call gets 401 + the challenge, so your client offers a Sign in button. |
https://hotels.flightpowers.com/mcp | The same, for hotels. |
…/mcp/oauth | The same server with the sign-in demanded on request one. For clients whose auth mode is fixed when a server is added, and for connectors saved on that URL before the change. |
Same tools, same product-per-hostname routing, same everything else. The alias is the identical tool registry behind a token check, not a second copy of the server.
The property that protects paying integrations: a request carrying a credential is never
challenged, and the credential order is unchanged (header, query, config blob, then OAuth
identity, then a connect token, then the RAPIDAPI_KEY env fallback). A wrong key is not
"nothing" either: it reaches the tools and comes back as the precise error RapidAPI gave,
because replacing that with a sign-in prompt would be a worse answer. A deployment with
RAPIDAPI_KEY set serves keyless callers off its own plan on purpose and is never challenged.
MCP_REQUIRE_AUTH=off turns the challenge off entirely, with no deploy.
Discovery stays open, and that was learned the hard way. For a few hours on 2026-09-09 the
challenge covered initialize and tools/list too. Glama re-checks every connector hourly by
opening an MCP connection and listing its tools, with no credentials; both paid listings were
marked unhealthy and ranked down the same evening, and Smithery's release scan, mcpservers.org
and M8ven probe the same way. So the line is drawn at spending, not at connecting: initialize,
notifications/initialized, ping, tools/list, prompts/list and resources/list are served
to anybody (src/discovery.py), and everything else needs a key or a sign-in. A batch with a
tools/call in it, an oversized body and an unparseable one are all challenged — the allowlist
fails closed. /mcp/oauth still challenges everything, which is the URL to give a directory
that wants a server always requiring auth, and the public, unauthenticated
/.well-known/mcp/server-card.json is still there for a scanner that reads a card instead.
What it is like to use. Paste the /mcp/oauth URL into your client. It registers itself,
opens a browser, you sign in with Google, and you approve that client by name on a page that
says exactly what it will be able to do: run searches billed to your own RapidAPI plan, nothing
else. The client never sees your RapidAPI key. If you have not connected one yet, the approval
still works and the first search comes back telling you to paste a key at /connect, with the
URL.
What you can revoke, and how. Press Disconnect on /connect: the stored key is deleted
and every OAuth token for that Google account is dropped in the same action. A client that was
connected stops working immediately. Individually, a client can call /oauth/revoke (RFC 7009).
| Route | Spec |
|---|---|
GET /.well-known/oauth-protected-resource and …/mcp/oauth | RFC 9728 |
GET /.well-known/oauth-authorization-server and …/mcp/oauth | RFC 8414 |
POST /oauth/register | RFC 7591, dynamic client registration, open |
GET /connect/authorize · POST /connect/authorize | RFC 6749 §4.1, PKCE S256 required |
POST /oauth/token | authorization_code and refresh_token |
POST /oauth/revoke | RFC 7009 |
The authorization endpoint is under /connect on purpose: the sign-in session cookie is scoped
Path=/connect so it can never be attached to a /mcp request, and putting authorize anywhere
else would mean either widening that cookie or making the user sign in twice.
Codes live 10 minutes and are single-use (DELETE … RETURNING, so two concurrent exchanges race
on one row and exactly one wins). Access tokens live 1 hour, refresh tokens 30 days with rotation.
Everything is opaque and stored as a SHA-256 hash, so a dump of the database contains nothing
that can be replayed. Tokens are not JWTs, deliberately: a signed token stays valid until it
expires whatever we decide afterwards, and Disconnect has to mean disconnect.
An access token is checked against the resource it was approved for before it is accepted, not
only when it is issued. Both products are the same deployment, the same database and the same
stored RapidAPI key per user, so without that check a token approved on the flights consent
page, which says "search live flight fares" and nothing else, would be accepted on the hotels
hostname and spend the user's hotels plan. The check is strict about the host and forgiving about
the path, because clients in the wild send the origin, /mcp and /mcp/oauth for the same
server. tests/test_oauth.py::TestATokenIsBoundToTheResourceItWasApprovedFor pins both halves.
