Three ways to interact
Use text for a conversation, a Task for durable work and a live call for voice. Voice notes and dictation are separate inputs. Each path retains its own state while remaining connected to the same agent.
Move from a message to work
The controller can answer directly, use permitted conversational tools or admit a Task or Process. It builds a self-contained request containing useful constraints, files and references. The agent’s Task harness is resolved only for the Task.
Keep the channel of origin
Native chat and external messaging bridges use a common conversation model. The source channel and receiving connection remain identifiable, so files and replies can follow the appropriate transport and permissions.
Two voice paths
A transcription–agent–speech pipeline uses configured STT, conversation and TTS resources. Native speech-to-speech uses a compatible realtime provider. Calls can use available context and tools; their functionality depends on the selected engine and transport.
Find a conversation or the right colleague
Chat home shows available agents and recent conversations, with unread indicators, the last message and previews according to your privacy preference. Lists are paginated with 50 items by default. Conversation names already in use, including archived conversations, receive a suffix to keep exchanges distinct.
An agent can consult the directory and authorised colleague profiles directly in conversation with agent_list and agent_get. This helps identify a specialist before assigning work. Reading the directory contacts no one and creates no Task.
Organise everyday conversations
The sidebar brings together Conversations, Documents, Tasks and Processes. Sections adapt to their contents while keeping your manual collapsed state; creation actions remain accessible according to your permissions. Show conversation filters when needed: hiding them preserves their values.
The Tasks panel combines work linked to the exchange with the agent’s running, queued or automatically waiting tasks, including work started elsewhere. It is useful even in an empty conversation; completed outside work and manually paused outside tasks are not added. The open document is highlighted in its list.
Name, search, archive or mute conversations. Reply to a specific message, dictate text, attach files and find linked Tasks, documents and Processes in chat panels. Unread state is stored on the server; configured Web Push can notify you when browser permission is granted.
Move from voice to lasting work
A call can start a Task while the exchange continues. Monitoring retains channel, duration, turns, interruptions and errors. Available transcripts and time to first text or audio help you assess the experience.
Native chat uses WebRTC; Matrix and Nextcloud Talk have their voice transports. Configure STT/TTS or speech-to-speech resources and network traversal. Revoking call permission stops audio.
Revise the document in front of you
On desktop, select a passage and write “rephrase this passage”. Your last selection, cursor position and an excerpt of the visible text accompany the message, even after you click into chat. This context covers editable text, source and Dataset JSON; embedded applications are not inspected. Excerpts are bounded and the agent must check the current revision before editing.
In chat, open a document in a desktop pane or a mobile dialog. The editor retains its permissions, versions and drafts. Sending a message waits for saving and includes the reference of the document actually displayed; closing it removes that context from subsequent messages.
The document_show function lets an agent request opening only in the current internal text chat. It does not run in Tasks, voice or external channels. It grants no access and is not proof that the recipient read the document.
Document continuity can also recall a reference from an older exchange beyond ordinary recent history. Recall checks current access, title and revision; it is separate from the open document’s context. Return to a document discussed earlier.
Keep the request and its useful media
When a conversation becomes a mission, Galaris keeps the original request and its attachments separate from the model’s supplementary context. A new independent request leaves previous work running or waiting; it does not silently replace it.
In chat using the internal engine, a compatible model can also receive attached images, audio, video and PDFs directly, subject to model and transport capabilities. Media remain tied to their original message. See media input conditions. An open document may contain HTML text, an application or a JSON Dataset; its editor follows the rules of its type.
The document remains usable in the chat panel: complete a form, adjust a calculator or manipulate a 3D scene while talking with your agent. Its content and embedded resources must respect the isolation constraints.
A correction that reaches the document
“Rephrase this passage” or “add the conclusions to the report”: in text chat, the displayed document gives the request a shared working context. The agent can integrate the correction into the existing document after checking its content and revision, then present it for review. A reply in the thread suits short exchanges. Work on the document.





