> ## Documentation Index
> Fetch the complete documentation index at: https://docs.wednesdayai.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Using the web client

> Chat with your WednesdayAI assistant, manage conversations, and choose what the web client displays.

# Using the web client

The web client is a page you open in your browser to chat with your WednesdayAI assistant. You
can send messages, find and organize past conversations, choose how replies are displayed, and
check on things your assistant is currently working on.

The person who manages WednesdayAI decides whether this page exists and sends you its address. If
you do not have a link, ask them for one.

## Before you begin

* The web address for the web client, from the person who manages WednesdayAI.
* A **setup code**: a block of text that lets this browser start talking to your assistant. The
  person who manages WednesdayAI creates one for you. Setup codes expire, so use it soon after you
  get it.

## Pair this browser

The first time you open the web client on a browser or device, it asks you to pair - a one-time
step that links this browser to your assistant so it can be recognized on later visits.

**1. Open the web client**

Open the web address you were given.

You should see **Pair this browser with \[your assistant's name]**.

If the person who manages WednesdayAI enabled sign-in with Google, Microsoft or GitHub, you can
select one of the **Continue with ...** buttons instead of using a setup code: signing in pairs the
browser for you.

**2. Enter your setup code**

Paste the setup code into the **Setup code** box. The box should contain your code.

**3. Pair this browser**

Select **Pair this browser**.

The button reads **Pairing...** briefly, then you land in the chat screen. If you see an error
instead, the setup code may have expired or already been used - ask for a new one.

You only need to pair once per browser. Pairing this browser does not affect any other browser or
device already paired with your assistant.

## Install the web client on your home screen

The web client can be installed like an app: it gets its own icon on your device's home screen
and opens in its own window, without the browser's address bar.

**1. Open the web client in your browser**

Use the same address you normally use, and pair if this browser has not been paired yet.

**2. Install it**

In Safari on iPhone or iPad, open the **Share** menu and choose **Add to Home Screen**. In Chrome
on Android, open the browser menu and choose **Install app** (or **Add to Home screen**).

You should see a WednesdayAI icon on your home screen.

**3. Open the installed app and pair once more**

Open WednesdayAI from the home screen. An installed app on iPhone or iPad keeps its stored data
separate from Safari, so it asks you to pair again the first time - paste a setup code as before.

The installed app then remembers the pairing, the way a browser does.

Two things the installed app does not do: there is no offline mode - it still needs a connection
to your assistant - and it does not show notifications while it is closed. On iPhone or iPad, the
installed app also needs iOS 18.4 or newer to keep the screen awake while you dictate or listen to
a reply being read aloud; see [Dictate and voice](#dictate-and-voice).

## Send a message

Type in the message box and press Enter (or select the send arrow). Press Shift+Enter to add a
new line without sending. On a phone or tablet, Enter always adds a new line instead of sending:
tap the send arrow to send. With a keyboard attached, Ctrl+Enter (Cmd+Enter on a Mac) sends too.
Your message appears on the right, and the reply streams in beneath it
as your assistant writes it.

While your assistant is working on a reply - in this window, or in another browser or tab - the
thread shows a three-dot indicator and the message box stays open so you can keep typing. The send
arrow becomes **Add to current turn** - or **Queue message** while a file is attached - and a
separate **Stop** button appears beside it. Select **Stop**
to end the reply early, including one that another device started. If the stop request fails, the
thread shows **Could not stop the reply:** followed by the reason, and the reply keeps running.
See [Keep chatting while a reply is streaming](#keep-chatting-while-a-reply-is-streaming) for what
Enter does in the meantime, and
[Use several browsers or devices at once](#use-several-browsers-or-devices-at-once) for the full
multi-device behaviour.

Your message is formatted the same way a reply is: markdown such as **bold**, headings, lists,
`code`, tables and links render in your message bubble. Each line break you typed stays a line
break.

If you have not opened a chat yet (for example, the first time you use a new browser), you see a
greeting (**Good morning**, **Good afternoon** or **Good evening**), a line that reads **What can
\[your assistant's name] help with?**, and three suggestion buttons: **Plan my day**, **Help me
think** and **Summarise what I missed**. Selecting a suggestion copies its words into the message
box so you can edit them - it does not send anything until you press Enter. Your first message
opens your conversation with the assistant automatically; it uses the same conversation your
direct messages on other channels use when the gateway links them.

Each reply, the message box (**Message \[assistant's name]...**) and the empty-chat line name the
assistant the open chat belongs to. A reply's avatar is that assistant's emoji, or the first two
characters of its name, in its colour. Until the assistant list loads, they show the deployment's
name and logo; the sidebar header always does.

Hover over a reply to reveal its buttons. **Copy** copies the reply with its structure intact -
headings, lists, tables and bold - ready to paste into a document or an email, where it takes on
that document's own styling; if the browser cannot write rich text, the reply's markdown source is
copied instead. **Reply** quotes it (see
[Reply to a message](#reply-to-a-message)). Hover over a message of your own to reveal **Reply**
too. **Read aloud** reads the reply out loud; see [Read a reply aloud](#read-a-reply-aloud). The
remaining buttons beside **Copy** are described in
[Controls that are not available yet](#controls-that-are-not-available-yet).

The strip under the message box shows the connection state (**Connected**, **Reconnecting...** or
**Offline**) and the line **AI can make mistakes. Always check facts and important information.**

Text you have typed but not sent yet is kept in this browser on your device, so it comes back even
after you close the tab, reload the page, or the browser crashes. Drafts are kept per chat - and
also for the message box before a chat is open - and are dropped once the message is sent, when
you log out, or after a week. Attached files are not kept this way: attach them again after a
reload.

### Reply to a message

Hover over a message (yours or your assistant's) and select **Reply** to quote it. A chip above
the message box shows who wrote it (**You**, or your assistant's name) and the first part of the
message, about 140 characters. Select the chip's **x** (**Dismiss reply**) to remove the quote
before sending. Your next message carries the quote, and your message shows the same quote above
it from then on, including when you reopen the chat later.

* **Reply** only appears on messages that have been saved to the chat. A message you have just
  sent, or a reply that has just finished, may not offer it until you reopen the chat: switch to
  another chat and back, or reload the page.
* A quote belongs to one chat. Switching to another chat discards it, and sending a message uses
  it up, even if that send fails.
* A quote on a message from an earlier session - one above a **New session** divider - keeps
  working: the reply still quotes the older message, and the quote stays with your message after
  the send.
* You cannot quote a message while a reply is still being written. If you try to add to the
  current turn, or queue, with a quote in place, the message box shows **Replying isn't supported
  yet while a reply is streaming. Wait for it to finish, or remove your reply and send a plain
  message.** Nothing is sent, and your quote and text stay where they are.

### Paste formatted text

Pasting into the message box keeps the structure of what you copied. Text copied from Word, an
email or a web page is converted to markdown as you paste: headings, lists, bold, links and code
survive, while scripts, embedded frames, HTML comments and pictures that are not on the web are
dropped. Pictures from a web page keep their web address; whether the conversation shows them is
the same rule as any other image (see [Screens with images](#screens-with-images)).

Press Cmd+Z (Ctrl+Z on Windows and Linux) to undo a paste: the text the paste replaced comes back
exactly as it was.

A paste of plain text longer than 10,000 characters does not go into the message box: it is
attached as a text file named **pasted-text.txt** instead, the same way an attached file is (see
[Attach files](#attach-files)), so a copy of a long document does not flood the conversation.
Shorter pastes, and pastes with formatting, are inserted as before.