No new environment variable. It comes on wherever /connect is configured, because it reuses
that Google sign-in and that key store. It needs one more table in the same database:
/health then reports oauth_enabled: true and oauth_mcp_endpoint; that URL is what goes in
a directory listing. MCP_OAUTH=off disables it while leaving /connect running; that is the
rollback that needs no code change.
Nothing has to change on the Google OAuth client. The redirect URI is still
…/connect/callback, because the MCP client's OAuth flow ends at our authorize page, and only
that page talks to Google.
/oauth/register is open, because the MCP spec requires it and because a client_id on its own
authorises nothing: every flow through one still ends at a consent page a signed-in human has to
press a button on. Open is not the same as unlimited, so three things stand behind it.
| Mechanism | Where it lives | What it stops |
|---|---|---|
| Rate limit | in memory, per instance | a burst: 10 registrations per address per 10 minutes, 60 token requests per minute, 10 /connect/save per hour. Over the limit is 429 with Retry-After. |
| Daily caps | Postgres, so every instance agrees | a slow drip: 30 registrations per address per day. The global cap is 5,000 a day -- a backstop against unbounded rows, not a defence: a global number set near real traffic is a lever an attacker pulls to refuse every new Claude or Cursor user for a day. Crossing 500 in a day logs and refuses nothing. MCP_OAUTH_DCR_MAX_PER_IP_PER_DAY, MCP_OAUTH_DCR_MAX_PER_DAY and MCP_OAUTH_DCR_WARN_PER_DAY move them without a deploy. |
| Sweep | on /oauth/register, at most once every 15 minutes per instance | the litter: expired codes and tokens, and registrations that never became an authorization within 7 days. A client with a live token, an outstanding code, or a consent page that has been rendered for it is never swept. |
The rate limit is per INSTANCE. On Vercel that means N warm instances allow up to N times those numbers between them, and a cold start starts the counters at zero. That is why the durable caps exist as well: they are counted in the database, where the number is the same everywhere.
The address every one of these is keyed on comes from x-real-ip first -- Vercel sets it to the
peer it accepted -- and otherwise from the LAST usable entry of x-forwarded-for, skipping hops
that can only be internal. Never entry 0: proxies append on the right, so the left-most entry is
whatever the caller wrote, and reading it would make every limit here one header away from being
bypassed. A request with neither header shares one bucket named unknown, which is rate limited
as a single caller and is not subject to the durable per-address cap (one missing header on the
edge must not lock the whole server out for a day).
A registration is stamped as in-use when its consent page is RENDERED, not only when the human
presses Approve. MCP clients commonly register when they are installed and authorize days later,
and the sweep must not delete a row while its consent page is on screen -- there is no foreign key
from codes or tokens back to the client, so the exchange that followed would fail
invalid_client with nothing naming the cause.
Requires one migration:
A refresh token is single-use: exchanging it issues a new pair and retires the one presented.
The retired row is kept and stamped, not deleted, because a deleted row and a token that was
never issued look identical -- and telling those apart is the point. Presenting an
already-rotated refresh token means either a client that lost the response or a copy in somebody
else's hands, and OAuth 2.1 §4.14.2 says to assume the second: the answer is invalid_grant, and
every token descended from that authorization is deleted. The honest client signs in again; the
thief's access token stops working at the same moment.
With one deliberate exception, for the case that is almost always the innocent one: the FIRST replay of the token we just rotated, from the same client, within 10 seconds, is answered with the pair that rotation already issued. It is an idempotent retry -- nothing new is created -- and it means a client whose response was lost to a dropped connection is not silently signed out. A second replay, or one after the window, is the real thing and still kills the family. The window is per instance and in memory, so a miss simply falls through to the conservative answer.
Revoking a refresh token through /oauth/revoke takes its access tokens with it, for the same
reason (RFC 7009 §2.1) -- and only if the token was issued to the client asking, which is the
other half of that section. A client presenting somebody else's token still gets 200 (§2.2) and
nothing is revoked.
client_id_metadata_document_supported: true is advertised in
/.well-known/oauth-authorization-server. A client may use an https URL as its client_id;
the document at that URL lists its redirect_uris, and we fetch and check it per flow instead of
writing a registration row. Smithery asks for this before it will proxy a remote OAuth server.
