You describe where you are going. It comes back with something genuinely good: an itinerary with the museums in a sensible order and a note about which day the market is on. You are pleased. Then you have to put it somewhere.

So you copy it into a travel app. By hand, because there is no other way. Then a friend recommends a restaurant, you add it, and you move two things around to fit it in. Now the app knows your real plan and the AI does not. Next time you ask it something you paste the whole itinerary back into the chat, prefixed with "I changed things, day 3 now looks like this".

Chat. App. Retype. Chat again.

Why the usual fix does not fix it

Travel apps have noticed, and the answer has been to build an assistant inside the app. That solves the copying and introduces a different problem: you get their assistant. Their model, their prompt, their idea of what a good itinerary looks like, and, depending on the app, a second subscription on top of the one you already pay for.

Gullivr does the opposite. There is no AI inside it. Instead it exposes everything it can do as a tool an outside agent can call: create a day, add a place, reorder stops, find a hotel, save a flight. You bring the assistant you already use and already trust.

So what is it, concretely

A trip planner with an ordinary, complete interface. You can use the whole thing and never connect an agent at all.

That is one day of a three-day Lisbon trip, the same itinerary as our Lisbon guide, which is where the notes come from. Two things worth noticing. The stops are numbered on the map in the order you walk them, so a day that zigzags looks wrong immediately. And between consecutive stops there is a real distance and time, computed from where they actually are: 1.7 km, then 1.0 km, then 751 m. Lisbon is the city where that matters: two stops a few hundred metres apart can be a hundred metres of climb.

Around that sit the parts a trip actually needs: flights and transfers, accommodation, a packing list, and expenses that settle up between whoever came along. Trips are shareable two ways. A public read-only link for people who only want to see the plan, and a real invite with an editor or viewer role for the people building it. We wrote about how that split works separately.

And then the agent

Connect one, and it acts on your trips with your permissions. "Find three restaurants near the hotel and put them on day 3" creates real entries in the real app, the same ones you would have added by hand, visible immediately to anyone else with the trip open.

Plenty of things will write you an itinerary. What is different here is where it gets written: into the plan you are actually travelling with, so there is nothing left to copy and nothing to re-explain. The agent works inside the trip, not in a separate window whose output you have to move by hand.

Three ways to connect, and the trips are the same whichever you pick:

• ChatGPT — Gullivr is a native app inside it. Enable it once, nothing to install.
• Claude — a desktop extension.
• Anything else that speaks MCP — point it at the MCP server.

The agents page has the actual steps and the config to copy.

Where to start

Make a trip and add a few places by hand first. It takes two minutes and it is worth doing before you connect anything, because it shows you what the agent is going to be manipulating: days, stops, an order. After that, asking it to fill in a day is much less mysterious.

It is free to start, and there is no card involved.

Gullivr has three separate answers to that. They do different jobs, and the three are easy to mix up. Here they are, shortest first.

A share link, for people who only need to look

Turn on sharing for a trip and you get a public read-only link. Anyone with the URL sees the itinerary as it stands — days, places, times — without an account and without being able to change anything.

Use it for the people who are coming but not planning: parents who want to know where you'll be, a friend joining for two days, the group chat that just wants the address of the hotel. Turn sharing off and the link stops working.

An invite, for people who are actually planning

Inviting a collaborator by email adds them to the trip properly. They get one of two roles:

• Editor — can add and move places, edit days, add flights and accommodation, change anything you can change.
• Viewer — sees everything inside the app, including the parts the public link does not show, but cannot edit.

Roles are not permanent. Promote a viewer to editor when they start pulling their weight, or demote an editor who keeps reordering day 4. You can also cancel an invitation before it is accepted, and remove a collaborator afterwards.

The rule of thumb: editor for the two or three people building the plan, viewer for everyone else. A trip with six editors is a trip where nobody knows who moved the museum.

Edits show up as they happen

When two people have the same trip open, changes appear on both screens without a refresh. It sounds minor. It is the difference between a plan you trust and a plan you keep checking against the chat, because you watched the place land on day 2 as someone else added it.

Where the AI agent fits

If you have connected an AI agent, it acts with your permissions on your trips. So asking Claude to swap day 2 and day 3 produces edits your collaborators see immediately, the same as if you had dragged them yourself.

It has the same access a collaborator does, and its edits arrive the same way.

