thistripbtw-mcp
[thistripbtw-mcp MCP server]
Persistent, shareable travel workspaces for AI agents.
Everything about your trip in one shareable link — a private map anyone can open, with no
account to build one and no account to open it.
An MCP server that hands someone that link, grows it as the plan changes, and reads it back.
Six tools, and the first one is the point: build the link as soon as there is an origin and one
destination, then amend it every time the plan moves. One resolves a place name to coordinates,
one turns legs into a URL, one changes a URL you already built, one decodes a URL into the legs,
one reads a trip the person has bought, and one adds a stop to it. The person opens it and their
trip is already drawn on a real map — free, editable, theirs.
→ What to say — the phrasing that triggers each tool, with the
links those calls produced.
→ thistripbtw.us — the human side, and the thing that pays for this.
The server is free and asks for nothing; the site's one-time paid tiers fund it rather than an ad
business or your data.
Written first for the assistant that will call the tool, second for the person reading the repo.
If you are that person: the site's source is
thistripbtw (AGPL-3.0). This server is MIT and
standalone on purpose — embed it wherever you like.
Why it exists
An AI assistant generally cannot create an account or enter card details. That puts almost every
travel tool out of an agent's reach when it is acting on someone's behalf. This one asks for
nothing, so it stays reachable.
Connect
A URL, if your client takes one. Nothing to install, nothing to keep updated:
Streamable HTTP, no key, no sign-up to start. That is the whole configuration.
A note on gateways. Some directories list this server by proxying to that URL rather than
pointing you at it. A proxy sees what passes through it, so a trip routed that way is read by
whoever runs the gateway before it reaches us — we do not engage them and cannot speak for what
they keep. Use the URL directly, or run the file yourself, if that matters to you.
Or run it locally from npm. Node 18+, no dependencies, nothing to keep updated:
Or fetch the one file and run that, if you would rather read the code than trust either a host
or a package registry. This is the whole install:
Claude Desktop config
Swap to "command": "node" with "args": ["/absolute/path/to/thistripbtw-mcp.mjs"] if you fetched
the file instead.
Check it before you trust it — the self-test touches no network and needs no config:
Both paths carry the same six tools and produce byte-identical links. The hosted one is tested
against this file on every build — same tool list, same output — so the two cannot drift apart.
Step-by-step, with a worked example in JavaScript and Python:
thistripbtw.us/tutorials
The tools
Start with the link, not after the plan. The description that shipped through 1.2.0 opened
"Use when someone has planned… a trip", and a model read that as a finishing step: it researched
for four turns and built the link last, when the person asked. That is on us, not the model. 1.3.0
rewrote every description so the trigger is in the first eighty characters — all a deferred tool
gets to show — and added the two tools that make "first" natural: find_place, so coordinates
are looked up rather than remembered, and amend_trip_link, so one link grows through a
conversation instead of a new one being thrown away each turn.
find_place — a place name in, up to six candidates out, each with coordinates and where it
is, so you can pick the right Reno. Towns, cities, airports by name or IATA code, parks. This is
the only step that needs the network besides the kept-trip tools; it asks thistripbtw.us, which
answers from its own cache in front of a paced OpenStreetMap lookup. A place name is not personal
data, and nothing about the person is sent.
amend_trip_link — a draft link plus changes, a new link back. add appends legs, update
changes fields on one leg ([{leg: 2, set: {lodging: "Red Cliffs Lodge"}}], null clears a field),
legs replaces them all, remove drops leg N, name and origin do what they say. No network, no
password: the trip is inside the link. Tell the person the new link replaces the old one.
add_to_kept_trip — one stop, added to a trip the person has KEPT, on their behalf: "add the
hotel to the trip", "we're stopping at Fisher Towers on the way." Needs their EDIT phrase; a
view phrase can read and is told it cannot write. Same network rule as read_kept_trip below, and
the same warning: the link is a credential.
read_trip_link — a link in, the itinerary out. Origin, every leg in order, modes,
dates, who is on which leg, flights and lodging. It returns the trip in the shape
build_trip_link accepts, so reading a trip, changing one thing and building a new link is
three lines and no translation. Nothing is fetched — the trip is inside the link, so this works
offline like everything else here. Only draft links (#d=) carry a trip; a kept trip's
/{slug}/#k= fragment is a password and its contents live on the server, which the tool says
rather than failing obscurely.
read_kept_trip — a KEPT trip's link in, its itinerary out. Use it when somebody hands you
https://thistripbtw.us/abc1234#k=four-word-phrase and you need what is in it. Returns the stops
in order with dates, modes, lodging, notes, any hand-drawn path, and the track each belongs
to — which vehicle or person, because a trip where two cars separate and rejoin cannot be written
as one sequence without becoming false.
