Chatting.

The New Chat page

A new chat opens on a scene, a mosaic picture, with the message box in the middle and a greeting above it. There are 29 scenes in six groups: times of day, nature, animals, places, seasons and celebrations. On Auto the scene changes on its own now and then, always one that suits the time of day, with a season's scenes in their season and a celebration's (Christmas, the New Year, Lunar New Year, the harvest moon) around its day. To keep one you like, click the pill in the bottom left corner and choose it; Off leaves the page plain, with the message box still in the middle. Settings › Appearance › Scenes sets how often Auto changes and which scenes it leaves out. Start a new chat with Ctrl N or the button at the top of the sidebar.

The Mosaik theme (Settings › Appearance › Theme) takes the whole window's colours from the scene that is up, and is light or dark with its picture: by day the pictures are light, so the room is light. The sidebar, the header and the chat become one room over the scene, parted by thin lines: the scene reaches behind the sidebar on the New Chat page and stays barely there behind a chat. The message box and windows are glass. Each of those has its own switch, the theme button at the bottom right of the sidebar picks Light, Dark or Mosaik, and Light and Dark stay plain.

Until you have chats of your own, the page also offers to bring your chats from ChatGPT or Claude.

The message box

Enter sends, and Shift Enter starts a new line. In an empty box, ↑ brings back your last question to edit, and / lists your prompts and short commands under the box, narrowing as you type: /new, /temp, /rename, /model, /search off, /think high, /compare, /export md and more. A command typed out in full runs when you press Enter. While an answer is being written, Enter queues your next message.

The small arrow beside the send button sends one message another way, and the chat's own settings stay as they are: Send without searching, Send to another model, Compare models (below), or Send with options…, where you pick the model, search, reasoning and model settings for that message alone.

The row under the box reads from both ends. On the left are the things that shape your request before it goes: Search (see Web search), and a … menu with the prompt library, Compare, Context, Share and Export. A Custom tag appears there when the chat has settings of its own, which you change with the Chat settings button at the top of the window. On the right are the things that answer it: a token estimate once your message gets long (it turns red when the message is too long for the model), the reasoning level for models that think, and the model itself.

Whatever you type and don't send stays with its chat. Go to another chat, or anywhere else in the app, and it's waiting when you come back, even after a restart. The New Chat page keeps its own, and a pencil in the sidebar marks every chat with words still unsent.

A very long question folds in the conversation, with Show all to open it.

Reading while you type

At each pause in your typing, a fast model reads the message and plans the reply: the model, how hard it thinks, what to search, the styling and the length. A search the message clearly needs can start before you press Send, so the answer starts sooner. The words read this way go to the AI provider before Send. What the read decides, such as the model and the searches, is kept with the answer, and Mosaik never writes chat content to its log, typed or sent. Each read is a small AI call, added to the cost of the reply to that message.

Its settings are in Settings › Chats › Message box › Reading ahead: Read messages while you type, whether it reads in temporary chats too, what it does on Free, and whether a small mark shows in the message box while a draft is being read.

The coordinator

That fast model is the chat's coordinator. It picks the models that answer and sets how they work. Greetings, thanks, small talk and short questions with a settled answer it answers itself, in a few sentences, so the reply starts at once and costs very little. The answer's details then say The coordinator answered this itself.

Prompts

Save what you ask often in the prompt library (Prompt library in the message box's … menu or the palette), then type / in an empty message box. Your prompts come first. Pick one and it goes in the box for you to read over and send.

A prompt can have blanks for the part that changes each time. Write them in curly brackets:

Picking a prompt with blanks opens a small form, one field per blank, with the message shown as you fill it in. Insert puts it in the message box.

The library also keeps chat instructions, which you apply from a chat's settings. See The prompt library.

@ mentions

Type @ in the message box to bring something into one message:

• A model answers this message. The chat's own model answers the next one.
• Another chat goes along as context, and stays in this conversation from then on.

What a message mentioned shows under it. Click a mentioned chat to open it.

Voice

Read aloud is in every answer's ⋮ menu. It reads the words, skipping code and files, and a bar under the answer has Stop. Settings › Voice chooses your computer's voice, which is free and stays on your computer, or a speech model.

Settings › Voice picks the model that writes out your recordings (automatic unless you choose), the language, whether words appear as you talk or all at once when you stop (Write out as you speak), and how answers are read: your computer's own voice, free and offline, or a speech model on OpenRouter for a more natural voice, at a small cost per answer.