Pasting a picture (for example a screenshot on your clipboard) attaches it as if you had used the
attach button, including the automatic shrinking described in
[Attach files](#attach-files). Ordinary text paste - including paste without formatting, such as
Shift+Cmd+V - works exactly as it always did.

### Read a reply aloud

Hover over a reply and select the speaker button (**Read aloud**) to hear that reply spoken. The
button plays the whole reply, however long it is: long replies are spoken in consecutive pieces
with no action needed from you. While it plays, the button becomes **Pause reading**; select it to
pause, and it becomes **Resume reading**, which continues from the same point. A separate
**Stop reading** button beside it stops the read. Pressing the microphone (**Dictate**) also pauses
a playing read instead of talking over it: the read stays paused when you stop the take, and only
**Resume reading** (or starting another read) continues it. Starting another reply, starting Voice
mode, or switching to another chat stops playback too. A dropped connection does not stop it: the
read continues from where it was once the connection is back.

Your assistant reads the text of the reply: paragraphs, headings and list items, spoken in order.
Code blocks, tables and embedded HTML are skipped, so a reply that contains only code or a table
shows **Nothing to read aloud in this reply.**

If the button is dimmed instead, hover over it to see why. **Read aloud — unavailable while
offline** means the connection is down; it comes back on its own once you are connected again.
**Read aloud — unavailable during Voice mode** means a Voice conversation is active (listening,
thinking or speaking); the button comes back when Voice mode is off. **Read aloud — unavailable on
this gateway** means the person who manages WednesdayAI has turned the feature off, or has not
configured a voice for your assistant to speak with. Ask them to enable it; see
[Text-to-speech](https://docs.wednesdayai.dev/next/users/tts.md) for the setting.

If reading fails, a short message appears beside the button: **Read aloud is turned off for this
assistant.**, **Couldn't play the audio in this browser.**, or **Couldn't read aloud:** followed by
the reason. Nothing is sent to your assistant when you use the button.

### Read replies aloud automatically

The speaker button beside the microphone turns on **Read replies aloud**. While it is on, every
new reply in the chat you have open starts playing on its own as your assistant writes it: the
first sentence is spoken as
soon as it exists, and the rest follows sentence by sentence without waiting for the reply to
finish. The button stays pressed while the setting is on, and this browser remembers it for your
WednesdayAI connection the same way it remembers the theme.

* Turning the toggle off stops the reading at once. Switching to another chat or starting Voice
  mode stops it too, but a dropped connection does not: the rest of the reply is still read out
  once the connection is back. Selecting **Stop** on the reply ends the reply but does not stop
  the voice: the rest of that reply is still read out. Turning the toggle off is what silences
  it.
* Long replies are spoken in consecutive pieces, like **Read aloud**; each piece is capped so the
  voice never sends too much text at once.
* If the button is dimmed, hover over it to see why. The reasons are the ones **Read aloud** uses:
  **unavailable while offline**, **unavailable during Voice mode**, or
  **unavailable on this gateway**.

While the toggle is on, dictation becomes hands-free as well. The tooltip reads
**Read replies aloud. Dictation sends when you pause**: when the microphone has heard you speak
and you then stay quiet for the chosen pause, the dictation ends by itself and the words are sent
straight away - one tap to start, speak, then just pause. A chip beside the speaker shows the
current pause - **1.5 s**, **2.5 s** or **4 s**, with **2.5 s** as the default. Select the chip to
choose a different pause; this browser remembers your choice the way it remembers the toggle, and
a change applies to the next dictation, not to one that is already listening. With the toggle off,
the chip is not shown and Dictate works as before: you select the microphone button to stop, edit
the text, and send it yourself.

While a reply is being read aloud, pressing the microphone pauses the reading instead of talking
over it; see [Read a reply aloud](#read-a-reply-aloud).

With auto-send on, a reply being written no longer holds your words back: the dictated text is
added to the current turn the way Enter adds it - or queued, when a file is attached. Only a
message box that cannot send at all (for example while a file is still uploading) keeps the text
in the box for you to send. A dictation that heard nothing (for example after denied microphone
access) never sends anything, including text you had already typed in the message box.

### Choose a model and Think

The button at the top of the screen, beside the menu button, shows your assistant's name and the
model this chat is using, for example **Assistant name · Model name**. Select it to open a menu
with **Assistant default** and every model your assistant is set up to use. A tick marks the
current choice. Picking a model applies to this chat only and takes effect on your next message;
other chats are not affected. If the change is refused, the reason appears under the button and
the tick stays where it was. Picking a model with no chat open shows **No chat is open**.

On a phone or a narrow window the header shows the conversation's **session key** under the
assistant·model button instead of the title - the same identifier each sidebar row carries, so you
can tell which conversation you are in when titles are hidden. Selecting the button still opens
the model menu. To leave the key room, the other header controls (the unread bell, **Display** and
**Canvas**) show as icons only.

The **Think** button in the message box opens a level menu (**Off**, **Low**, **Medium**, **High**)
plus a **Send reasoning** switch. These apply to this chat only, like the model. A tick marks the
current level. **Send reasoning** controls whether reasoning is delivered with replies; it does
not control what this browser displays.

**Think** is dimmed when the chat's current model does not support extended thinking. Hover over it
to see **Think — \[model name] does not support extended thinking**. Switch to another model, or
pick one that does, to use it. A refused change shows its reason under the button.

### Choose what the conversation displays

**1. Open Display**

Open **Display** in the top bar.

You should see **Show reasoning** and the **Hide**, **Minimum**, **Collapsed** and **Full**
choices under **Tools**.

**2. Choose reasoning visibility**

Turn **Show reasoning** on if you want to see reasoning included with replies.

Replies that contain reasoning show a collapsed **Reasoning** section that you can expand.

**3. Choose tool details**

Choose how tool activity appears.

The conversation updates immediately: **Minimum** shows a one-line summary per tool call,
**Collapsed** shows compact expandable rows, **Full** shows each call with its result, and
**Hide** removes normal tool details while leaving errors and failures visible.

The display choices work as follows:

* **Show reasoning** is off by default. Turn it on to reveal collapsed **Reasoning** sections
  when a reply includes reasoning, then expand a section to read it.
* **Tools → Minimum** is the default: each tool invocation appears as a single line, such as
  `📝 Edit: in ~/Code/x/README.md`, and failed calls end in **· Failed** with the error shown.
* **Tools → Collapsed** shows each tool invocation as a compact expandable row, with its result
  paired when available. Expand it to inspect arguments and the result. (This was called
  **Minimal** in older versions; an older saved **Minimal** choice loads as **Collapsed**.)
* **Tools → Full** shows each call with its result inside the same card, including their
  structured details.
* **Tools → Hide** removes normal tool call and result details, but keeps tool errors and failures
  visible so you still know when something went wrong.

These choices are remembered in this browser for this WednesdayAI connection. They do not change
the model's Think level or **Send reasoning**. **Show reasoning** only changes whether reasoning
already included with a reply is displayed; it cannot generate reasoning or add it to a saved
reply. Completed replies load their saved tool details without requiring a page reload.

