Setting up Janus Chat

Turning Chat on

Janus Chat runs only when the gateway starts with JANUS_CHAT_PREVIEW=true (false by default). Without it, Chat is hidden from everyone and Administration → Chat shows a banner saying so; settings you save there apply once it is on.

Administration → Chat → General → “Janus Chat is on” turns Chat off for everyone without a restart. Nothing is deleted: turning it back on restores every conversation. Administrators configure Chat but can never read anyone’s chats, memories or files.

General

Default model is what a person starts with until they pick their own; left on First available model, it is the first model in the picker. Instructions for every chat (up to 8,000 characters) are sent first in every conversation, before each person’s own instructions. Suggestions on the empty chat screen are up to 8 starter prompts of up to 200 characters each.

Under Features, Share links lets people share a read-only copy of a chat with others signed in to this instance; turning it off stops new links, and links already made keep working until their owner stops sharing or deletes the chat. Export lets people download chats as Markdown or JSON; printing stays available either way. Personal plans in Chat lets people chat with their own connected vendor plans (ChatGPT, Grok…); those are paid by their owner, show $0, and still count in usage and reports. Turned off, plan models leave the picker and are refused.

Keep temporary chats for sets how long an idle temporary chat is kept before it is deleted: 1 hour, 6 hours, 1 day (the default), 3 days, 7 days or 30 days.

Models

The Models tab lists every model granted to someone, and how it appears in the Chat picker. People only ever see the models they are granted. In Chat hides a model from the picker and Chat refuses it; it stays usable through the API. Category (General, Fast, Reasoning, Coding, Vision, Local & private, Creative) becomes the model’s group in the picker, Order (0–999) sorts it, lower first, and Recommended adds a badge.

Sees images, Hears sound and Sees video decide whether Chat sends the model the images, sound and video people attach. Automatic uses the model’s listed modalities, or guesses from its name. Set Yes for a self-hosted model served under your own name, for example a vision model on vLLM, so attachments reach it. A model set to No gets a note, a transcript or a description instead (see Tools).

The friendly chat name (up to 60 characters) and description (up to 200) are set per grant, with Chat name and description on a grant row in Administration → Gateway → Grants, so the same model can read differently for different teams. When several grants apply, the most specific one with a name or description wins; reports keep the catalog name.

Tools

Tools from the model’s provider (Web search, X search, Read web pages, Run code) are on by default. Web search, X search and Run code run on the provider’s own servers, and only for models whose provider offers them; any extra provider fee is included in each reply’s cost. Read web pages works with every chat model: Claude and Gemini use their provider’s reader, and every other model uses Janus’s own, which opens only public internet addresses, fetches up to 5 MB and reads the first 30,000 characters. Switching a tool off here turns it off for everyone; people can also turn any tool off for one chat.

Create images (Let chat models make images) lets any chat model make pictures. Image model for Create images is used when the chat is not on a personal plan with its own image model; left empty, it is the first image model the person may use, billed to their Working under.

Create files, Memory, Create videos and Create music are run by Janus for any chat model that can call tools. Video model and Music model choose the models for videos (xAI Grok Imagine, Google Veo, OpenAI Sora) and music (Google Lyria); left empty, the first one the person may use. Videos and music are their own metered requests, billed like images; files and memory cost nothing extra. The video and music settings need a video or music model that is enabled and granted.

Transcribe sound with picks the speech-to-text model that turns attached sound into a transcript for models that do not hear it. Describe videos with picks a model that watches video (set Sees video on the Models tab) to describe attached videos for models that do not watch them; without one, a video’s sound track is transcribed. Both are made once per file.

Web search for self-hosted models

Models whose provider runs no search (vLLM, llama.cpp, Ollama…) can search through a service you set up under Web search for other models. It is off by default; providers with their own search keep using it. Choose a search service: SearXNG (self-hosted and free; add json under search → formats in its settings.yml and enter its address), Brave Search, Exa or Tavily (paste an API key; it is stored encrypted and never shown again, and Remove key deletes it).

Results per search is 3, 5 (the default), 8 or 10; more results give the model more to work with but cost more tokens. Test search runs a sample search with the values on screen before you save. Then turn on Use this search for models without their own. The status beside the heading shows whether search is in use, and with which service.

To set search up with the deployment instead, run SearXNG next to Janus and pass its address in JANUS_CHAT_SEARCH_SEARXNG_URL (in the Helm chart, under extraEnv). At startup Janus turns search on with it once and shows it as Installed with Janus. It never replaces a search service you already use, and your later changes are kept across restarts.

Who can use Chat

On Janus Business, Who can use Chat limits Chat to Everyone or to Only these groups, teams and people. Message for people without access (up to 300 characters) is shown instead of Chat; without one, they are told to ask an administrator. Administrators can always use Chat. People outside the restriction can still open share links others send them (read-only). Someone who loses access can still list and revoke the links they made, but cannot make new ones.

On Community, everyone with an account can use Chat, and only the instance-wide on/off switch applies. A restriction saved under a Business license keeps applying after the license expires or the gateway is downgraded to Community; the Who can use Chat tab then shows a warning with Let everyone use Chat, which lifts it. Saving a new or changed restriction needs Business. Chat being turned off is always honored.

Search index, audit and environment

Chat search stores keyed hashes of words, never the words. After changing JANUS_ENCRYPTION_KEY, rebuild the index with POST /api/v1/admin/chat/reindex. Every change to these settings is recorded in the audit log as chat_settings_updated; search API keys are never written there.

JANUS_CHAT_PREVIEW (default false) must be true for Chat to run. JANUS_CHAT_AUTO_TITLE (default true) is the starting point for chat naming until an administrator chooses in Administration → Chat → Tools → Chat names; set false to start with naming off. JANUS_CHAT_MAX_UPLOAD_BYTES (default 20 MiB) caps one attachment, and never exceeds JANUS_STORAGE_MAX_OBJECT_BYTES. JANUS_CHAT_SEARCH_SEARXNG_URL is the address of a SearXNG installed with the deployment. Attachments and made files are kept in Administration → Settings → General → File storage.