01738 260265 Talk to us

APP CASE STUDY · OUR OWN PRODUCT

zxn Chat: website chat with AI and a team inbox

zxn Chat is website chat that we designed, built and run ourselves. A chat window on your site answers from your own information, and a person steps in when it matters. The team replies from a web inbox or a phone. It has run on zxn.uk since 10 November 2026.

Book a free 20-minute callSee the product and prices

A sample bakery website with the zxn Chat window open: a visitor asks about gluten free cakes, the AI answers and a person joins the chat.
The zxn Chat window on a phone: the AI has handed over and a person from a sample bakery has replied.
A visitor asks about a cake. The AI answers from the bakery's own information, then a person joins. Real screens from the product. The chats on them are samples, and the visitors are made up.
The zxn Chat team inbox in a web browser, sample data: a list of chats on the left, the open conversation in the middle and the visitor details on the right.
The zxn Chat iPhone app, sample data: the inbox, with the Mine, Unassigned, All and Closed tabs and a list of chats from visitors.
The same chat in the team inbox, on a computer and on an iPhone. Real screens from the product. The chats on them are samples, and the visitors are made up.

At a glance

Who it is for
Small organisations that want visitors to get an answer straight away.
What it is
A chat window, a team inbox on the web, iPhone and Android apps, and a way in for AI tools.
Built on
Cloudflare. Chat history is kept in the EU.
Where it is today
Live on zxn.uk and in the web inbox. The phone apps are in testing.
First version
Built overnight, on 8 November 2026.

Why did we build our own website chat?

We answer visitors on zxn.uk every day, and we build sites for clients who want the same. Three things were missing.

What does zxn Chat look like? Real screens

Each picture is a capture of the running product. Under each one is the same thing said in plain words.

  1. 01On the website

    The visitor asks a question

    The chat window sits in the corner of the page. It carries the site's own colours and name. The AI answers from the information the business gave it, and says it is the AI.

    In plain EnglishLike a shop assistant who knows the shop, and who fetches the owner when a question is too big.

    A sample bakery website. The chat is invented.

    A sample bakery website with the zxn Chat window open: a visitor asks about gluten free cakes, the AI answers, a person called Ellie joins and says Friday is fine.
  2. 02On a phone

    It works on a phone, in light and dark

    On a phone the chat opens full screen. It follows the visitor's own light or dark setting.

    In plain EnglishMost visitors are on a phone, so the phone version is the one we check first.

    Sample chat, light and dark.

    The zxn Chat window on a phone in light mode: the AI has handed over and a person from the sample bakery has replied.The zxn Chat window on a phone in dark mode, showing the same chat.
  3. 03In the browser

    The team answers from one inbox

    Every chat is in one list. Tabs split them into Mine, Unassigned, All and Closed. The AI's replies carry a label. A private note, shown in yellow, is seen only by the team. A visitor panel shows where they are and which page they are on.

    In plain EnglishOne shared mailbox for chats, that shows who said what, and when.

    Sample chats and made-up visitors.

    The zxn Chat team inbox in a web browser, sample data: a list of chats on the left, the open conversation in the middle with an AI reply, a yellow private note and a person's reply, and the visitor details on the right.
  4. 04iPhone app, in testing

    The same inbox on an iPhone

    A native iPhone app, written in SwiftUI. Swipe a chat to close it or take it for yourself. Light and dark both work.

    In plain EnglishThe inbox in your pocket, so a chat is never waiting for you to get back to your desk.

    In testing. It is not in the App Store yet.

    The zxn Chat iPhone app, sample data: the inbox with the Mine, Unassigned, All and Closed tabs and a list of chats.The zxn Chat iPhone app in dark mode, sample data: the same inbox.
  5. 05A chat the AI is handling

    Take over from the AI at any moment

    While the AI is replying, a bar says so and offers Take over. A Suggest button writes a draft reply for you to edit. The AI stops by itself after 12 replies in one chat.

    In plain EnglishThe AI does the first answers. You are always one tap away from taking the chat.

    iPhone app, and the Android app beside it. Both are in testing.

    The zxn Chat iPhone app, sample data: a chat where the AI assistant is replying, with a Take over button.The zxn Chat Android app, sample data: a chat where the AI assistant is replying, with a Take over button.
  6. 06Android app, in testing

    Android gets its own native app

    Written in Kotlin with Jetpack Compose, not copied from the iPhone. It uses Android's own controls, so it feels like the rest of the phone.

    In plain EnglishNot one app squeezed onto two phones. Two apps, each made for its phone.

    In testing. It is not in Google Play yet.

    The zxn Chat Android app, sample data: the inbox with chats from visitors.The zxn Chat Android app in dark mode, sample data: the same inbox.
  7. 07Visitor details

    Know who you are talking to

    The team sees the visitor's name, email, town, local time and the page they are on. Past chats with them open from the same screen.

    In plain EnglishYou answer a visitor on the pricing page differently from one on the contact page.

    Sample visitor. The email address is made up.

    The zxn Chat iPhone app, sample data: the visitor screen with a name, an example email address, Leeds, local time and the page the visitor is on.

How does one message travel?

Here is what happens when a visitor types a question. Five steps, in the time it takes to blink twice.