### Keep chatting while a reply is streaming

You do not have to wait for a reply to finish before sending your next message:

* Press **Enter**, or select the send arrow (**Add to current turn**), to add your message to the
  current turn: your assistant reads it before its next step, without waiting for the reply to
  finish. While the reply is still streaming, the message waits at the bottom of the thread
  marked **Queued** until the assistant takes it - it then joins the reply where the conversation
  has reached, or starts its own turn. While one of your messages is already waiting in the
  queue, Enter queues your message behind it (in the order you sent them) instead of adding it
  to the current turn. On a phone or tablet, Enter adds a new line instead - tap
  the send arrow. If your assistant cannot take the message right now, it is queued instead and
  marked **Queued**.
* With a file attached, the send arrow reads **Queue message** instead, and queues the message and
  the file for after the reply.
* To queue on purpose, select the small arrow beside the send arrow (**Send options**) and choose
  **Send after reply**, or press **Ctrl+Enter** (**Cmd+Enter** on a Mac) on a computer. A queued
  message appears in the thread marked **Queued** and is sent as soon as the current reply
  finishes; queued messages run in the order you sent them. The label goes away as soon as your
  assistant starts on the message: the first streamed text, tool activity or reply of that run
  clears it, and so does the start of a later message of yours, which proves the earlier queued
  ones have run. Until then, queuing keeps working.
* **Stop** ends the current reply early at any time, including a reply started in another browser
  or tab.

While the reply is in progress, the message box shows **Add to \[your assistant's name]'s current
turn...** as its hint text.

Whichever way you send it, your message keeps its place in the thread. Messages appear in the
order they were created and never reshuffle: a message sent while a reply is streaming waits at
the bottom until it joins that reply or starts its own turn, and one sent while another message
of yours is already waiting queues behind it. Once a message is part of the
saved conversation its position is fixed - reconnecting or reloading the page shows the same
thread in the same order.

A queued message whose run has already started is part of the saved conversation, so it comes
back after a reload in its place, without the **Queued** label. One that has not started yet is
not shown after a reload in the same browser tab until the assistant starts on it - the client
cannot see the gateway's waiting queue - but the message itself is not lost: it still runs when
its turn comes, and its reply appears then. If the gateway itself restarts first, the waiting
message is gone.

Your chat is one continuous conversation, even when the assistant starts a fresh session behind
the scenes. Depending on your session reset settings (`session.reset`), WednesdayAI starts a
fresh session after a period of inactivity or once a day, and always when you send `/new`, so
the assistant can run its end-of-session journal and memory work. When that
happens, a **New session** divider appears in the thread where the fresh session began; every
earlier message stays above it in its original order, and **Load older** at the top of the thread
keeps going into those older sessions.

This depends on the server being recent enough. If the gateway is older than the web client, the
thread shows only the current session: earlier sessions are not fetched, no **New session**
divider appears, and **Load older** reaches only the current session. Ask the person who manages
WednesdayAI to update the gateway.

One known gap is worth knowing:

* After a connection drops and comes back, a queued message can keep its **Queued** label, and
  the message box can keep behaving as if a reply were still running. The label clears when the
  conversation is fetched again (switching chats or reloading does it), or when a later message of
  yours finishes: the client then checks on the earlier message and clears it once its reply has
  finished.

Adding to the current turn needs some text, and cannot carry a file. Attachments go with a queued
message, but the **+** button is dimmed while a reply is being written, so attach files before you
start or wait for the reply to finish.

If your assistant refuses the message, the thread shows **Could not add to the current turn:** or
**Could not queue that message:** followed by the reason (for example `dropped_by_queue_policy`,
when the assistant's queue settings discard the message). The text you typed returns to the
message box so you can send it again; any attachments do not come back. If the reason is unclear,
ask the person who manages WednesdayAI.

### Dictate and voice

Two buttons beside the send arrow turn your voice into a message: a microphone (**Dictate**) and
a wave (**Voice mode**). Whichever one is on stays pressed and highlighted, so the buttons show at
a glance whether they are off or on. Either may be dimmed, independently, depending on what the
person who manages WednesdayAI has configured — see
[Controls that are not available yet](#controls-that-are-not-available-yet) if both are dimmed for
you.

**Dictate** turns your speech into typed text in the message box, without sending anything on its
own:

1. Select the microphone button and allow microphone access when your browser asks.

   The button stays pressed while Dictate is on.

2. Speak. Your words appear inside the message box as you talk, greyed out and replacing
   themselves as they are recognized. While Dictate is on, the message box cannot be edited and
   Send is unavailable; while it is still empty, its hint reads **Listening...**.

3. Select the microphone button again to stop. The button reads **Transcribing...** briefly while
   your last words are recognized, and cannot be pressed then. The words are then placed where
   your cursor was - or at the end of what you had typed - and you can edit them and send as
   usual.

If stopping fails, the words heard so far stay in the message box, so nothing you said is lost.
Your first words are kept too: the microphone starts the moment you select the button, and anything
captured while the transcription service is still connecting is sent when the service is ready. On
a slow connection, stopping can take a few seconds while that service catches up with your words.

If a take fails part-way - for example your microphone stops sending audio, or the transcription
stops unexpectedly - the microphone button leaves its recording state within a few seconds and a
red message below the message box names what happened (**Your microphone stopped sending audio.
Tap Dictate to try again.** or **Transcription stopped unexpectedly. Tap Dictate to try again.**).
The words heard so far stay in the message box. Select the microphone again to start a fresh take;
you never need to reopen the chat.

When you send a dictated message, your assistant is told the text came from speech. That way it can
allow for mis-heard names and similar-sounding words. This note is not shown in the chat bubble;
you only see the text as dictated.

Pressing the microphone while a reply is being read aloud pauses the reading instead of stopping it
(see [Read a reply aloud](#read-a-reply-aloud)). Stopping the take does not restart the read: it
stays paused until you select **Resume reading** on that reply.

**Voice mode** has a live, spoken conversation with your assistant instead:

1. Select the wave button and allow microphone access when your browser asks.

   The button stays pressed while Voice mode is on. Above the message box, a pulsing dot shows
   **Listening**, **Thinking** or **Speaking**.

2. Speak normally. Your assistant answers aloud, using the same chat and its recent history as
   typed messages. Live captions show what it heard you say and what it is saying. Once each turn
   is saved, the exchange appears in the thread as text.

3. To interrupt, start talking once the actual reply is speaking. Playback stops almost immediately
   and you can keep talking. The chat keeps only the part of the reply you heard, marked
   **interrupted**.

4. Select the wave button again (**Stop voice mode**) to end the conversation and release the
   microphone.

The microphone starts the moment you select the wave button: your first words are captured while
the connection to the voice service is still opening and are sent once it is ready, so nothing
you say first is lost.

If your microphone stops sending audio - for example another app has taken it over - Voice mode
ends the conversation on its own after about three seconds of silence and shows **Your microphone
stopped sending audio. Tap to start voice again.** Select the wave button again to start a fresh
conversation.

A short phrase may play while your assistant is still thinking (for example "One moment."), and
another when it starts using a tool on your behalf (for example "Working on that."). Both are
normal. Talking during those short phrases does not interrupt the reply. You can interrupt once
the actual answer starts speaking. A cough, a short "mm-hmm", or your assistant's own voice
reaching the microphone should not cut the reply off either.

Your WednesdayAI may be set up for a faster, more natural voice that talks with you directly. In
that setup, an interrupted reply appears in the chat only as **\[interrupted]** (what you said is
still saved), and your assistant can look things up but cannot make changes by voice, for
example editing files. Type those requests instead. In that setup the voice service detects your
speech itself, so talking over the short acknowledgement can cut it off too.

If the strip under the message box changes to **Reconnecting...** or **Offline**, the microphone
is released. Voice mode returns to the unpressed wave button, even if it was still starting.
Dictate stops too, and the microphone button's tooltip reads **Voice input stopped because the
connection dropped.** The words heard before the drop stay in the message box, ready to edit and
send. Wait until the strip says **Connected**, then select the button again.