Attachments

Drop files anywhere on the window, paste them, or pick them with +. Each one shows as a chip in the message box, and you can send files with no message at all.

Almost anything attaches: pictures, PDFs, Word, Excel and PowerPoint files, OpenDocument files, RTF, EPUB books, e-mails, Jupyter notebooks, audio, video, ZIP and tar archives, and anything that holds text, like code, notes, CSV, logs and subtitles.

To mention a file in your message, drag it over the text. The message box lights up and a line follows your pointer to show where the file will go. Let go and it lands right there as a little box with its icon and name, between two words or at the end, so you can ask "does the lease agree with the budget?" and point at each one. A long name shows its first part and its extension. Backspace takes it out again.

How each file reaches the model

Models take in different things. Some see pictures, some read PDFs, a few hear audio or watch video. Mosaik checks what the model you picked can take, and sends each file the best way it can:

• As itself, when the model takes that kind of file.
• As its text, for documents, spreadsheets, slides and archives. They're read on your computer. A scanned PDF goes as pictures of its pages.
• Read by another model first, when the model can't take it. A quick, low-cost model that can see, hear or watch describes the picture, writes out the recording or describes the video, and its words go with your message, marked as its account of your file, so the answer talks about what your picture shows. It names what it recognises, like a painting, a landmark or a plant. That's done once per file and kept. When a question needs more, the answering model can ask the one that sees about the picture again, for example to read a signature, and you'll see each look in the answer's work. To name a picture, Mosaik searches by the picture itself.

Click a chip to see how it will be sent. A sparkle marks a file another model reads first. When there's more than one way, Send as lets you choose, for example a PDF's text or another model reading its pages.

If FFmpeg or LibreOffice is installed, Mosaik uses them for more: HEIC and AVIF photos, uncommon audio and video, frames from a video for a model that only sees, and old .doc, .xls and .ppt files.

Long pastes

Paste 4,000 characters or more and they're attached as Pasted text.txt instead of filling the message box. Its menu has Put back in the message if you wanted it inline. Settings › Answers › Files sets the length, or turns this off.

Files you've sent

Sent files sit above your question as thumbnails and cards. Click one to open it, save a copy, or see exactly what the model got. They stay with the chat: every later answer sees them, and so does a branch. Mosaik keeps its own copy, so moving or deleting the original doesn't break the chat.

Files the model makes

Ask for a file and you get one. A spreadsheet, a slide deck, a document, a picture or a whole folder of code shows up in the answer as a card with Open and Save.

No model writes a Word file directly. Each writes what it's good at, like Markdown, CSV or an SVG drawing, and Mosaik builds the real file on your computer. For pictures you can choose instead to have an image model draw them from the model's description, in Settings › Answers › Files. A picture that was only described has Draw it on its card either way.

Posters, cards and other pictures made of words are set from a page, so their words stay sharp type. They are edited in layers: ask to change what the photo shows and only the photo goes to the picture model, with the words and shapes set back over it where they were. A change to the words, shapes or layout rewrites the page and calls no picture model.

Analysis on your files

Attach a spreadsheet or a data file and ask about it: which region sold the most, what changed month to month, which rows are missing a value. Instead of reading numbers off the file's text, the model runs Python over the file itself and answers from what the code worked out. Ask for a chart or a cleaned copy and it comes back as a file in the answer.

Your files are only read. The Python runs in the same locked-down sandbox as the exact tools: it can't open, change or delete anything on your computer, and the file is handed to it by Mosaik. Click the work line above the answer to see the code it ran. It works with any model that can use tools, in plain chats.

The work line

While the model works, a single line above the answer says what it's doing: connecting, thinking, searching for something, reading a site. When the answer starts, the line sums up the work, for example Searched the web for 14s or Created a file, ran code and 2 more steps. Click it to see every step in order as a short line: Created guide.docx, Searched cone 6 glazes, code by what it does, and each stretch of the model's thinking by what it was doing, such as Planning the chapters. Open a step to see all of it: the results, the file, the code, every word of the thinking.

While a reply works, the line shows the kind of step it is on (reading your message, thinking, researching, outlining, writing, checking, editing or drafting) with a small mark of its own, a short label and what the step is about, such as Researching heat pump running costs. A part of a long answer written out of sight drafts a paragraph in its place, line by line. The marks stand still under Reduced motion.