Diagram of one message: a visitor's chat window sends it to the Worker, which saves it in the EU database, tells the live room, asks Claude and sends push messages. The live room updates the web inbox, the iPhone app and the Android app. AI tools reach the Worker through MCP, A2A and the API.
  1. 01 of 05

    A visitor sends a message

    The chat window on the website sends it to our Worker, a small program on Cloudflare. The window loads nothing until the visitor opens it.

    In plain EnglishLike posting a note through a door that only opens when someone knocks.

  2. 02 of 05

    The message is saved first

    The Worker checks which website it came from and writes the message to the database. The database is in the EU. Saving comes before anything else, so nothing is lost if something later fails.

    In plain EnglishWrite it down before doing anything else.

  3. 03 of 05

    A live room tells everyone

    Each team has one live room, a Cloudflare Durable Object. It tells every open inbox at once, on the web, an iPhone or an Android phone. A room with nobody in it costs nothing.

    In plain EnglishOne room, one announcement, every screen updates.

  4. 04 of 05

    The AI answers, or asks for a person

    The AI, Claude from Anthropic, reads the information the business gave it. It either answers or asks for a person. A limit on replies per chat and per day keeps it in check.

    In plain EnglishAnswer if it knows. Fetch a person if it does not.

  5. 05 of 05

    The team replies and the phone buzzes

    A person takes over from any inbox. The visitor sees the reply at once. Phones are sent a push message through Apple or Google.

    In plain EnglishThe reply reaches the visitor while they are still on the page.

What is it built with?

One platform, Cloudflare, and a few well-known tools. Each part has one job.

How did we make the big choices?

Six decisions shaped it. Each one had a real alternative. This is the record of what we weighed.

01 Decision

Native phone apps, or one app for both?

We wanted apps that feel at home on each phone.

  • ChosenTwo native apps: SwiftUI and Jetpack Compose

    Each phone keeps its own feel: swipe actions, the back gesture, and a reply typed straight into a notification. The cost is two code bases to keep.

  • RejectedOne app in Flutter

    Less code to keep. But the native controls would need a bridge between the app and the phone.

02 Decision

Where should the chats live?

Live rooms are fast, but we also need a safe record we can copy and restore.

  • ChosenThe D1 database is the record. Live rooms only pass messages on

    Our copy and restore checks already work on D1, and questions about the data are simple to ask.

  • RejectedKeep the chats inside each team's live room

    Possible, but we would have had to build the copy and restore checks again.

03 Decision

How is a returning visitor recognised?

The chat window runs on the customer's website, not on ours.

  • ChosenA random code kept in the window's own storage on that site

    Only a scrambled copy of the code is saved on our side. It works in every browser.

  • RejectedA cookie from our own domain

    On a customer's site that is a third-party cookie, and Safari and Firefox block those.

04 Decision

How should AI tools connect?

Other software, and AI assistants such as Claude, should be able to work the inbox.

  • ChosenScoped API keys for the MCP server, and an A2A address for each site

    Keys work in every MCP tool today. A2A lets one AI agent talk to a business's chat the way a visitor would.

  • RejectedSign-in with OAuth for MCP from the start

    A good idea for later. For the first version it would have left some tools unable to connect.

05 Decision

How do we keep the web inbox safe?

It holds other people's conversations.

  • ChosenA protected session cookie, a check of where each request comes from and a strict content policy

    Nothing secret sits in the browser's storage. The build fails if any page contains inline script.

  • RejectedA token kept in the browser

    Simpler, but a script that slipped in could read it.

06 Decision

How do we stop the AI running away?

An AI that answers everything can cost a lot and say wrong things.

  • ChosenAnswers only from the business's information, 12 replies per chat, a daily limit per site

    It stays on topic, hands over to a person, and cannot run up a bill.

  • RejectedNo limits

    Fewer rules to explain, but one bad day could cost real money.

What has it done so far?

Only things we can show. No made-up numbers.

What is not done yet

zxn Chat is live, and it is young. These are the honest edges.

zxn Chat: common questions

What is webchat?

Webchat is a chat window on a website. A visitor types a question and gets an answer straight away, from an AI or from a person.

Is zxn Chat a chatbot or live chat?

Both. The AI answers first, from the information you give it. A person can take over at any moment, and the AI asks for one when it is not sure.

Does the AI make things up?

It is built not to. It answers only from your information and asks for a person when it is unsure. It stops after 12 replies in one chat.

Where is the chat history kept?

In the EU. The AI that writes replies is Claude, made by Anthropic, so a visitor's message goes to it. The privacy notice says what is kept and for how long.

Does it work on any website?

Yes. Paste one line of code into the page, or install the WordPress plugin.

Can other software or AI assistants use it?

Yes. There is a documented API, webhooks, an MCP server and an A2A address for each site.

Is there an iPhone or Android app?

Both exist and are in testing. They are not in the stores yet. The web inbox works today, on a computer or a phone.

Want webchat on your website?

Book a free 20-minute call. We will show it working, and tell you if it is not right for you.

Book a free 20-minute callSee the product and prices

Facts checked on 15 November 2026 against the product's own decision records and its production check. The chats on every screen are samples, and the visitors are made up.