If either button shows an error (for example after you deny microphone access, or if this browser
cannot use a microphone), the button gets a red outline and its tooltip shows the reason. Select
it again to retry. Both release the microphone fully when you stop them, when the connection
drops, or when you leave the page.

While Dictate, Voice mode, or a read-aloud is playing, the web client asks your browser to keep
the screen awake, so a phone or tablet does not lock in the middle of a sentence. This is best
effort: the browser may refuse it (for example while Low Power Mode is on), and home-screen apps
on iPhone or iPad need iOS 18.4 or newer; nothing on the page changes when the browser cannot or
will not hold the lock.

### Share your location with messages

The map-pin button beside the microphone turns on **Share location**. While it is on, every
message you send - typed or dictated - carries your device's position when it is sent (a
position your browser read in the last minute may be reused). The button is off by default, and this browser remembers your
choice for your WednesdayAI connection the way it remembers the theme.

1. Select the map-pin button and allow location access when your browser asks.

   The button stays highlighted while sharing is on.

2. Send a message as usual.

   Your assistant receives the position with that message and that message only: it is not kept
   in the chat history, and it does not appear on later turns. A new position travels with each
   message only while sharing stays on.

Your assistant is told where you were when the message was sent, including how accurate the
position is, together with a note to use it only when your request depends on where you are -
directions, travel times, nearby places, local time or weather.

Notes and limits:

* **The button is dimmed** when this page is not on HTTPS (or `localhost`), with a tooltip
  explaining that HTTPS and browser location access are needed - browsers only offer location to
  secure pages. Ask the person who manages WednesdayAI for the HTTPS address.
* **If you denied location access**, sharing turns itself off and a message under the message box
  says so. Allow location for this site in your browser's site settings, then select the map pin
  again.
* **If a position cannot be read when you send** (for example a weak signal indoors), the message
  is still sent, without a position, and a short message explains that. Location never blocks or
  delays a message by more than a few seconds.
* **Voice mode** conversations do not attach your location yet.
* Turning the map pin off stops the position from travelling with your next messages; the
  assistant is not told anything about the change.

### Attach files

Select the **+** button (**Attach files**) in the message box, or drag files onto it. You can
attach several at once, up to 32 files, each up to 50 MB (images show a thumbnail; other files
appear as a card). There is no limit on the total size of a message's attachments. Each file is
sent to your assistant's server over its own secure, authenticated upload while you keep typing:
the attachment card shows the upload's progress, and Send stays unavailable until every file has
finished uploading. Selecting a card's **x** while it is still uploading cancels that upload and
removes the card. Large pictures are shrunk automatically first: an image wider or taller than
2048 pixels is scaled down to fit inside that size and re-encoded, so a photo that is too large
attaches instead of being refused. Pictures are never enlarged, animated GIFs are left untouched,
and an image that still cannot be made small enough is rejected like any other oversized file.
Images are described automatically when the chat's model cannot see images. A file that is empty
or larger than 50 MB, or a message with more than 32 files, is rejected before anything is sent,
with a message under the message box explaining why. The **+** button is dimmed while a file is
uploading or a reply is being written. Attached files belong to the chat you attached them in:
opening another chat before you send removes them (and cancels any upload still running), so a
file never goes to a different assistant by accident.

Inside the WednesdayAI app, files you share to the app from another app (for example a photo from
your gallery) appear in the message box as attachment cards, already uploaded and ready to send
with your next message. The 32-file limit counts them too; a file over the limit is named in a
message under the message box.

### See older messages

Scroll to the top of a conversation and select **Load older messages** to fetch earlier history.
The button reads **Loading...** while it works. Your place in the conversation stays put while
older messages load. The button disappears once there is nothing older to load. If loading fails,
the thread shows **Could not load older messages:** followed by the reason; select the button to
try again.

The conversation follows the newest message on its own, but only while you are near the bottom. If
you scroll up to read something earlier, new text does not pull you back down.

## Switch between chats

Each separate conversation is a **session**. Your sessions are listed in the sidebar under
**Sessions**, with the most recently updated first. Each row shows the conversation's title and its
**session key**, the full identifier of that conversation, such as
`agent:main:telegram:direct:example-user`. Long keys are shortened visually; hover over a row to
read its full key. A colour dot identifies the assistant, and a small tag identifies conversations
from another app, such as Telegram or Discord. A session whose assistant is working on a reply, or
that has unseen replies, also shows a small dot beside its title - see
[Use several browsers or devices at once](#use-several-browsers-or-devices-at-once).

With no sessions yet, the list reads **No sessions yet**.

Scheduled jobs, isolated heartbeat checks, subagents, conversations your assistants hold with
each other, and other internal conversations are hidden by default. Turn on **Show internal**
beside **Sessions** to include them. They appear in their own **Internal** group at the end of the
list. A normal conversation that also contains heartbeat activity is not hidden.

A message that reached this chat through another conversation - for example, when your assistant
handed part of a task to a specialist and the specialist's answer is relayed back to you - shows
**From \[assistant name]** above it. If the sending assistant is not one you can see in your
assistant list, the label reads **From another assistant**.

Use the **Sort sessions** control to choose **Newest**, **Oldest**, **Title A-Z**, **Title Z-A**,
**Key A-Z** or **Key Z-A**. Sorting, assistant filtering, search and **Show internal** apply to your
entire conversation history before a page of results is shown, not only to sessions already loaded
in the sidebar. Time-based sorts group sessions into **Today**, **Yesterday** and **Earlier**;
alphabetical sorts show one ungrouped list.

* Select a chat to switch to it. If the chat cannot be loaded, the thread shows **Could not load
  this chat:** followed by the reason.
* Select **Browse assistants** at the top of the sidebar to open the assistant picker: a list of
  every assistant you can talk to, each with its own colour. Some assistants work behind the
  scenes — a specialist your assistant hands part of a task to, for example — and those are not
  listed; you reach them through the assistant you normally talk to. Hovering an assistant, in the
  sidebar or the picker, shows its description when it has one. Pick an assistant to open its
  conversation right away; the assistant stays selected afterwards. Close it with **Close**,
  **Escape**, or by clicking outside it.
* Or start typing in the message box with no conversation open: the first send opens the
  conversation for the selected assistant automatically.
* Type in the **Search sessions** box at the top of the sidebar to filter the list. The list
  updates shortly after you stop typing; select the **x** in the box (**Clear search**) to see
  every conversation again.
* Once you have more chats than fit on one page, a **Load more** button appears at the bottom of
  the list.

Switching chats while a reply is still being written is fine: the new chat's message box works
straight away.

While **Dictate** or **Voice mode** is on, switching chats is blocked: selecting another chat, the
unread bell, or an assistant only shows a notice - **Stop dictation to switch chats.** or **Turn
off voice to switch chats.** Stop the one that is on, then switch.

On a wide screen, the button in the sidebar header collapses the sidebar to a thin strip; select
the strip's button to bring it back. On a phone or a narrow window the sidebar is hidden. Open it
with the menu button at the left of the top bar, and close it with the **x** at the top of the
sidebar or by tapping outside it. On a phone the page itself never scrolls: the conversation
scrolls inside its own area, and the top bar and message box stay where they are - including with
the on-screen keyboard open - and typing into a box never zooms the page. Wide content - a long
link, a wide table or a large picture - never pushes the page sideways: it scrolls or shrinks
inside its own box instead.

The **Assistants** section lists assistants with an eligible chat updated within the last
**4 hours**, by default. Choose **1h**, **2h**, **4h**, **8h**, **16h**, **24h** or **All** with the
activity-window control. **All** includes assistants with no chats. The assistant you have selected
stays pinned and visible even when its chats are older than the window. An assistant you pick in
the **Browse assistants** window also stays in the list for the rest of this browser session,
even once its chats fall outside the window; that memory is kept for the page you have open, so a
reload returns the list to the ordinary rules. The assistant for the chat
you currently have open also stays visible when you have not selected a filter.