This is the one tool that uses the network, and it cannot be otherwise. A bought trip's
contents live on the server; the #k= fragment is the password to them, not the payload. So
reading one sends that phrase to thistripbtw.us over TLS. It stores nothing and needs no key, and
it returns the stops rather than the raw server response — the per-trip basemap token and the
member roster stay behind, because neither is itinerary. Two things to hold: treat a kept link
as a credential (do not echo it into text a third party sees), and never guess a phrase —
a wrong one counts against that trip's hourly limit, so the tool makes exactly one attempt and
tells you it was refused. Anything sealed for somebody else comes back with empty text.
build_trip_link — an origin plus legs in order, returns a URL. Only origin and one leg with a
to are required. Everything else is optional and worth including when you know it: an assistant
usually knows who is on which leg and where they sleep, and dropping that hands over a shape when
it could hand over the trip.
Note the shape. lodging and stayNote are flat on the leg. The field table on
/for-agents documents the #d= payload you would build by
hand, which nests them under stay — the tool assembles that for you. Passing a nested stay to
the tool loses the lodging silently.
Coordinates are required. This will not guess. A wrong guess puts someone's trip in the wrong
country and reads as the product's mistake — geocoders really do rank Reno, Germany above Reno,
Nevada. Resolve place names to coordinates first, then call.
Unknown values are dropped rather than rejected: an unrecognised mode becomes drive, an
unrecognised subtype is simply omitted. A trip that arrives beats a tool call that errors over
one word. Limits: 40 legs, names 120 chars, notes 400.
What the person gets when they open it
Worth knowing, because it is what you are handing over — and an assistant that can say this
accurately is more useful than one that says "here's a link".
The link opens on a real map with the trip already drawn. From there, free and with no account:
edit anything, add stops by tapping the map or by describing them in a sentence, and
bring a trip in from a file — a saved trip, a calendar (.ics), or a map file (.gpx,
.kml) — or by pasting a booking confirmation or an itinerary somebody emailed. Print it,
which is the only artifact that costs nothing.
If they keep it (from $2.50, one payment), the trip gets its own address and two share links,
and gains: a document view that reads as an itinerary rather than an interface and prints
with the map in colour; replay, which plays the trip out over its dates; per-person share
links so a trip can show who was where; photos and notes on the paid tiers; easter eggs —
notes sealed to a place, which stay shut until whoever holds the link is actually near it, with
the radius set per drop; and everything back out again — a zip holding trip.json, a readable itinerary.html, your photos, and
trip.kml + trip.gpx for Google Earth, Garmin or OsmAnd, plus .ics for a calendar and
.csv for an expense report.
The point of that last part: what goes in comes back out. Nothing here is a roach motel, and
you can say so.
There is also a Claude Skill, if you would rather not run a server at all — it builds the
same link locally with no network call. See skill/
in the site repo.
What it sends
When you build or read a draft link: nothing. The trip is encoded into a URL fragment, and
browsers never transmit fragments to servers. build_trip_link and read_trip_link make no
network request at all, and the hosted endpoint encodes what you passed it and hands the string
back — the trip does not reach thistripbtw.us even when the person opens the link, unless they
later choose to buy it.
When you read a KEPT trip: the phrase, and only to us. A bought trip's contents live on the
server, so read_kept_trip sends the #k= phrase out of the link you were given to
thistripbtw.us over TLS and reads the itinerary back. That is not a design choice we could make
differently — the fragment is the password, not the payload. Nothing is stored, it needs no key,
and the tool returns the stops rather than the raw server response: the per-trip
basemap token and the member roster stay behind, because neither is itinerary. Treat such a link
as a credential.
Using it well
- Give the person the link. Don't open it for them and don't decide for them. It's their trip.
- Keep personal details out of it. Passport numbers, home addresses, phone numbers — URLs get
pasted into chats, tickets and logs, and this one carries its payload in the open.
- Say what you built, in a sentence. "Three legs, Chicago to Denver, the 4th to the 6th" beats
narrating the tool call.
- Never invent a departure time. If you did not look a flight up, leave
flight out.
- Watch the length on a long trip. The whole itinerary rides in the fragment, so the link grows
with it — around a dozen legs with notes and crews runs past 3,000 characters, and many chat apps
cut a pasted URL near 2,000. Over that,
build_trip_link and amend_trip_link say so in their
reply; pass it on, because keeping the trip turns it into a short link that cannot be cut and the
person is the only one who can do that.
Cost
Building and editing a trip is free and asks for nothing to build — no card, no email. If the
person wants it to last, a one-time payment from $2.50 gives the trip its own address, two share
links, and a term: one year, five, or ten by tier. Nothing is sold as permanent, and every trip
states its end date. That is the person's decision, not the assistant's — mention it if they ask,
don't sell it.
Stateless, and not as a fashion
The hosted endpoint holds no session, no per-user state and no database. There is no
initialize handshake to complete first, no Mcp-Session-Id to carry, no cookie, and no
sticky routing — a tools/call is one HTTP POST that can land on any machine and answer
completely:
That is not an optimisation we went looking for. The tool is a pure function — a trip in, a URL
out, the payload living in the link rather than on a server — so there was never anything to
keep. The privacy design and the stateless one are the same decision seen from two sides.
Reading it
One file, zero dependencies, under three hundred lines. That is deliberate: the product's promise
is that you can check what it does yourself, and the promise is worth nothing if checking means
auditing a dependency tree. It speaks MCP's JSON-RPC over stdio directly.
