Using the app

Chat

A conversation with the model running on your computer. When you grant folders or email, the same chat becomes an agent: it understands by itself when it has to work on your files or your email.

WritingImagesAttached documentsReasoningContext and fidelityRepliesHistoryLong conversationsMemoryRolesAgentAccessMissing accessConfirmations and log

Writing and sending

  • Enter sends, Shift+Enter starts a new line.
  • Stop halts the reply: the part already written stays, and the model remembers it afterwards. It also stops a tool at work right away, such as a search through files: while it works, the tool's step shows two animated lines.
  • Under each reply: number of tokens, speed in tokens per second and duration, plus Copy to copy the text.
  • At the top, next to the model name, the engine state and a reminder that everything stays local.

Images

With a model that sees images, you can show the chat a photo, a screenshot or a scanned page. These models are marked Sees images on the Models page.

  • To attach: the paperclip in the message box, dragging into the chat, or Ctrl+V to paste one (a screenshot, for example).
  • Attached images appear as thumbnails above the message box; × removes one. You can send them with or without text.
  • Large images are reduced to at most 1600 pixels per side; formats the engine can't read, such as WebP, are converted.
  • Images are saved with the conversation and shown in your message. They stay in the conversation, so you can ask follow-up questions about them.

With a model that doesn't see images, attaching an image shows a message suggesting a model marked Sees images; documents instead can be attached with any model (see below). If you reopen an old conversation with images using such a model, the model gets a placeholder instead of each image.

Like the rest of the chat, images stay on your computer.

Attached documents

You can attach documents too, with any model: their text reaches the model together with your question ("summarize these minutes", "how much did I spend in total?", "make an anonymized copy of this Excel file").

  • They are attached like images: paperclip, dragging or Ctrl+V. Each document appears above the message box with its name; × removes it.
  • Formats: text files (txt, md, csv, code…), PDF, Word (.docx and old .doc), Excel (.xlsx and old .xls, one sheet at a time, with numbered rows) and PowerPoint (.pptx and old .ppt, one slide at a time, with notes). Binary files are reported and skipped.
  • When you send, if the file is not in a folder already granted (to the chat or the agent), the confirmation card appears: Only this file reads it once for this message, The whole folder adds the folder to Access (so the model can also work there, for example save a file next to it), Cancel doesn't attach it. Reads and refusals go to the log.
  • A long file arrives in part: about a quarter of the model's context, shared among the attached documents. The model knows the file continues and, if the folder is granted, can read the rest by itself.
  • The document stays in your message and in the conversation: when you reopen it, the model finds the text it had read, even if the file has changed or moved in the meantime.

For the model to create or edit documents it needs granted folders: see What an agent can do (the chat works the same way).

Reasoning

Many models reason before answering. While they do, the chat shows Thinking…; then the reasoning stays available in the Reasoning section, collapsed to keep things tidy. Whether to turn it on and how long to let it run is decided in Settings: on gives more accurate answers to problems, calculations and code; off gives immediate answers.

Context and fidelity from the chat

The button at the top right of the chat, next to Access, shows what the engine is using, for example Context 32K · Balanced. Click it to change:

  • Context length: how much text the model keeps in mind. With an expert map, context takes VRAM that would otherwise go to the experts: on Qwen3-235B with an RTX 5080, going from 32K to 16K raises generation from about 4.3 to 5.1 tokens/s.
  • Automatic context priority (only with Automatic): Speed, Balance or Long context, explained in Settings.
  • Selective fidelity (only when the model has a map): Balanced is recommended; Fast is about 15% faster but loses some quality, more on topics far from the map. Details in Selective fidelity.

They are the same settings as in Settings and apply after reloading the engine: the panel shows the Reload the engine button and how long the last load of this model took. If the conversation is longer than the new context, the panel warns you: after reloading, its oldest part will be summarized (see Long conversations).

What replies can show

  • Headings, bold and italics, bulleted and numbered lists, quotes and tables.
  • Code blocks with a Copy button.
  • Simple math formulas, shown as readable text.
  • Text you can select by dragging, across paragraphs, lists, code and even several messages (list bullets are copied too): it goes straight to the clipboard and Copied to clipboard appears (you can turn this off in Settings). A double click selects a word; Ctrl+C and the right mouse button copy too.

Stopped replies

MessageWhat happened
Stopped: the model was repeating…The app notices when the model repeats the same text in a loop and stops it. Rephrase the question.
Stopped: the reply length limit…The reply reached the maximum length per reply.
The engine did not respondThe engine stopped. If it happens again, reload the model from the Models page.
The request does not fit in the model's context…Even after compacting the conversation, the request is too big (for example one huge message). Start a new chat or increase the context in Settings.

History

  • The clock button opens the panel on the right with the conversations, newest first, and a box to search them by title.
  • Reopening a conversation lets you continue it: the model finds the earlier questions and answers.
  • The bin next to each conversation deletes it.
  • Conversations are saved only on this computer.

Long conversations

A model can keep only so much text in mind: its context. When a conversation gets close to the limit (about 70%), the app makes room in two steps:

  • First it removes the results of older tool uses (files read, searches), keeping the latest three. The model can read them again if needed. It's instant and loses nothing of what you and the model wrote.
  • Only if that isn't enough it has the model itself summarize the oldest part, and keeps your current request and the latest exchanges as they are. Then a note appears in the conversation: Conversation compacted: the oldest messages were summarized to stay within the model's context.
  • The history keeps all the messages: only what the model receives is summarized.
  • Details from the summarized part can be less precise than in the original messages. If one matters, repeat it.
  • If a request still doesn't fit, for example one huge message, the reply tells you to start a new chat or increase the context in Settings.
  • Long files are read in parts, so a long PDF doesn't fill the whole context.