Use **Sort assistants** to choose **Recent**, **Name A-Z** or **Name Z-A**. **Recent** lists
assistants with a chat updated in the last 24 hours first, most recent first, then everyone else
alphabetically; assistants updated at the same moment are also listed alphabetically. The list
holds still while you use it and re-sorts when the page loads, when you come back to the window
or tab, or when you change a sort or activity-window choice. An assistant that newly appears is
added at the bottom until then. Recent uses each assistant's most recently updated eligible chat.
Internal conversations contribute only when **Show internal** is on. The activity window narrows
the assistant list, not the age of chats shown after you select an assistant.

Select an assistant to highlight it, show its chats and open your conversation with that
assistant: the same one your direct messages on other channels use when the gateway links them.
It opens with its existing history, and the browser reopens it on the next visit. Selecting the
assistant you are already talking to does not reload the conversation.

Selecting the same assistant again keeps it selected; it does not clear the selection. Select
**All assistants** when you want to clear the assistant filter and show chats from everyone. Your
currently open chat stays open. Sorts, filters, the selected assistant and remembered chats persist
across reloads in this browser. Each assistant's colour also marks its chats and appears beside its
name in the top bar.

### Rename a chat

Select the chat's title at the top of the screen (it reads `/` followed by the title) to edit it in
a box labelled **Chat title**. A title can be up to 60 characters.

* Press **Enter** to save. Clicking elsewhere also saves. Press **Escape** to cancel.
* An empty or unchanged title closes the box without saving anything.
* If the change is refused, the reason appears beside the box and the box stays open. For example,
  `label already in use: [title]` means another chat already has that title.

The sparkle button beside the box (**Suggest a title**) writes a title for you from the first
message shown in the thread and puts it in the box. It does not save: press **Enter** to keep it,
edit it first, or press **Escape** to discard it. Until the chat has a message, the button does
nothing and its tooltip reads **No message to summarise yet**.

Suggesting titles is limited to 3 per minute. Past that, the box shows **rate limit exceeded for
sessions.title.prepare; retry after 60s** (the number of seconds varies). Wait that long and select
the button again, or type a title yourself. If no title could be produced, the box shows **Could
not suggest a title**.

### Archive a chat

Select **Archive chat** in the top bar to tuck the chat you are reading away. Nothing is deleted:
the conversation and its history stay exactly as they are. On an archived chat the button reads
**Unarchive chat** and puts it back.

* Archived chats are hidden from the sidebar. To see them, turn on **Show archived** in the menu
  beside **Sessions** (the menu that also has **Show internal**): they appear in their own
  **Archived** group after **Internal**. The chat you are reading stays in the list even when it
  is archived.
* Archiving is shared across your devices: archive a chat on your phone and it disappears from
  the sidebar on your computer too.
* A new message in an archived chat does not bring it back on its own, and neither does the
  fresh session the gateway starts behind the scenes after an idle period, once a day, or when
  you send `/new`; unarchive the chat if you want it listed again.
* If the gateway refuses the archive or unarchive, a red message appears beside the archive
  button and the chat stays where it was; try again.

### This chat changed elsewhere

Each send is checked against the latest message this window has shown you. If the chat has moved on
since then - because you or your assistant added to it from another browser, another device or
another app - your message is still sent: the thread refreshes to show the messages you had not
seen, and the message is sent again automatically against the refreshed chat. Attachments and a
reply quote go with the resent message. You normally notice nothing except the new messages
appearing above yours.

A banner reads **This chat changed elsewhere. The latest messages are shown below.** only when the
chat changes a second time while that automatic retry is under way. The text you typed is put back
in the message box (an attached file or a quote is not), so read what is new and press Enter again
if you still want to send it. If you switch to another chat while the automatic resend is being
prepared, the resend is cancelled and your text is put back in the message box instead. Select
**Dismiss** to hide the banner; it also clears on your next send or when you switch chats. Queued
messages, and messages added to the current turn, are not checked this way.

## Use several browsers or devices at once

You can have the web client open in several browsers, tabs and devices at the same time. A reply
appears in every one of them without reloading the page: the window that sent the message shows
the reply as it is written, and the others show the working indicator until it is ready, then the
full exchange.

While your assistant is working on a reply anywhere, the chat shows it in three ways:

* The thread shows the three-dot indicator described in
  [Send a message](#send-a-message), including in a chat that does not have any messages yet.
* The chat gets a small pulsing dot in the sidebar.
* The message box offers **Stop**, **Add to current turn** and **Send after reply** as usual, and
  they work on that reply even though another device started it.

Sending into a chat that gained messages elsewhere meanwhile is covered in
[This chat changed elsewhere](#this-chat-changed-elsewhere).

A chat counts as **unread** in this browser (or installed app) when an assistant reply arrived
since you last had that chat open here. Replies a chat already had the first time this browser
sees it do not count: the unread baseline starts at what is on screen. Unread chats show a small
dot in the sidebar - on the chat's row and on the chat's assistant - and an assistant with unread
replies stays listed even when its newest chat is older than the activity window. The bell in the
top bar shows a count of unread chats. Select the bell to see the list: choose a chat to open it,
or choose **Mark all as read** to clear every unread mark at once. That choice is offered only
while something is unread; with nothing unread the panel reads **You're all caught up**. The
tab's title shows the number of unread chats, for example **(2) WednesdayAI**. Unread marks are
kept in this browser (or installed app) and are not shared with your other devices, so they
survive closing the page, the device sleeping and a dropped connection. Opening the chat clears
its mark, and so does returning to the tab while it is open. If a chat is deleted elsewhere while
you are away, its unread mark is dropped the next time the chat list is read in full on one page
and that chat is not in it. These work whether or not you turn on the notifications described
below.

To also get a desktop notification when a reply arrives while the tab is hidden, open the account
menu (select your name at the bottom of the sidebar), find **Notifications** and choose **On**.
Your browser asks for permission; choose **Allow**. While it is on and the tab is hidden, a
notification names the assistant that replied and the chat it arrived in - never the reply text -
and selecting the notification opens that chat. The choice is remembered in this browser for this
WednesdayAI connection, like the theme.

If your browser has blocked notifications for this site, the menu shows
**Blocked in browser settings** and the choice stays **Off** until you re-allow notifications for
this site in the browser's settings. On browsers without desktop notification support - most
mobile browsers - the menu shows **Not supported in this browser**; there, the bell, the sidebar
dots and the title count are the notification.

Inside the WednesdayAI app, which shows this web client in its own window, notifications come
from the app instead of the browser. Choosing **On** asks for no browser permission; the app shows
the notification (your phone or computer may ask once whether the app may notify you), and
selecting it opens that chat. The app also follows the chat you have open.

Coming back to a tab you left in the background also catches the open chat up: messages that
arrived elsewhere are fetched and shown without a reload.

## Forms and other assistant screens

Your assistant can do more than write text: it can build small screens - a form to fill in, a list
of buttons to choose from, a table, a simple chart, or a progress bar - and show them right in the
reply that produced them, directly after the step that showed the screen. The screen stays there
when the reply finishes and after you reload the page. A screen from a reply another device started,
or from the reply to a form you sent, appears when that reply finishes.

### Fill in a form and send it

When your assistant needs an answer, it may show a form instead of asking a plain question.