The short version

Travel guides

Itineraries and destination write-ups. Not the SEO-shaped kind that lists twelve cities and tells you nothing about any of them, but the kind you would actually copy into a planner and follow. Where the good coffee is. Which museum is worth the queue and which one is not.

How to use Gullivr

There is more in the app than you see on the first pass: shared trips with editor and viewer roles, flights, transfers, accommodation, a packing list, expenses that settle up at the end, public share links. Most of it is discoverable, but "discoverable" is not the same as "discovered". These posts take one feature at a time.

What's new

Short notes when something ships, and the occasional longer one about how it was built. This post is the first of them.

Where to start

There is already one of each of the first two: a guide, three days in Lisbon arranged so you are not always climbing, and a how-to on planning a trip with other people.

Everything is published in English and Italian. The switch at the top of any post changes language and keeps you on the post you were reading, rather than dropping you back on the index.

Lisbon is built on hills. An itinerary that ignores them is a day of pointless climbing. This one is arranged by neighbourhood and by altitude: each day stays roughly on one level, and you descend rather than ascend wherever there is a choice.

Day 1 — Baixa, Chiado, Bairro Alto

Start low and end high, once.

Begin in Praça do Comércio, at the river. It is flat, it is enormous, and it is the only part of central Lisbon laid out on a grid, the result of the 1755 earthquake that flattened everything and gave the Marquês de Pombal a blank sheet.

Walk north through Baixa to Rossio, then take the Elevador de Santa Justa up to Chiado if the queue is short and the stairs beside it if it is not. The lift is a nice piece of iron and a bad use of forty minutes.

Chiado is for the afternoon: bookshops, the Bertrand (the oldest still trading anywhere), coffee standing up at a counter. Walk uphill into Bairro Alto as it gets dark. The neighbourhood is dead until about nine and unrecognisable by eleven.

Evening: dinner in Bairro Alto, then down to Cais do Sodré if you want the night to continue.

Day 2 — Belém, and only Belém

Belém is six kilometres west and people try to fit it into a morning. It does not fit.

Take the tram or the train along the river. The three things worth your time are the Mosteiro dos Jerónimos (go at opening or after three; the middle of the day is a queue), the Torre de Belém from the outside (the interior is a stairwell with a view you can get elsewhere), and the Padrão dos Descobrimentos.

Eat a pastel de nata at Pastéis de Belém because you have to, then eat a better one somewhere else later in the trip. The original is good. The queue is not proportional to how much better it is.

Afternoon: the MAAT if you like modern architecture, the Coach Museum if you want the strangest collection in the city. Head back east along the river as the light goes.

Day 3 — Alfama, downhill

The mistake with Alfama is entering from the bottom. Do the opposite.

Start at Castelo de São Jorge, at the top. Get there for opening; the castle itself is mostly walls and peacocks, but the view is the point and it belongs to whoever arrives first.

Then walk down through Alfama, with no route. This is the one part of the city the earthquake spared, which is why the streets make no sense: they are medieval, and nobody ever redrew them. Getting lost is the intended experience. You will come out somewhere near the river eventually; the hill only goes one way.

On the way down: Miradouro de Santa Luzia for the tiles, the Sé cathedral if it is open, and any of the small places serving grilled sardines that do not have a menu in four languages.

Late afternoon: the Feira da Ladra flea market if it is Tuesday or Saturday.

What to skip

• Tram 28 as a sightseeing trip. It is a commuter tram full of tourists and pickpockets. Ride one stop if you want the photograph, then walk.
• Fado dinners with a set menu. The food subsidises the music and neither is good. Find a small place in Alfama that starts at ten.
• Sintra squeezed into three days. It deserves its own day and you do not have one.

Copying this into Gullivr

Three days, three neighbourhoods, one each. If you are planning this trip in Gullivr, the shape that works is one day per section above, with the sights added in walking order. The route distances will then tell you where you have been optimistic. Lisbon walking times run longer than the map says, because the map is flat and the city is not.

Or skip the typing. If you have an AI agent connected, ask it to build the trip and describe the shape: three days in Lisbon, one neighbourhood per day, the stops in the order above. It creates the days and adds the places itself, in your account, and what you get is the same trip you would have built by hand. Then you can argue with it. Move Belém to day 3, drop the Coach Museum, ask what else is near the castle. The itinerary is a real thing in a real app the whole time, not a block of text you have to keep re-pasting.

Any agent that speaks MCP works; the agents page has the setup. We wrote about why the app is built this way round separately.