What is checked, every time: https only, a public hostname (no IP literals, no localhost, no
credentials in the URL), the hostname resolved and every address it answers with required to
be public unicast (a name is not a control: 127.0.0.1.nip.io has a dot in it and points at
loopback), no redirect followed, a 5-second timeout, a 64 KB cap enforced while the body is read
rather than after it is buffered, a client_id inside the document that matches the URL if it is
present, and -- the one that matters -- the redirect_uri in the request must be listed in the
document. The lookup is rate limited on its own (60 per address per 10 minutes), because it is
the only outbound fetch in this server that a caller can trigger before signing in. Dynamic registration is unchanged
and still the default: a client_id that is not an https URL is looked up in the table exactly as
before.
| Tool | What it does |
|---|---|
search_oneway_flights | Real-time one-way fares. Input: origin IATA, destination IATA or a list, and either one departure date or a date range. Returns price, airline, duration, stops, buy_link, and Google's historical price range so you can judge the fare. Use for any one-way question, including open-ended ones: one call with a range, never one call per date. |
search_roundtrip_flights | Real-time round-trip fares priced as paired legs, not two one-ways. Input: origin, destination(s), a departure date or range, and either a return_date or a trip length in nights (a number or a list like [5,6,7]). Returns total price, per-leg airline/stops/duration, and one buy_link for the trip. |
The hotels deployment serves search_hotels, find_hotel_by_name and compare_hotel_rates instead; see Hotels: providers and compare_hotel_rates.
search_oneway_flightssearch_roundtrip_flightssort_by is applied by this server across the merged result set from every search it ran, so it
is predictable regardless of how many combinations were expanded.
User: "I'm in Tel Aviv. Cheapest week-long trip to Rome or Athens, leaving any day in the first half of May."
One call:
That expands to 15 dates × 2 destinations = 30 combinations, which is exactly the per-call cap. The response shape (field names are real; the values below are illustrative, not a quote, run the call to get live fares):
Other response shapes to expect, all of them normal:
results: [] with a message: Google Flights genuinely
returns nothing for some route/date combinations. Not an error. Try nearby dates or a
nearby airport. use_fallback will not change this and is left unset by default: the
backend accepts the field, but the second flight-data source it selects is gated behind
USE_FALLBACK_FLI (fallback_available()), which is not switched on for this API, so
none of its three values has any observable effect on a search today. The automatic
retries the backend does on an unreadable page are unconditional and are not affected
by it.partial field says how many of the executed searches failed, and
the results cover the rest.search_coverage.truncated: true plus a note. The range is sampled
evenly across the whole window (first and last kept), not cut short, so the sample is
representative, not the first N days. Raise max_searches or narrow the range for fuller
coverage.needs_api_key: true, zero spend, with the fix. A valid RapidAPI
key that is not subscribed to this API is the most common cause.quota_exhausted: true with api_usage, plus a reminder that narrowing
the range makes remaining quota go further.POST /search prices exactly one stay, so "cheapest three nights in Rome in May" used to be
31 tool calls -- or, in practice, one call on a date the model picked and an answer presented as
the cheapest. Both hotel search tools now take the flights shape instead:
Same machinery as the flights fan-out (src/fanout.py): one backend call per stay, capped at
max_searches_per_tool_call (30, hard max 60), sampled evenly across the range when it does
not fit, and reported in search_coverage. nights derives each check-out date, so it replaces
checkout_date rather than joining it. A fixed checkout_date against a range of check-in dates
is allowed and means "out on the 4th, whenever I arrive"; the impossible pairs are dropped.
The response is bounded on purpose. Every property of every stay is ~25 KB per stay (measured: 25,892 bytes for 25 properties, 18,989 of them URLs), so each stay reports its cheapest property, its per-night rate and its median, and the full property list comes back for the cheapest stay only:
reason on a stay is a fact about our pipeline, never a guess about the hotel:
reason | search_status | Means |
|---|---|---|
ok | ok | priced |
no_availability | empty | searched, answered, nothing came back |
no_price | empty | properties came back, none carried a price (available: false lands here) |
search_failed | degraded | the search errored, so nothing is known -- not "no rooms" |
not_searched | not_searched | the cap sampled it away |
Counts are null rather than 0 on the last two: zero reads as "nothing there", and neither case
knows that. Top-level search_status is ok / partial / empty / degraded over the stays;
every stay failing raises instead of answering with an empty list.