**1. Read the screen**

The form appears in the reply right after the step that showed it, with a title and the fields
your assistant chose - text boxes, checkboxes, a date picker, a table or a chart. You can change
any field before you send.

**2. Select the button**

Select the button the form provides, for example **Book**. Your answer is sent to your assistant
and its next reply takes your answer into account.

If your assistant is still busy with a reply, the button's label changes to **Queued** until the
answer can be sent.

After your answer is delivered, the screen is locked so it cannot be sent twice: the fields stop
accepting input, the button stays dimmed, the form shows **Submitted**, and your answer shows in the thread as a chip
that reads **You submitted "Book"** (the button's label). One form accepts one answer; if you need
to change it, ask your assistant for a new form.

If the button is dimmed while your assistant is still working, it may be waiting for the answer to
be accepted; it unlocks again on its own. If sending fails, a short message explains why - for
example **Could not reach the assistant. Try again.** or **This form was already submitted.** Ask
your assistant for a fresh form if the message stays.

### When a form is out of date

Your assistant may update a form while you are still looking at an older copy of it. The older
copy keeps your place but shows a banner that reads **Updated: open latest in Canvas**. The newest
version of the screen always lives in the Canvas panel (see below), and answering the old copy is
refused with a message that says the form has been updated. Open Canvas, find the newest version
there, and answer that one instead.

## The Canvas panel

The **Canvas** button in the top bar opens a panel beside the chat that collects every screen your
assistant has sent in this conversation - including older ones you scrolled past. Select **Canvas**
again to close it.

**1. Open the panel**

Select **Canvas** in the top bar.

The panel opens showing the screens your assistant has published in this chat, newest first, and
any screen you pinned at the top.

**2. Work with the screens**

Each screen shows its title and the same fields as the copy in the chat. Fill one in and select
its button to send your answer the same way as inline.

You should see the same **Queued** and **Submitted** labels and the same messages as in the chat.

* **Pin a screen** with its **Pin** button to keep it at the top of the panel. **Unpin** puts it
  back in date order. Your pins are kept even after you reload the page.
* **Open in Canvas** - the button on a screen in the chat - jumps to that screen in the panel and
  highlights it, so you can find a specific form quickly.
* On a phone or narrow window, the panel opens as an overlay over the conversation; select the
  close button (or press Escape) to return to the chat.
* A new conversation starts with an empty Canvas: **No assistant UI in this chat yet.**

## Watch your assistant work

While your assistant works on a reply, each tool it uses appears in the reply as it happens, as a
tool line in your **Tools** display form, between the parts of the reply text it came between.
While the reply is still being written, the line shows the tool's name and what it is doing (for
example the file it is reading or the command it is running); when the reply finishes, the same
line gains the step's result without moving. Each sentence of reply text arrives in full before
the next tool line appears. Nothing you have seen disappears or jumps when the reply finishes,
and a reload shows the reply the same way. A reply that used no tools shows no tool lines. With **Tools → Hide**, tool lines
stay hidden, except errors and failures. A hidden announcement area tells your screen reader the
name of each tool step as it starts; when tool lines are hidden, nothing is announced there.

Each reply is one block: one name, one time and one set of actions. **Copy** and **Read aloud**
cover the whole reply, and **Reply** quotes its last message. Replies saved before this version
of WednesdayAI can still show as several blocks, one per saved message, because they were stored
without a record of which reply each message belongs to.

If a reply fails partway, the text already written stays and the error appears below it. The error
keeps its place when later replies arrive; it is not kept after you reload the page.

## Read and triage your inbox

The **Inbox** brings mail and messages from your connected accounts (for example Gmail or
Microsoft 365) into one list, so you can deal with what needs you without opening each app. Your
assistant reads the same inbox and handles what it can; the Inbox shows you the rest.

The **Inbox** entry appears in the sidebar only when your WednesdayAI setup has an inbox. Its
number counts the messages still unread in your own accounts.

**1. Open the Inbox**

Select **Inbox** in the sidebar. The Inbox opens beside your chat, showing the queue: the
messages that need you.

**2. Work through what needs you**

Each card is one conversation that needs you. Use the buttons on the card to deal with it.
The card leaves the queue once nothing in it needs you any more.

**3. Look for something specific**

Select **Explorer (legacy)** at the top of the Inbox to browse every conversation by account
and folder. Select **Queue** to go back.

### The queue

On the left, **Needs you** shows the same number as the sidebar: your unread messages. Below it,
one row per account you own (for example `gmail · you@example.com`) shows that account's unread
messages; select a row to open that account in the explorer.

A conversation needs you when one of its messages is unread, when your assistant flagged a message
for your attention (the card shows the flag, for example `financial`), or when your assistant could
not decide what kind of message it is (the card shows `Unclear`). If nothing does, the queue says
**Nothing needs you.** Messages your assistant already dealt with are collected under **Handled by
agent**, closed by default, with their count; select it to see them.

### Act on a message

A card shows only the buttons that message's account supports:

| Button | What it does |
| - | - |
| **Mark read** | Marks the message read in its own account, for example in Gmail. |
| **Mark unread** | Marks a read message unread again. |
| **Archive** | Archives the message in its own account. It stays findable in the explorer. |
| **Label** | Type or pick a label, then select **Apply**. |
| **Reply (needs approval)** | Write your reply, then select **Send**. The reply goes out only after approval. |

There is no delete button: the Inbox never erases a message.

While an action is running, the other action buttons are dimmed. If the action is refused, for
example because a reply was not approved, the reason appears next to the buttons and your reply
text stays in the box so you can try again.

### Browse with the explorer

**Explorer (legacy)** lists every conversation, newest first:

* The tabs along the top pick an account type: **All**, or one tab per source such as `gmail`.
* The folders pick which conversations to list:
  * **Inbox**: everything not archived.
  * **Unread**: conversations with a message still unread in its own account.
  * **Review**: messages your assistant flagged for your attention or could not classify (`Unclear`).
  * **All incl. archived**: also archived conversations.
  * **All incl. deleted**: also messages deleted in their own account. They are marked **deleted
    at source**.
* Under **Labels**, select a label to show only conversations with that label; select it again
  to show all.
* Select a conversation to read it. Each message shows whether you **Received** or **Sent** it,
  with the same buttons as in the queue.
* Select **Load older** at the end of the list to fetch earlier conversations.

### When the Inbox is empty

An empty Inbox says why it is empty and what fixes it:

| What you see | Why | What to do |
| - | - | - |
| **This device isn't linked to a person.** | WednesdayAI does not know who uses this browser, so it cannot tell which accounts are yours. | Send the person who manages WednesdayAI the command shown on the screen. It includes this browser's device id: `wednesdayai devices link <deviceId> --identity <name>`. |
| **No inbox accounts are bound to you.** | WednesdayAI knows who you are, but no connected account is marked as yours. | Ask the person who manages WednesdayAI to run `wednesdayai inbox owners bind <your name> <source:accountId>` (`wednesdayai inbox owners list` shows the accounts). |
| **No inbox sources are connected.** | No mail or message account is connected yet. | Ask the person who manages WednesdayAI to connect one. |
| **Sources connected, nothing here yet.** | Your accounts are connected but no message has arrived or been fetched yet. | For Gmail, Microsoft 365, BlueBubbles or WhatsApp, the person who manages WednesdayAI can fetch existing messages with the command shown. If Microsoft 365 is signed out, they sign it in from the Control UI. Other sources fill up as new messages arrive. |
| **Couldn't load the inbox.** | The Inbox could not reach WednesdayAI. | Select **Retry**. If it keeps failing, see [The page shows a connection error](#the-page-shows-a-connection-error-and-never-loads-chat). |

## Screens with images

Replies can contain images, and the web client is careful about which ones it loads:

* Images your assistant serves itself, and images attached to the conversation, are shown as
  pictures.
* Images from other websites are not loaded on their own - they appear as a link you can follow,
  labelled with the image's description (or its web address when it has none). This keeps an image
  from silently reporting your visit to a
  third party.
* The person who manages WednesdayAI can allow specific trusted websites, so images from those
  load as pictures too. If a picture you expect is showing as a link, ask them whether that site
  is on the allow list.

