Agent chats

8 min read

An agent's Chat tab lets you keep several independent chats. Each session retains its own history and context, so you can handle different matters without mixing them.

Work with sessions

In the conversation sidebar, you can create a session, switch between sessions, search them, or view archived ones. To search within the open conversation, use the magnifying glass in the header or, in the Chat tab, press Ctrl + F (⌘ + F on Mac). The actions menu lets you change a session's name, icon, and color, archive it, or delete it. Ctrl + O (⌘ + O on Mac) opens a new session with the cursor in the chat box.

Option	When to use it
New session	Start a matter without carrying over another chat's context.
Fork	Try another path from a point in the conversation without losing the original.
Rename	Identify conversations that you will resume later.
Archive	Remove a completed thread without deleting its history.
Delete	Only when you do not need to keep the conversation.

Delete a session

Deleting a session removes its history. Archive it if you want to keep it for reference.

Fork from a message

Hover over any of the agent's replies and click the branch icon next to the time. A new session, "name · bifurcada", opens with the conversation up to that message, and the chat switches to it. The agent carries on there with the same context it had at that point: it remembers what was said up to that message and nothing that came after.

The original session does not change, so you can continue both along different paths. The copy keeps the original's model, effort, and context limit. Attachments are not copied: they stay in the original session.

You can only fork from one of the agent's replies, not from one of your own messages.

Sessions the agent manages

The agent itself can also organize its sessions while it works. It can create one to separate a matter, change its icon or model, and prepare the information needed to continue work in that thread.

When a session has no name, the agent gives it one within the first few messages. If it already has one, the agent doesn't change it on its own: it proposes the new name (and the icon, if it changes that too) as an approval in the chat itself. If you approve it, the session is renamed; if you reject it or don't respond within 24 hours, it keeps its current name. If you rename it yourself while the proposal is pending, your name wins.

This is useful when a conversation becomes an independent task: the agent can open a dedicated session with the relevant context already included, without mixing it with the original thread. You can then find it, rename it, archive it, or continue the conversation from the dashboard.

The agent can also consult context from other sessions or channels when it helps complete work. That access respects your permissions: it cannot retrieve information from conversations or channels you cannot access.

Context remains independent

A new session does not automatically share all another session's history. The agent provides the context it considers necessary when creating it; from that point, every session maintains its own thread.

All the agent's conversations

The session panel in Chat only shows your own conversations with the agent. The Conversations tab on its page gathers every conversation of the agent that you can see: yours, those from external channels, from event channels, and from scheduled tasks. Other members' personal conversations remain private. It is the same list as Conversations in the sidebar, but only with that agent's: you can filter them by channel, status, or model, change the model or effort of each one (or of several at once), and open any of them to read its history.

Settings for each conversation

A session can use a model different from the one configured for the agent and have its own context limit. Both changes apply only to that conversation.

• Change the model when the matter needs more capability or you want to control the cost of a specific task.
• Adjust the context limit when an extensive conversation should compact sooner or later. You change it from the session's chat or from the Context column of Conversations —the general page or the agent's tab—, which is the only way for conversations on event channels —a webhook, GitHub— that are not opened from the agent.

The picker offers the models of every connection in the organization, not only those of the agent's provider. If you choose one from another provider, the conversation uses that provider's default connection; connections that carry their own model, such as OpenRouter or Ollama, appear by name. The change applies from the next message and does not restart the agent.

Changing the model or the effort clears the cache

The history is kept, but the context cache is not: the conversation starts a new session from its latest messages, and each model also has its own cache. The next message is processed in full from scratch, so it takes longer and costs more. The same happens to conversations that use the agent's model or effort when you change them in its settings. The panel warns you before applying the change; if you tick Don't remind me again, it stops asking on that device until you turn it back on in My account › Profile › Shortcuts and warnings.

Choosing the connection of a conversation requires the View permission of Model providers, the same one needed to set the connection of an agent. If the agent switches model on its own, it stays within the provider of the connection you chose.

See Models to choose a provider and model, and Usage and limits to understand compaction and limits.

Messages, files, and voice

You can attach files to a message. The agent can review compatible images and documents, and receives other files in its workspace. When it finishes a task, it can return files through the same channel.

From the composer, you can record audio, and with Call, in the header of the agent's chat, start a voice call when the organization has configured this capability. The transcript is recorded in the thread; the agent adapts its response to voice and leaves links or technical content in writing in the chat.

During the call you can ask the agent to speak faster or slower, or to switch voices: the change applies from its next sentence. If you have permission to edit the agent, it stays as its usual voice for every conversation; if not, it only applies to that one. The agent’s usual speed is set in its configuration, next to the voice. When you say goodbye or ask it to hang up, the agent says goodbye and hangs up itself.

The agent can also call you when you ask it to, for example “call me when you finish”. It can only do so if you ask from your chat with it in the panel, and the call rings in the panel or the desktop app; if you answer, it tells you the result out loud. If the panel is not in front of you, you also get a notification on your phone or in the desktop app, with Answer and Decline: answering opens Kujira with the call picked up, and declining opens nothing. If you don't pick up, the notification turns into Missed call; if you answer or decline it on another device, it goes away. If you do not answer within 30 seconds, decline it or have no microphone available, it leaves the result in writing in the chat. To stop agents from calling you, turn off Allow calls from agents in My account → Notifications.

Configure voice and audio

In Settings > Voice, add the keys for the services the organization will use:

Service	Purpose
ElevenLabs	Voice replies and call transcription; audio notes too if you don’t set up Groq.
Groq (optional)	Transcribes audio notes. When set up, it is used instead of ElevenLabs for notes.

If you create the ElevenLabs key with restricted permissions, it needs Voices (read), Text to Speech and Speech to Text. Without the voices permission, the agent’s voice picker shows up empty. The Test button on each key checks that it works and, for ElevenLabs, tells you which permission is missing. It spends no credits.

For a voice call, ElevenLabs is enough. Enable calls in the agent configuration and select an ElevenLabs voice; the play button next to the picker lets you listen to each voice before saving it. If audio is not understood, first check transcription setup.

Debug with support

You can authorize Kujira support to review one specific internal chat from the session list. This authorization is limited to that thread, expires after 24 hours, and is not included in general temporary support access.

Keep one goal per session

Use one session for each customer, project, or problem. More focused history helps the agent stay on track and makes it easier to resume work later.