Janus Chat
Chat with every model you can use, in the browser, metered and protected like API traffic.
Starting a chat
Open Chat in the sidebar. If it is not there, your administrator has turned Chat off or limited it to certain groups, teams or people; ask them for access.
The model picker in the message box lists the models you can use while working under the current “Working under” selection, plus the chat models from your own connected plans under “Your subscriptions” (unless your administrator turned personal plans off for Chat). Your administrator decides how the picker is grouped and may give models friendly names, a short description and a Recommended badge. “Billed to” says who pays: the team or personal account you work under, or your own plan for a subscription model. New chats start on the model you last started a chat with, or on your organization’s default model if that one is no longer offered.
Press Enter to send and Shift+Enter for a new line. Replies keep going if you close the tab; reopen the chat to see them. Esc stops a reply, Ctrl/Cmd+Shift+O starts a new chat, and / jumps to the message box. An unsent message is kept per chat while the tab is open.
While a reply streams, the thread follows it until you scroll up to read; Jump to latest follows it again. Edit a message you sent, or regenerate a reply, to make a new version; the arrows under it switch between versions.
After three replies, or once a chat has been quiet for 10 minutes, Janus gives it a short name in the background, once. By default the model that wrote the last reply names it, so the conversation goes nowhere it has not already been. An administrator can choose one model to name every chat, or turn naming off, in Administration → Chat → Tools → Chat names. A chosen model gets only what the person typed (the first and latest messages, trimmed), never replies, files or instructions; chats the Security Gateway had any finding on are not sent to it. Naming is a system task, so people don’t need access to the naming model. Naming asks the model not to reason (a title needs none). Chats you renamed and temporary chats are never named. Naming is a separate request, metered like the chat and shown in the request log as janus-chat-title; when it follows a reply its cost is added to that reply’s footer, and it does not count as a request toward rate limits or request quotas.
Your models
Tools in the message box opens Tools & models. Under Your models you choose your own Chat, Images, Video and Music models (the last three appear when you have such models). They are your defaults in every chat for as long as they stay available to you. Automatic uses your plan’s model when you chat on a plan, otherwise the organization’s.
Files, sound and video
Attach files with the paperclip, by dropping them on the message box, or by pasting an image. You can attach up to 10 files per message, each up to 20 MiB unless your administrator changed the limit. Files are checked by their contents: programs and archives are refused even when renamed, and so are images over 40 megapixels. Text files must be UTF-8. Files you attach but never send are removed after a day.
Documents (PDF, Word and text files) are converted to text and sent in full with every message of the chat, so the box shows about how many tokens they add before you send.
Images (PNG, JPEG, GIF, WebP) go to models that see images. Other models are told an image was left out, and the chat says so.
You can also attach sound (MP3, WAV, M4A, OGG, FLAC, AAC, Opus) and video (MP4, MOV, WebM) and play them in the thread. A model that hears sound or watches video gets the file itself. Any other model gets a transcript of the sound, or a description of the video written by a model that watches video. This text is made once per file, and the reply shows Transcribed or Watched the video. Your administrator decides which models see images, hear sound and see video.
Instructions and thinking
The settings button in the chat header holds instructions for this chat, your own instructions for every chat (your name, tone, language, format), a Thinking level and your Memory. Your organization’s instructions, set by an administrator, come first in every request, then yours, then this chat’s. The model sees them; the thread does not show them. They are stored encrypted.
Thinking (model default, low, medium or high) is sent as reasoning_effort. Janus fits it to what each model takes, and the note under a reply says when it was adjusted. Models that do not reason ignore it.
Tools
Tools for this chat, in Tools & models, lists what this chat’s model may use, with a switch for each: Web search, X search, Read web pages, Run code, Create images, Create files, Create videos, Create music and Memory. A model is offered only the tools its provider or Janus can actually run, and it decides when to use them. Every tool is on unless you switch it off for this chat or your administrator turned it off for everyone. If the model cannot use tools, the tray says so. If Child Safety is on for your account, some tools may be missing; a note under Tools for this chat says which.
Web search and Run code run on the provider’s own servers for OpenAI, xAI, Claude and Gemini models; code never runs on Janus. X search searches posts on X and is offered to Grok models. Any extra fee the provider charges for a tool is included in the reply’s cost. Self-hosted and other models without a search of their own can search the web when your administrator has set up a search service. The reply shows what was searched and lists the sources it cited.
Read web pages works with every chat model: give it a link and it reads the page. Claude and Gemini use their provider’s reader. Every other model uses Janus’s own reader, which opens only pages on the public internet and reads the first 30,000 characters. The page is listed under the reply.
Pages, search results and the text made from an attached sound or video file are checked when Chat reads them, against your Security Gateway policy. One that contains instructions aimed at the assistant is withheld from the model, the reply says “Withheld something it read”, and the reply goes on.
Making images, files, videos and music
Ask a chat model for a picture and it makes one in the same reply, up to four per reply. It uses your own image model if you chose one under Your models, otherwise your plan’s image model on a subscription chat, otherwise the organization’s.
You can also pick a model from “Make images” in the model picker (it lists the image models you are granted under the current “Working under”). The message box then asks for a description; send it and the reply is the image. Generation takes from a few seconds to about a minute. Under the image are its size, a download link, the description the provider actually used, and what it cost.