Two deliberate refusals: a check-in range with providers naming more than one source (a fan-out
times a per-source fan-out, billed to two subscriptions, that neither search_coverage nor
api_usage can describe honestly today), and max_searches on a single stay, which would silently
do nothing.
A single stay is byte-identical to what it was before this existed -- same request body, same
response keys, no stays, no search_coverage, no search_status. Asserted in
tests/test_hotel_date_range.py::TestTheOldShapeIsUntouched against a frozen expectation captured
from the previous code.
providers and compare_hotel_ratesThe hotels deployment (hotels.flightpowers.com, the same code selected by the Host header)
serves three tools. search_hotels and find_hotel_by_name also take a check-in range (above);
search_hotels gained one optional argument for sources and there is one new tool.
| Tool | What it does |
|---|---|
search_hotels | Live rates for a destination and dates, or for every stay a check-in range expands to. providers names the sources to price on: ["booking"] (the default), ["airbnb"], or both. |
find_hotel_by_name | One named property, Booking.com only. Airbnb's room page carries no price, so a name lookup there would resolve to something that cannot be priced. |
compare_hotel_rates | The same stay priced on every source you have a key for, one row per source: cheapest total, median total, how many places were priced, currency, and when the rows were read. |
search_hotels with no providers argument sends the same upstream request it always sent, to the
same host, and answers with the same keys. So does providers: ["booking"]. Both are asserted in
tests/test_providers.py::TestTheDefaultDidNotMove, upstream request body included -- a default is
only a default if keeping it costs nothing.
Naming a second source changes the response shape, and only then:
booking goes straight to booking-live-api.p.rapidapi.com on the caller's key, billed by
RapidAPI to their own subscription. Unchanged.airbnb goes through our own front door, POST https://api.flightpowers.com/v1/hotels/search
with provider: "airbnb" in the body (flight_rabbi #472), because the Airbnb backend is not on
the RapidAPI edge. The call carries the caller's key as x-rapidapi-key and identifies itself
with X-FP-Client: mcp-hotels/<version> -- the front overwrites body attribution with its own
conclusion, so the header is the only thing that says who called (rule 11). Override the origin
with API_FRONT_BASE_URL for a preview front; it holds no credential.The Airbnb listing does not exist on RapidAPI yet, and the front ships with the provider off, so
today a real providers: ["airbnb"] call comes back as a degraded row saying so. That is the
designed answer, not a bug: an unmetered source reachable by anyone with any valid key is a gateway
we pay for.
providers_skipped with
reason (no_key, not_subscribed, key_rejected) and the URL where you subscribe. Each
listing is a separate subscription, so a 403 from the Hub means "not subscribed to that
listing", not "bad key".search_status: "degraded", count: null, no
prices -- and the other source's rows still come back. "Booking did not answer" and "Booking had
nothing" are opposite answers and must be sayable as different sentences.cheapest_total and median_total cover only the rows that carry a price, and count is
that number. A price_string is never parsed into a number.caveats says the totals are not comparable. Nothing
is converted.Almost everybody has one RapidAPI key subscribed to several listings, and that key is used for every source with no extra configuration. A caller who genuinely holds two can name one per source, and the source-scoped name wins:
The order is the one src/credentials.py already documents -- source-scoped names, then the
unscoped specific names, then the config blob, then generic names last and skipped entirely when a
config parameter is present, because a gateway's own key under a generic name is not ours.
filters and price_as_seen_from are Booking-onlyOn a mixed search they are sent to Booking and not to the front. On an Airbnb-only search they are refused, not dropped: a silently discarded filter returns more properties than you asked for and nothing says so.
outputSchema, structuredContent, isError)Every tool declares an outputSchema, and every result carries the payload
twice: once as structuredContent, once as the serialized JSON in a text
content block. The MCP spec asks for the duplicate --
For backwards compatibility, a tool that returns structured content SHOULD also return the serialized JSON in a TextContent block.