Compacting on request

Type /compact in the message box to summarize the conversation right away, for example before starting a new long task. You can add what to keep: /compact keep the figures of the Rossi quote. The command isn't sent to the model as a message.

  • While you type /, a hint appears under the box: /compact [what to keep]: summarizes the conversation now to free up context.
  • If there is nothing new to summarize, a note says: Nothing to compact: the conversation is already summarized or empty.

Agents work the same way, /compact included.

Memory across conversations

If you turn it on in Settings › Memory, the chat remembers short notes about you and can search your past conversations when you mention them. It is off by default.

The chat that works: roles

When you have granted folders or linked email (see Access), for every message the chat works out whether it needs a role: a set of rules and tools for a certain kind of work. For an ordinary question it activates none and answers as usual.

RoleWhat forWhat it needs
FilesFinding, opening, reading, summarizing and comparing your documents; answering about them; writing new files.A granted folder
Tidy upPutting a folder in order: it proposes a structure, waits for your yes, then creates folders and moves files. It deletes nothing.A granted folder
CodeWorking on a software project: reading and changing code and, if you allow it, running builds and tests.A granted folder (and commands, if you want)
EmailChecking what arrived, searching, reading, replying, writing drafts, sending, saving attachments.A linked email account
ScreenLook at an open window or the screen: read an error message, explain what a program shows. Looking only, no clicks.The "Look at the screen" access; a model that sees images (on Windows the text read with OCR is also enough)
IntegrationUse the tools of an external service or program, for example GitHub. Each ticked integration is a role of its own, with its name, and is activated only when the request is about that service.The integration ticked in Access: see Integrations
  • Automatic choice. In the message box, Role: automatic means the chat chooses by itself. A request can activate two roles together, for example reading a file and sending it by email.
  • The role stays while the task goes on. If tidying up proposes a plan and you answer "ok, go ahead", the role is still there; it drops when you change subject.
  • You can choose it yourself. Remove a role with × or add one with + role: the next message uses your choice.
  • Under each reply you see the roles used, and the tool steps (for example Reads, Searches the mail) can be opened to see their result.
Roles grant no permissions

A role only uses what you granted in Access. Without a granted folder the Files role does not exist; commands work with every file role (Files, Tidy up, Code) once the commands switch is on.

Agent

The chat with the Photo tidy-up agent chosen in the selector: the Tidy up role proposes how to arrange the photos and waits for the go-ahead.
The chat with the Photo tidy-up agent chosen in the selector: the Tidy up role proposes how to arrange the photos and waits for the go-ahead.

The Agent selector at the top tells who you are talking to. General is the chat with its own access; an agent brings its own instructions, folders, email and integrations. Changing it in a chat that has already started opens a new chat, after a confirmation; the previous one stays in the history, with the agent's name under the title.

Access

The Access panel: granted folders, commands and email accounts.
The Access panel: granted folders, commands and email accounts.

The Access button at the top opens the only place where you grant something to the chat. Changes apply from the next message.

  • Granted folders: the chat can search, read and write only inside them (subfolders included). The first one is the starting point.
  • Can run commands: for every file role. Builds and tests, but also unlocking a file kept read-only by version control (for example p4 edit with Perforce). On Windows commands run in PowerShell (or cmd, if the model asks for it). Each command is shown to you and asks for confirmation, unless you allowed it forever.
  • Email: tick the accounts the chat may use, or add a new one. See Email.
  • Integrations: tick the integrations the chat can use; the Set them up on the Integrations page link leads to the page where you add them. If you tick a remote one, the header line "· local, nothing leaves your PC" becomes "· also uses online services:" with their names. See Integrations.
  • Permissions: the actions you allowed forever, each with Revoke.
  • Look at the screen (Windows, and Linux with X11; Mac is coming): the model can take a screenshot of a window or of the screen. You see every screenshot in the confirmation card exactly as it will reach the model, and it goes only if you allow it: there is no "always allow" here. With a model that does not see images, on Windows it gets the text read in the screenshot (OCR): the confirmation card says so and the model reminds you in its reply, because layout, colors and charts escape it. The screenshot stays in the tool step, as a thumbnail: one click opens it full size.

When access is missing

A question about email with no linked account: the chat offers to link it, with the instructions for Gmail.
A question about email with no linked account: the chat offers to link it, with the instructions for Gmail.

If you ask for something that needs files or email you haven't granted yet, the chat doesn't answer "I can't": it shows a card to set them up right there, in the conversation.

  • Email: tick an account you already have, or type your address. The app recognizes the service, sets the servers and tells you what is needed (for Gmail, the app password, with a link to the right page).
  • Files: Add folder… and choose the folder to work on.
  • Screen: turn on Look at the screen in the card. If the loaded model does not see images, the card says so: you need a model labeled Sees images.
  • As soon as the access is there, your request starts again by itself. With Answer without it the model answers with what it has.

The password is typed only in its protected field: it never goes through the chat or the model.

Confirmations and log

The chat asks for permission before deleting a file.
The chat asks for permission before deleting a file.

Risky actions stop on a confirmation card with three choices: Allow, Always allow (for that kind of action) and Don't allow. If you refuse, the model doesn't insist and asks what you would rather do. Which actions ask for confirmation is explained on the agents page: the chat follows the same rules.

The Log tab next to the history: every read, write or attempt, with its outcome.
The Log tab next to the history: every read, write or attempt, with its outcome.

In the panel on the right, the Log tab lists everything the chat read, wrote or ran, including the actions you denied. Files that were created, changed, moved or deleted can be restored with Undo.

Small models

Smaller models use tools less reliably: sometimes they don't open the right file or stop before finishing. For work on files and email, a model marked Good for agents is the better choice.

HotMoE is a Virsion project. This guide describes the app as it is today: what is still on the way is marked as such.