## Choose a theme

Select your name at the bottom of the sidebar to open the account menu, then pick a **Theme**:

| Theme | What you get |
| - | - |
| **Auto** | Follows your device: dark or light, or high contrast if you have asked your device for more contrast. |
| **Dark** | Always dark. |
| **Light** | Always light. |
| **High contrast** | Brighter text and stronger outlines, for low vision or bright rooms. |

The change applies immediately and is remembered on this browser. If you previously chose the older
**Black** or **White** theme, WednesdayAI carries that over to **Dark** or **Light**. The theme also
sets the colour of your browser's own scroll bars and form controls, so they match the rest of the
page. Animations are skipped if your device is set to reduce motion.

## Controls that are not available yet

Some controls are visible but switched off because the feature behind them is not ready. They look
dimmed. Hover over one (or focus it with the keyboard) to see what it is waiting for. Selecting one
does nothing and sends nothing to your assistant. (**Think** can also be dimmed, but for a
different reason: see [Choose a model and Think](#choose-a-model-and-think).)

| Where | Controls |
| - | - |
| Sidebar | **Add extension** |
| Account menu | **Settings**, **Keyboard shortcuts** |
| Under each reply | **Feedback**, **Regenerate** |

**Canvas** in the top bar is not on this list: it opens the Canvas panel. If it is dimmed instead,
the tooltip reads **Canvas — Assistant UI is off** - the person who manages WednesdayAI has not
turned assistant screens on yet; see
[Forms and other assistant screens](#forms-and-other-assistant-screens).

**Memory** and **Calendar** are not on this list either: they turn on when the person who manages
WednesdayAI allowlists a chat UI plugin that provides them - see
[Chat UI plugins](https://docs.wednesdayai.dev/next/developers/plugins/chat-ui). Until then they stay dimmed, and their tooltips read
"coming with the extension panel SDK" followed by a tracked issue number, just like the tooltips
for **Add extension**, **Settings**, **Keyboard shortcuts**, **Feedback** and **Regenerate**.

**Dictate** and **Voice mode**, the microphone and wave buttons beside the send arrow, are not on
this list — see [Dictate and voice](#dictate-and-voice). Either can still be dimmed when the
person who manages WednesdayAI has not turned on and configured that control's voice service yet;
the tooltip then reads **unavailable on this gateway**. Ask them to enable it if you want to use
it.

The **Share location** map pin beside the microphone is not on this list either: it works when the
page is on HTTPS (or `localhost`) and the browser offers location, and is dimmed with a tooltip
naming what is missing otherwise - see
[Share your location with messages](#share-your-location-with-messages).

**Read aloud** is not on this list either: the button works, or is dimmed with the reason beside
it - see [Read a reply aloud](#read-a-reply-aloud). The composer's **Read replies aloud** toggle
is live the same way, with the same dimmed reasons - see
[Read replies aloud automatically](#read-replies-aloud-automatically).

## Sign out

Select your name at the bottom of the sidebar to open the account menu, then select **Log out**.
The browser always discards its stored credential and key, so you (or whoever uses this browser
next) will need a new setup code to sign back in here, or a sign-in provider (Google, Microsoft
or GitHub) if your administrator enabled one. It also asks the gateway to forget this paired
device and unlink it from you; if the gateway was unreachable at that moment, ask your
administrator to remove the device (`wednesdayai devices remove <deviceId>`). It does not affect
any other browser or device paired with your assistant. Logging out also clears the drafts and
unsent messages this browser kept on the device.

The same menu has **Linked accounts**, which shows who this browser is signed in as and lets you
link your chat accounts (Telegram, Slack, Discord and friends) by entering a `/link` code - see
[Identities](https://docs.wednesdayai.dev/next/admin/gateway/identity).

Inside the WednesdayAI app the same menu has **Devices**: your glasses and this phone, each with
its status. Select one to open its settings in the app. A plain browser has no Devices section.

## Troubleshooting

### The bottom of the screen says "Offline" or "Reconnecting..."

The web client lost its connection and reconnects on its own - this is normal after a network
hitch, when your computer wakes from sleep, or when the assistant's server restarts. Returning to
the tab or coming back online retries immediately. Nothing is lost while this happens: a reply that
finished while you were away appears once, a reply cut short by a restart shows a notice (see the
next item), and an unsent message returns to the message box. A reply that stopped moving while
the connection silently broke picks up again on its own within about 15 seconds, or within a few
seconds of you returning to the tab - without reloading. If it still has not recovered after a
minute, reload the page.

### A notice says the reply was interrupted

If the assistant's server restarted while a reply was being written, the thread shows **The gateway
restarted before this reply finished. Your message was delivered; send it again to retry.** Your
message arrived - nothing was lost. The old reply does not continue on its own, so send the message
again if you still want an answer.

### Dictate or Voice mode will not start, or the button is dimmed

A dimmed microphone or wave button means the person who manages WednesdayAI has not turned that
control on yet. Ask them to enable it. If the button is available and then shows an error, allow
microphone access in the browser's site settings, reload the page, and select it again. If the
tooltip says this browser cannot capture microphone audio, use a current desktop browser over
HTTPS (or `localhost`).

### Read aloud will not start, or the button is dimmed

Hover over the dimmed speaker button to see the reason: wait for the connection to come back
(**unavailable while offline**), turn off Voice mode (**unavailable during Voice mode**), or ask
the person who manages WednesdayAI to turn the feature on (**unavailable on this gateway**). If
the button is available but a message appears beside it instead of audio, the message says what
went wrong - for example **Read aloud is turned off for this assistant.**, or
**Couldn't play the audio in this browser.**, which usually means the browser is old or missing
audio support: try a current desktop browser. See
[Read a reply aloud](#read-a-reply-aloud). The composer's **Read replies aloud** toggle follows
the same rules; see [Read replies aloud automatically](#read-replies-aloud-automatically).

### The location button will not turn on, or location is not shared

The dimmed map-pin button means this page cannot offer location at all: browsers only give
location to secure pages, so the address must be HTTPS (or `localhost`) - ask the person who
manages WednesdayAI for the HTTPS address. If the button is available but turning it on fails:

* **A message says location permission was denied**: allow location for this site in the
  browser's site settings, then select the map pin again. Sharing turns itself off when a send
  meets a denied permission, so it needs re-allowing once.
* **A message says your location could not be read**: the device could not get a fix in time
  (common indoors or with location services off). Turn location on in your device's settings or
  move somewhere with a clearer signal, then send again. The message itself was still sent.
* **The reply ignores your position**: your assistant only receives coordinates on the message
  that carries them, and is told to use them when the request depends on where you are - ask for
  directions, travel time or nearby places to see it used.

### The screen goes dark while I am using my voice

While Dictate, Voice mode, or a read-aloud is playing, the client asks the browser to keep the
screen awake. This is best effort: browsers can refuse it, for example while Low Power Mode or a
power-saving setting is on, and iPhone or iPad home-screen apps need iOS 18.4 or newer. Nothing
indicates a refused lock, because there is nothing the page can do about it; opening the page
fresh in the browser (rather than the home-screen app) or charging the device often helps.