The line under each answer opens its details. How it was made lists every AI call and tool step behind the answer with its model, time and cost, starting with the reads and searches made while you typed. Ask in the chat how an answer was made, or whether your message was read while you typed, and the AI answers from those same facts.

Edits while an answer is written

The AI can change words of an answer it is still writing. A part you haven't scrolled to yet changes without a trace. A part you have already read keeps its old words on the page, struck through, before the new ones. The answer ends with a count such as 2 edits: click it to list them, or choose Commit edits there or in the answer's ⋮ menu to keep the changes and remove their marks. Opening the chat again commits them too. The saved answer is always the clean text.

New text follows the bottom of the page only while you're at the bottom. Scroll up to re-read and the page stays put however much is written below. An arrow above the message box takes you back down.

Bigger tasks

Some messages are a job rather than a question: a book outline with a file for each chapter, a script that has to run, a pile of notes to sort. The model works through them step by step and writes as it goes. A sentence says what it's about to do, its steps fold into one line such as Created 2 files, and the next sentence follows. Click a line to see its steps.

A quick question stays quick. The model starts with a short note of what it can do, and reads the full instructions for a set of tools, like files or reminders, only when a task needs them. It reads them rather than guessing how a tool works.

Long jobs and Keep going

Once the model starts working with files or code, one message may take up to 100 steps. A step is one search, one file written or one piece of code run. When it hits that limit, it stops and says in its own words what's done and what's left. A line under the answer says it hit the limit and offers Keep going, which carries the work on in the same answer. Change the number in Settings › Chats › Working.

Chats that keep working

Open another chat while one is working and it carries on. A small star beside its name in the sidebar says it's still at it, and a dot says it finished while you were elsewhere. Open it to see the answer, or the work so far carrying on live. Up to three chats work at once; another waits its turn and says Waiting for another chat to finish. To stop a chat you're not in, choose Stop from its menu in the sidebar. Closing Mosaik while chats are working asks first.

Settings › Chats › Working sets the steps one message may take and how many chats work at once, and lists the reminders waiting, each with Cancel.

Sending while it answers

You don't have to wait for an answer to end. Type your next message and press Enter: it waits under the answer and is sent the moment the answer finishes. Send several and they go one at a time, each after the answer before it.

A waiting message has Send now, which stops the answer being written and sends it straight away, and buttons to put it back in the message box to edit or to take it back. Its … menu and a right-click have the same, and Copy. If you stop an answer, or it fails, waiting messages stay where they are until you send them. A waiting message belongs to its chat and waits there if you go somewhere else.

Answers that stop early

Every answer has a length limit, 16,384 tokens to start with (about 12,000 words, thinking included). When one reaches it, a line under the answer says This answer stopped at its length limit. Continue has the same model write the rest, which is added to the same answer. Raise limit beside it lets answers run twice as long from then on, and continues this one. The limit itself is Maximum answer length in Settings › Models › Model settings, or in the chat's own settings for one chat. A limit costs nothing by itself: an answer is paid for by what it writes.

When a long answer is ready

A long answer is a good moment to do something else. When one finishes while you're in another window, Mosaik's taskbar button flashes and Windows shows a notification with the chat's name. Click it to open the chat.

Settings › Chats › Notifications chooses how long an answer has to take first (10 seconds unless you change it), a notification or just the flash, a soft chime, and whether the notification shows the start of the answer. That's off unless you turn it on, since notifications stay on screen where anyone passing can read them. Try it shows one.

Maths

Formulas are drawn the way a textbook sets them, in a sentence or on a line of their own. Models write them in TeX, and Mosaik draws them everywhere an answer appears: the chat, quick-ask, Ask from anywhere, shared pages and PDFs. A dollar sign followed by a number stays a price.

Right-click a formula for Copy formula: As LaTeX, to paste into a LaTeX document or another chat, or As MathML, which Word and other editors turn into an equation.

Copying, editing and branching

Point at a message to see Copy, Edit or Regenerate. The newest answer keeps them showing. The ⋮ button and a right-click open a short menu with the rest, and its More holds what you need less often.

If a request fails, a card under your question says what went wrong in plain words, with what the provider said under Details. It offers Retry when the failure might go away, and a way round it when there is one: free models, more credit, another model or your key's settings. The card stays with the chat, so it's still there when you come back.

Versions

Nothing you regenerate, rewrite or edit is thrown away. The new answer, or the edited question, is a version beside the old one, and ‹ 2 / 3 › under the message flips between them. Each version keeps the conversation that followed it, so going back to an older answer brings back everything you said after it. The model only ever sees the version on screen.