-- and it is load-bearing here rather than ceremonial, because clients that predate structured output read the text block and nothing else. (Verified against spec revision 2026-07-28; structured output arrived in 2025-06-18.)
The schemas are deliberately additionalProperties: true with only results
required. The spec puts the obligation on the server -- "Servers MUST provide
structured results that conform to this schema" -- and these tools have
several legitimate exits that carry different keys -- a zero-result answer, the keyless
needs_api_key reply and the quota_exhausted reply. A tighter
schema would look better and would make the server non-conformant on a path
it ships on purpose.
search_status, and why degraded is an errorFlight results carry search_status, mirroring the backend's own
X-Search-Status vocabulary:
| value | meaning |
|---|---|
ok | every combination searched returned results |
empty | the search completed; Google genuinely has no itineraries. A real answer |
partial | some combinations returned results, some failed. The list is incomplete |
degraded | every combination failed. The search did not happen; an empty list means nothing |
A degraded result is also flagged isError: true. It is the only one
that is. The spec classifies "API failures" as tool execution errors and says
clients "SHOULD provide tool execution errors to language models to enable
self-correction", while nothing in the spec obliges a host to show
structuredContent to the model at all. A failure carried only by a field
inside the payload is therefore a failure the model may never see, which was
the whole problem search_status was added to solve.
The payload still rides along with the error -- structuredContent and the
text block are both present, so nothing is lost. api_usage in particular: a degraded
search still spent the caller's own RapidAPI requests, and hiding that would
hide a charge they have to pay. empty and
partial are not errors: one is a true negative and the other carries
results a caller can use.
api_usage)The money is yours, so the meter is visible. Every successful response carries:
| Field | Meaning |
|---|---|
requests_used_by_this_call | Billed upstream requests this one tool call consumed. |
plan_requests_remaining | What is left on your RapidAPI plan this period. |
plan_requests_limit | Your plan's limit for the period. |
note | The same thing in a sentence, so the model can relay it to you before you ask. |
plan_requests_remaining and plan_requests_limit come from the upstream response and are
omitted when upstream does not report them; the note adapts. The rule the model should state
out loud: one date × one destination = one billed request.
Cost control knobs, in order of bluntness: max_searches per call (lower it to spend less on a
wide question), a narrower date range, a shorter destination list.
The underlying REST API takes exactly one (origin, destination, date) tuple per call. Against a
one-date-per-call passthrough, "cheapest to Sri Lanka anywhere in October" is 31 separate tool
calls: 31 round trips through the model, 31 chances to lose the thread, and a bill the user only
discovers afterwards.
Here it is one tool call. The fan-out happens server-side, concurrently, capped, evenly
sampled, deduplicated on buy_link, merged, sorted by your sort_by, and reported honestly in
search_coverage and api_usage.
| This server (paid) | Free server | |
|---|---|---|
| Fan-out per call | 30 (hard max 60; raise or lower per call with max_searches) | 15 |
| Ads | none | one disclosed sponsored card per result |
| Key | your own RapidAPI key | none needed |
| Spend reporting | api_usage in every response | n/a |
| Directory-listable | yes | no |
This server carries no ads at all, not by taste but by constraint: Anthropic's connector directory policy and OpenAI's app guidelines both prohibit advertising and sponsored content in tool results, so an ad-carrying server can never be listed there and this one can.