Images are kept in Janus, encrypted, so they stay after the provider’s own link expires; only you can open them, and deleting the chat deletes them (for good 30 days later, like the chat). You can switch back to a chat model in the same chat to talk about the image; the chat model is told an image was made, but does not see it. Image models make new images from a description; editing an attached image is not supported yet.
When a model needs a detail to answer well (which environment, how often, which of two approaches), it can stop and ask you. The question appears as a card under its reply with a few choices, the recommended one first, and sometimes room to type your own answer. Pick and press Send answer, or just type a reply. Turn Ask questions off in the Tools menu for a chat, or for everyone under Administration → Chat → Tools. Not used when comparing models or in deep research.
A model can hand you whole files, such as code, web pages, games, documents or data, up to 10 per reply and 2 MiB each. A complete HTML or SVG page in a reply becomes a file card too. A file card has Preview (in a sandbox with no access to Janus), Download and Copy, and code blocks have Download.
Files stay part of the chat. Later in the conversation, any model you switch to can see the files made earlier: small text files are included in full, and it opens larger ones, and files you attached, by name when it needs them (shown as “Read a file”). So you can ask one model to check or change files another one made.
With Create videos and Create music, a model makes short videos or music clips with the organization’s video or music model, or your own under Your models, at most two per reply. They play in the thread and can be downloaded. Each one is its own metered request, billed like an image, and a video can take a few minutes.
Memory
When you ask the assistant to remember something, it saves it to your memory and reads it in later chats; ask it to forget something and it removes it. A Memory updated note under the reply lists what changed, with a Manage memory button.
Memory in the settings drawer lists everything saved. You can add, edit or delete entries, delete them all, or switch off Remember things across chats. Memories are encrypted and only you can see them; administrators cannot read them. Temporary chats never use or update memory, and an administrator can turn Memory off for everyone.
Deep research
Deep research, under This message in Tools & models, has the model run many searches (up to 20), read the most relevant pages and write a cited report. It takes a few minutes. The report is also saved on the reply as a Markdown file you can download.
It applies to the next message only and shows as a chip above the message box, which you can remove before sending. It needs a chat model and cannot be combined with Compare.
Comparing models
Compare, under This message in Tools & models, sends your message to two or three chat models at once and shows their answers side by side, each with its output tokens, cost and time. Each model is billed for its own reply, and Stop stops all of them. Image models cannot be compared.
The first model you pick is marked Preferred, and the chat goes on from its answer. Choose Prefer this answer under another reply to continue from that one, with its model, instead; the other answers stay as versions you can switch to. Compare stays on until you remove its chip above the message box.
History, pinning and temporary chats
Chats are listed by date with pinned chats at the top. Each row has a menu to rename, pin or delete; double-click a title to rename it. When you delete a chat it disappears at once; Janus removes it and its files for good 30 days later. Search finds chats by whole words in titles, messages and attached documents.
A Temporary chat (the toggle in the header of a new chat) is never listed or searchable, and is removed once it has been idle for as long as your administrator set (a day by default; hover the toggle to see the current setting). It never uses or updates memory, and starting one does not change the model new chats start on. It still goes through the gateway and is metered like any other request.
Saved prompts
Type / in the message box to pick a saved prompt; Enter or Tab inserts it, arrows move between them. A prompt can ask for values with {{name}}: you fill them in before it is inserted, and blank ones stay in the text to fill by hand. Manage your prompts from the prompts button in the chat header, or turn a message you sent into one with “Save as prompt”. Prompts are yours only and stored encrypted.
Cost before you send
While you type, the message box shows about how many input tokens the next message sends and, for priced models, what that input costs. Every message re-sends the conversation, which is why long chats and large files cost more per message. The reply is billed on top at the model’s output rate; models from your own plans show tokens only. If a quota you are under is close to its limit or used up, a line above the box says so and when it resets.
The tokens and cost under a reply cover the whole reply: every tool and research round, images, video, music, transcripts and descriptions it made, and the chat’s title. The chat header shows what this chat has cost so far — every reply, including other branches and Compare answers — and how much of the model’s context window it fills. When the conversation no longer fits, the oldest messages are left out of the request, and the thread says how many earlier messages were not sent.
What the notes under a reply mean
Each reply shows the model that answered, tokens in and out, cost, speed and a link to its row in the request log. A short note appears when something worth knowing happened: an auto-routing router or alias answered with a different model (and why), a fallback model answered, Janus adjusted the thinking level, a security policy flagged, redacted or stopped the reply, or Child Safety’s tutor held back the answer. A reply a policy declines is shown as a calm notice, not an error.
In the request log, a reply’s first call has the client janus-chat; its extra calls are janus-chat-tool (tool rounds), janus-chat-research (deep-research rounds), janus-chat-media (images made by a tool, video, music, transcripts and descriptions) and janus-chat-title. All calls of one Chat conversation carry the conversation as X-Janus-Session, so auto-routing keeps the conversation on one route.
Tool use is shown with the reply: what was searched and the sources, pages read, code that ran, images, files, videos and music made, sound transcribed or video watched, and memory changes.
Administrators cannot read chats. Messages, instructions and files are encrypted at rest. Administrators set up Chat in Administration → Chat; see Setting up Janus Chat.