Select part of an answer and choose Rewrite in the bar that appears to change only that passage. The result is a new version too.

Asking a named AI, or several

Name a model in your message, such as Ask Claude or have GPT write it, and that company's model answers, within your Cost limit.

Ask for the views of several AI companies, such as What do three different AIs think, and one model from each of two to four companies answers in the same reply. Each answer stands in a box of its own, headed by the company and the model, exactly as the model wrote it. The boxes appear as soon as the models are asked and fill as they write, all at once, each with Waiting, Writing or Done at its right. Once every answer is in, the coordinator writes where they agree, where they differ and a short summary under the boxes.

When the Cost limit allows none or fewer of the models asked for, a line above the reply says so, and the models within the limit answer.

Comparing models

Talk to two or three models at once and read their answers side by side, right in the chat. Choose Compare models in an answer's ⋮ menu, in the send button's arrow, in the message box's … menu, or tick models in the model browser's Compare. The models sit as chips at the top of the message box. Change one, add a third or take one away from its chip. Nothing is asked until you send.

Compare from an answer and that answer is the first model's, as it is. The others wait with an Ask button. Send a message and every model answers it, each in its own column under your question.

• Keep talking and each model keeps its own conversation: the next question goes to each of them with its own earlier answers.
• Columns hold still while they grow, so you can read while the models are still writing. Under each: how long it took, how many tokens, and what it cost.
• Keep this one carries the chat on with that model. Every other answer is kept as a version, one arrow away. Stop comparing keeps the first model's conversation.

Designs

Ask for something designed, like a website, an app screen, a poster, a menu or a small game, and the model makes it as a design. While it's being written, the answer shows a card that says how far along it is. Nothing is drawn half-finished. When it's done, it opens in a window beside the chat.

• Keep chatting to change it. Ask for a darker header or a pricing section and the model writes it again. The window moves to the new version once it's done.
• Every version is kept. ‹ 2 / 3 › in the window steps through them, and each one has its own card in the chat.
• Copy the design's HTML, save it as a web page, or make the window wider from the top.
• A design runs in a sealed frame, like the pieces in styled answers. It can't see Mosaik, your files or your key.
• Esc closes it. Click its card in the chat to open it again.

Designs are separate from styling. Answers still use colour and layout to read well, but anything you ask to have designed goes in this window, and you only get a file when you ask for one.

Open in editor

Answers stay in the chat. When you want to work on one, a long draft or a block of code, choose Open in editor in its ⋮ menu. If the answer has code, pick the whole answer or one of its blocks. It opens beside the chat, set the way the chat sets an answer.

• Read, Edit and Changes at the top. Read shows the draft formatted and code highlighted. Edit shows the text to type in. Changes marks what the last change added and took out.
• Select part of it, in Read or Edit, and ask for a change underneath. Only that part is rewritten. With nothing selected, the whole document changes.
• Every version is kept: the answer as it was, each change you asked for, and your own edits. ‹ 2 / 3 › steps through them, and you can carry on from any of them.
• The word count, or the line count for code, sits under the title. In code, Tab indents.
• Copy and Save as a file are at the top. Esc closes it, and opening the same answer again brings it back where you left it.

Changes are made by the chat's model and count toward your spending like any other request. They show as Editor on the Spending page.

Finding and organising chats

• Search every chat with Ctrl K. It looks in titles and in everything that was said, and shows the passage that matched.
• Find in this chat with Ctrl Shift F.
• Names. A chat names itself after its first answer. Rename it from its menu in the sidebar, or double-click the title at the top. Mosaik never renames a chat you named.
• The sidebar groups chats by Today, Yesterday and Earlier. Right-click a chat, or use its … menu, to rename or delete it. Drag the sidebar's edge to resize it, or fold it away with Ctrl B.
• Pins and folders. The same menu pins a chat to the top, and its Folder group files it in a folder of your own or takes it out. A folder's menu renames or deletes it. Deleting a folder keeps its chats.
• Several at once. Ctrl-click or Shift-click chats to pick them, then pin, file or delete them together from the bar at the bottom of the sidebar. Esc lets go.

Memory

Tell Mosaik something about yourself worth keeping, like your job, the tools you use or how you like answers, and it remembers it for every chat. Memory is kept on this computer. When a reply saves something, a line under it says what, with Undo.