Point a client at the local process the same way:
Tests (620 passing, verified):
Configuration lives in example.env; every variable is documented there. The ones that matter:
| Variable | Default | Why it matters |
|---|---|---|
MAX_SEARCHES_PER_TOOL_CALL | 30 | Per-call fan-out cap. Clamped to a hard maximum of 60. |
MAX_CONCURRENT_SEARCHES | 10 | Concurrency of the fan-out. |
MAX_HTTP_CONNECTIONS | 60 | Connection-pool ceiling; serverless instances share a file-descriptor pool. |
REQUEST_TIMEOUT_SECONDS | 75 | The upstream function's Timeout (60) plus a 15s edge-relay margin, so this side never gives up on an answer that is still coming. |
DEFAULT_RESULT_LIMIT | 10 | Results requested per individual upstream search. |
MCP_PRODUCTS | both | Which product this deployment serves: flights, hotels or both. Selects the tool set, the server instructions, the service name, the policy pages and the RapidAPI listing a keyless or unsubscribed caller is sent to. A hotels deployment left on the default introduces itself as a flights server. |
SIGNUP_URL | listing matching MCP_PRODUCTS | Quoted back to users who arrive without a key. On both, the hotel tools quote the Booking listing regardless: one URL cannot be the Subscribe button for two APIs. |
MCP_PRODUCTS_BY_HOST | default | Which product each hostname serves, so one deployment can carry both paid domains and each listing still gets exactly its own tool set. default is the built-in map of flightpowers.com aliases; off disables host routing entirely (the no-code rollback); or an explicit host=product,… map. An unmapped hostname falls back to MCP_PRODUCTS. |
MCP_PUBLIC_URL | http://localhost:8000/mcp | Reported by /health and the origin of every policy-page link. MCP_PUBLIC_URL_FLIGHTS / MCP_PUBLIC_URL_HOTELS override it per product on a deployment serving both. Without them the hotels hostname would advertise the flights one. SIGNUP_URL_FLIGHTS / SIGNUP_URL_HOTELS work the same way. |
RAPIDAPI_KEY | (empty) | Leave empty in production. If set, every keyless caller is served on, and billed to, that subscription. The server logs a warning at startup and /health reports server_side_key_configured. |
METRICS_TOKEN | (empty) | When set, /metrics requires an x-metrics-token header. |
LOG_PATH | (empty) | Empty disables the file sink; stdout MCP_CALL lines remain the record. Correct on serverless. |
Operational routes: GET /health (public, unauthenticated: registries poll it),
GET /metrics, GET /metrics/calls?hours=24,
GET /.well-known/mcp/server-card.json (see below).
GET /.well-known/mcp/server-card.json returns this deployment's metadata as a standalone
JSON document: serverInfo, description, transport, capabilities, authentication,
instructions, and the full tools and prompts lists exactly as tools/list serialises
them. Public, unauthenticated, Cache-Control: public, max-age=3600, open CORS.
It exists because the URL we publish in directories is /mcp/oauth, which always answers
401, so an automated scanner pointed at THAT url cannot read the tool list off the wire.
(Since 2026-09-09 a scanner pointed at /mcp can: read-only discovery is served without a
credential — see src/discovery.py.) Smithery's publish
page names this document as the way out: "If automatic scanning can't complete (auth wall,
required configuration, or other issues), you can provide server metadata manually via a
static server card at /.well-known/mcp/server-card.json". The field list follows
SEP-1649.
Two things worth knowing:
tools/list returns. tests/test_server_card.py compares
the two on both products.hotels.flightpowers.com returns the hotel
card, both flights hostnames return the flights card, and every URL inside is on that
product's own origin. transport.endpoint is /mcp/oauth where OAuth is configured, with
the keyed /mcp endpoint listed under _meta as an alternative.Deployment target is Vercel via api/index.py (FastAPI wrapper handing FastMCP its lifespan,
stateless_http=True). The canonical MCP path is /mcp, no trailing slash.
Never commit a real key. example.env ships with placeholders; keep it that way.
The hosted server needs nothing installed. This is for self-hosting, and it is what lets Glama run its build test and cut a release.
The container serves streamable HTTP on ${PORT}/mcp, the same transport as the hosted
deployment, via python -m src. api/index.py is the Vercel wrapper and is not used here.
No secret is baked into the image. Every search is billed to the caller's own RapidAPI
subscription and their key travels with the request as x-rapidapi-key. RAPIDAPI_KEY is
optional and is a server-side fallback: when set, a caller who sends no key of their own is
served on, and billed to, that subscription. Leave it unset unless that is what you want;
/health reports server_side_key_configured either way.
Every variable in the table above works as -e NAME=value. HOST defaults to 0.0.0.0 and
PORT to 8000 inside the image. The HEALTHCHECK polls /health on $PORT, so overriding
PORT keeps working.
This is an independent API that returns publicly available flight pricing. It is not affiliated with, endorsed by, or sponsored by Google. "Google Flights" is used only to describe the public data source. Fares are supplied by the upstream provider, change constantly, and are not guaranteed. Always confirm the price on the airline or booking site before purchase.