### A dictated message sent without me pressing send

If the **Read replies aloud** toggle is on, this is expected: dictation ends itself after a quiet
moment and sends the words it heard, as long as the message box was free to send. While a reply is
being written, the words go to the current turn instead of waiting. How long that
quiet moment is - by default two and a half seconds - is set by the pause chip beside the speaker,
which offers **1.5 s**, **2.5 s** and **4 s**. Pick a longer pause there if messages kept going out
before you finished. If you prefer to edit before sending, turn the toggle off; dictation then
keeps its old behaviour, stopping when you select the microphone button again.

### Talking does not interrupt the reply

Interruption works only while the actual reply is speaking, not during the short "One moment" or
"Working on that" phrases, and not for a cough or a brief "mm-hmm". Wait until the reply itself
is playing, then speak a real sentence. If your first words after a reply are missed, pause
briefly or use a headset so speaker echo is not mistaken for you.

### Voice mode or Dictate stops by itself

Check the strip under the message box. **Reconnecting...** or **Offline** releases the microphone
and returns Voice mode to the wave button. That is expected. Select the button again after the
strip says **Connected**. If the strip stays **Connected** and Voice mode still ends with **Your
microphone stopped sending audio. Tap to start voice again.**, your browser stopped sending
microphone frames - the same situation Dictate describes below. Select the wave button again to
start a fresh conversation. If the conversation ends with no message, try once more and tell the
person who manages WednesdayAI the time it failed.

### Dictate says the microphone stopped sending audio, or transcription stopped

While recording, a red message below the message box reading **Your microphone stopped sending
audio. Tap Dictate to try again.** means your browser stopped sending microphone frames (for
example after a phone call or another app interrupted the microphone), and **Transcription stopped
unexpectedly. Tap Dictate to try again.** means the transcription service ended the take. Either
way the microphone button leaves its recording state within a few seconds and the words heard so
far stay in the message box. Select the microphone again to start a fresh take - reloading the
page is not needed. If the first message repeats, check that no other app is holding the
microphone, then try a headset.

### On iPhone or iPad, Voice mode starts but you hear no reply

On iPhone and iPad, Safari only lets a page start playing audio from inside a tap. Voice mode
connects to the voice service after you select the button, so on iOS the spoken reply can be
silenced even while the status shows **Speaking**. If that happens, stop Voice mode and use
**Dictate** or typed messages instead; typed chat is unaffected. This is a known gap on iOS.

### I cannot switch chats while Dictate or Voice mode is on

That is expected: while one of them is on, selecting another chat, the unread bell, or an
assistant only shows a notice - **Stop dictation to switch chats.** or **Turn off voice to switch
chats.** Select the pressed microphone or wave button once to stop it, then switch. See
[Switch between chats](#switch-between-chats).

### You are suddenly asked to pair again

Short interruptions - the assistant's server restarting, a flaky connection - no longer sign this
browser out; it reconnects on its own and you keep your chats. If you are genuinely asked to pair
again, this browser's access was removed or has expired. Ask the person who manages WednesdayAI for
a new setup code.

### A file will not attach

Check that the file is under 50 MB, is not empty, and that the message carries no more than
32 files. Any file type is accepted; there is no limit on the total size of a message's
attachments. A picture is shrunk automatically first - scaled to at most 2048 pixels on its longest
edge and re-encoded - so a large photo usually attaches after all; an image that still cannot
be made small enough (and any other file type over 50 MB) is rejected.

### The message box will not send, or shows an error

While a reply is being written, Enter adds your message to the current turn, or queues it, instead
of sending it straight away; see
[Keep chatting while a reply is streaming](#keep-chatting-while-a-reply-is-streaming). If the
thread shows **Could not add to the current turn:** or **Could not queue that message:**, your
text is back in the message box. The same happens when an ordinary send finally fails - for example after repeated
connection trouble: the thread shows the reason as a system message and your text returns to the
message box, ready to send again. Text the gateway never confirmed also comes back after the
fact: even a reload, a crash or the assistant's server restarting cannot lose it (a queued
**Send after reply** message is the exception) - reopening the
chat puts it back in the message box with a notice that reads **Restored a message that was not
sent.** (or **Restored N messages that were not sent.**). If the box says **Replying isn't
supported yet while a reply is streaming...**, remove the quote with its **x** or wait for the
reply to finish. If the bottom of
the screen says **Offline** or **Reconnecting...**, see the first item in this list.

### A notice says the chat changed elsewhere

Your message was still sent: the thread refreshed to show what this window had not seen, and the
message went out again automatically (see
[This chat changed elsewhere](#this-chat-changed-elsewhere)). The banner appears only when the
chat changes a second time during that retry. Your text is back in the message box; an attached
file or a quote is not, so add them again before sending.

### Suggest a title says "rate limit exceeded"

Suggesting titles is limited to 3 per minute. Wait for the number of seconds shown and try again,
or type a title yourself.

### I am not getting desktop notifications

Check four things, in order: **Notifications** is **On** in the account menu; the browser has
granted notification permission for this site - if the menu reads **Blocked in browser settings**,
re-allow notifications for this site in the browser's site settings; the tab is actually in the
background, because notifications are only shown while it is hidden; and the browser supports
them at all - most mobile browsers do not, and the menu reads
**Not supported in this browser**. The bell, the sidebar dots and the title count work in every
browser.

### A chat stays unread after I open it

Opening a chat clears its unread mark in the browser that opened it. If the dot is still there,
you probably opened the chat in a different browser or the installed app than the one showing the
dot: unread marks are kept per browser or installed app and are not shared between them. Reload
the page that shows the dot, and open the chat once there.

### A reply arrived but nothing shows as unread

The bell, the dots and the title count need to know when each reply was written. If the gateway
is older than the web client, replies can only be counted while the page is open, so a reply that
arrived while the page was closed or the connection was down is not marked. Ask the person who
manages WednesdayAI to update the gateway. If your browser clears site data between visits (for
example some private-browsing modes), unread marks kept for this site go with it too.

### A control does nothing when I select it

It is probably one of the [controls that are not available yet](#controls-that-are-not-available-yet).
Hover over it to see when it is planned. If it is **Think**, see
[Choose a model and Think](#choose-a-model-and-think).

### Reasoning does not appear after I turn on Show reasoning

**Show reasoning** can display only reasoning available in the reply; it cannot add reasoning to
an existing reply. **Send reasoning** controls delivery for new replies, separately from this
browser's display choice. Check that setting in **Think** before sending a new message. Some
models or replies may not produce reasoning even when that switch is on.

### Display, sorting, or assistant choices reset after a reload

The web client saves these choices in this browser. If it cannot read saved choices, it starts
with the defaults: reasoning hidden, tools shown in **Minimum** form, assistants from the last
**4 hours**, assistants sorted by **Recent**, sessions sorted by **Newest**, internal sessions
hidden, and no assistant filter. If it cannot save a change, your choice still works while the page is
open, but a reload may restore an older saved choice or a default. You can still navigate and
chat. Check your browser's privacy or storage settings if you want your choices remembered.

### The page shows a connection error and never loads chat

The address that hosts the web client may not be allowed to talk to your assistant yet, or the
address itself may be wrong. Contact the person who manages WednesdayAI and share the exact web
address you tried.

## Related

* [Web client hosting](https://docs.wednesdayai.dev/next/admin/web/user-client.md) - for the person setting this up
* [Pairing QR and setup codes](https://docs.wednesdayai.dev/next/users/cli/qr.md) - for the person who creates your setup code


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.