Settings › Instructions › Memory lists everything it remembers. Edit or delete any of it, or add your own.

To make it forget something, delete it there, or just tell it in a chat.

Temporary chats

Click the ghost at the top of a new chat (or type /temp); it turns dark while it is on. A temporary chat doesn't appear in your chats or in search, doesn't use memory, and is deleted when you leave it. It has its own black and white mosaic, which stays faintly behind the conversation, so you always know where you are.

Click the ghost again to leave it. Want it after all? Click the bookmark beside it (Keep this chat). It moves into your chats like any other, gets a name, and is remembered from then on.

Long chats and context

Each model can only read so much at once. The ring at the bottom of the window shows how full that is. Click it to see where it goes: Mosaik's guidance, memory, tools, summaries, the conversation, and the room kept for the answer.

Show the next request lays it out exactly as it will be sent, with its size in tokens, message by message, each marked as sent, summarized or left out. Click a part to read it. Copy as JSON copies the request itself.

When a chat gets long, older messages are summarized rather than dropped, and the latest ones are kept word for word. A line in the chat marks where the summary starts; click it to read the summary, edit it, or undo it. To do it yourself, open the context detail and use Compact now, from Light (keeps the most) to Strongest (keeps only the last exchange). It shows how much room each setting frees before you choose.

A message's ⋮ menu, under More, has Always include, which keeps it in full however long the chat gets, and Leave out, which stops it being sent without deleting it. Settings › Chats › Long chats turns automatic summaries on or off, sets how full the window gets first, the usual strength, and which model writes them.

Exporting and sharing

Ctrl E saves the open chat as a Markdown file. Export in the message box's … menu also saves it as JSON or a web page. The files include the model's reasoning and the chat's own settings. Share › Copy whole conversation, in the same menu, puts the chat on the clipboard as Markdown.

Share as a page… and Save as PDF…, in the same … menu, save the chat as one file that looks the way it does in Mosaik: styled answers, charts, citations linked to their sources, and the pictures and files the answers made. Nothing is uploaded; send the file to whoever you like.

Choose whether the page is light, dark or follows the reader's setting, and whether reasoning, pictures, attached files and sources go in. Under Advanced, leave Mosaik's fonts out for a smaller file.

Importing from ChatGPT and Claude

Your chats from ChatGPT or Claude can come with you. First ask for your data:

Each emails you a link to a .zip file. Download it, then choose Import chats from ChatGPT or Claude in the palette, or Import chats… in the sidebar's folder menu, and pick the file.

The chats arrive in a folder of their own, From ChatGPT or From Claude, with their own titles and dates. Every edited question and every answer version comes too, as versions to flip through, and each chat opens on the one you last saw. A model's thinking comes as its reasoning, where the export kept it. Searches, tool use and hidden instructions are left out, and files and pictures are named where they were, since exports don't include them.

The file is read on your computer and nothing is sent anywhere. Import a newer export later and only the chats you don't have yet are added.

Quick-ask

For a side question that shouldn't derail the chat. Select some words in any message (prose, a list, a line of code) and a small bar appears next to them. Click Ask, or press Ctrl Shift K. The same bar has Quote, which puts the passage in your next message, Rewrite and Copy.

A small window opens by the passage. Your question goes to the chat's model with the passage and the recent conversation, so the answer has the context, but nothing is added to the chat. You can ask follow-ups in the window.

• Esc closes it and forgets the answers.
• Open as chat keeps the side conversation as a new chat.
• The main chat's model is told what you asked on the side (not the answers), so a later “like I asked a moment ago” still makes sense.
• Drag the window by its top to move it out of the way.

Ask from anywhere

Press Alt Space in any app and a small Mosaik window opens over it. If you had text selected, it's attached as a quote. The camera button adds a picture of the screen. Ask, and the answer appears right there; follow-up questions carry on the same conversation.

• Esc hides the window. Open it again within ten minutes and the same conversation is there.
• Every answer is kept as a chat. Open in Mosaik carries on with it in the full app.
• The model pill at the top picks another of your pinned models.
• Drag the window by its top.

Settings › Window and keyboard › Ask from anywhere changes the shortcut, whether the selected text and a screenshot come along, and the model that answers. To get the selection, Mosaik copies it for you and then puts your clipboard back as it was.

The shortcut works while Mosaik is running. Turn on Keep Mosaik running in the tray and closing the window leaves it in the tray, ready; quit from the tray icon's menu.