Chat

Orbi and agent chats

Orbi is the general assistant: a conversation with no agent attached. Its composer reads Message Orbi..., and its mascot pulses while it is thinking. You can also chat with one of your own agents. In an agent chat, the agent's avatar replaces the model picker, because the agent brings its own model, tools, and skills.

On the home screen, a Chat / Studio switch in the composer moves between chatting and generating media. See Studio. Under an empty composer, suggestions (such as Keep track of my follow-ups) fill in a detailed request without sending it. None of them needs a connected account.

What you can ask

In a single thread, you can ask Orbi to:

• Build a workflow, an agent, a table, or an app from a description of the job.
• Change a step: swap an integration, adjust a condition, shorten a prompt.
• Run the workflow and show you the results.
• Inspect a past run to explain what happened or why it failed.
• Connect an integration or wrap the workflow in an app.

A workflow always starts with a trigger. If you describe the steps first, Orbi adds a trigger before them. It tells you what it just did and what it suggests next.

Running and debugging from chat

Ask Orbi to run the workflow and the run trace streams into the thread, step by step, with each node's output. If something fails, ask “why did that fail?”. Orbi reopens the run and reads it node by node, so it can tell you which step broke and why (a bad credential, a missing field, an empty result) and offer a fix.

You:   Run it with a test contact.
Orbi:  The run stopped at "Send welcome email". The first three steps
       succeeded, but the mail server rejected the credential.
       Want me to point the step at a different one?

Two kinds of steps do the work. Plain steps (an integration call, an HTTP request, a data transform, a bit of code) do exactly one predictable thing. Agents reason: they decide, summarize, classify, or handle open-ended input. Use an agent when the task needs judgement, and a plain step when the operation is fixed. Plain steps are faster, cheaper, and easier to audit.

The composer

The message box grows as you type. Enter sends, and Shift+Enter or Ctrl+Enter adds a new line. A screenshot pasted in the box attaches to the message, as a file dropped on it does. While Orbi is answering, the send button turns into Stop, and pressing Escape also stops the answer (unless Escape is closing a menu or dialog first).

Stop also stops the workflows and applications launched from this conversation in the last 24 hours that are still running. Three kinds of run are left as they are: one waiting for a person (an approval, an interface), one also running something it was not launched for (a test from the builder, a webhook or schedule on production), and one you could not stop from its own page.

As Orbi works, its progress renders live in the thread:

• A thinking indicator shows while it prepares its next move.
• Each tool call renders as a tool card that updates as the call completes. Back-to-back calls to the same tool share one card. Only the latest steps stay expanded; older ones fold behind a toggle.
• Results that are more than text render as previews: tables, interfaces, applications, generated images and files, workflow diagrams, and run traces. Several previews in a row are grouped into a carousel. Click a workflow diagram to open it in the side panel.

On a narrow screen or a narrow side panel, the attach and tools buttons merge into a single More actions (+) menu.

Model picker and reasoning effort

The model picker sits in the composer and keeps to the essentials: each model shows its provider icon, a profile (Most capable, High performance, Balanced or Affordable), whose hover says when to use it and when to switch to another profile, and, on the cloud, the estimated credits per exchange, shown in the open menu only. The i button opens the full picture: when to use the model, what it can and cannot do, how much it takes in at once, and its technical details. Models covered by the Free plan's monthly credits carry a Free mark. The footer of the menu has Profile and Provider filters, remembered in your browser, and a Reset button.

Your model choice applies across the whole app, not only to the current conversation: it is used for your next message in any chat, and your browser remembers it.

When the selected provider supports it, a Reasoning effort control appears in the menu footer, next to the filters. Like the model, it applies app-wide, but it resets to Auto when you reload the page. The chat in the side panel has no effort control.

The levels are Auto (the model's own default), then minimal, low, medium, high, xhigh, and max, from the lightest to the deepest reasoning. Deeper levels are slower and cost more. When a model does not support a level, the nearest level it supports is used instead.

Using your own API key

On the cloud, you can run chats and agents on your own model provider key from Settings > AI Providers > Your keys. It requires the Enterprise plan (or Team, for the accounts already on it). Your provider bills the tokens, and LiveContext charges a small fee per turn. See Models & providers.

Attachments

Attach files (the paperclip) adds files to a message, and so does dragging files onto the message box. You can attach several files to one message, along with your text.

Two limits apply, and they are not the same thing. You can upload a file of up to 100 MB. The model reads a file's contents straight from your message only up to 10 MB (below that, a very long text can still be cut and a very large image may not be shown to the model, depending on the model). A larger file is still uploaded and saved with your files: the assistant receives the file itself rather than its contents, and can hand it to a tool or a workflow step that takes a file, such as Extract from File, which reads a large document in parts. Images get a thumbnail that opens full size on click. PDFs have their text extracted so the model can read them, text files are read as they are, and the assistant opens Word and Excel files with its file tools. If a file is refused or its upload fails, the reason shows under the file in the composer. An upload cut by the connection or a server error is tried again the next time you send; a file that was refused is not: remove it and attach it again.

Tools, skills, and chat settings

The Tools button in the composer opens a panel with three tabs:

• Tools: the integration tools available to the conversation.
• Skills: turn skills on or off for this conversation. Untouched, the list follows the platform's default skills. In an agent chat, this tab edits the agent's own skill list instead. See Skills.
• Options: the chat settings described below.

The settings apply at one of three scopes: Conversation settings, Agent settings (in an agent chat), or Defaults for next conversation. Workspace defaults for new Orbi and agent conversations live on the Agents page, in its Settings tab.

Below them, Advanced mode is a collapsible section, closed at first and the same one the agent editor shows. While it is collapsed, its header still says when web search is off or memory is recall only (Web search off, Recall-only memory). It holds:

The Advanced mode section also holds per-turn safety stops that prevent runaway loops:

When a turn reaches one of these limits, or Max iterations, its reply says which one and offers Raise the limit, which opens these settings on it. What was done is kept: send continue to resume. In a chat with an agent, the limits are the agent's own, so raising one there changes it for every conversation of that agent.

A microphone button (shown when your browser supports speech recognition) dictates into the composer and adds the transcript to what you already typed.

Approvals: authorization cards

Before Orbi does something sensitive, a card asks you to authorize or decline it. That includes: installing or running an application, continuing a paused interface, resolving a pending approval step, running or restarting part of a workflow, setting or removing a production version, running an agent, calling an integration directly, sending or deleting email, and creating or updating an agent with a schedule. Installing an application shows its marketplace listing in the card.

• Authorize (Install for an application) lets the action go ahead. If the assistant had already paused, it resumes as soon as it is free.
• Decline blocks that one action, and the card reports that the agent stopped. If the assistant is still holding the call, it is told you refused and can go on with other work or end its turn. If its turn had already ended, nothing resumes it: you take over from there.
• The Don't ask again in this conversation checkbox turns on the same setting as Run sensitive actions without asking.

A separate card groups the external services a workflow needs. If every credential is already connected and healthy, it approves and closes itself. If one may be expired or invalid, it stays open with a way to manage or retry it.

Connecting an app from a card

When a tool needs a credential you have not connected, a connect card opens the credential wizard for that service. After an OAuth sign-in, you come back to the conversation with the wizard showing the connected state. The services card resumes the assistant automatically. After a plain connect card, tell the assistant to continue if it does not pick up on its own. When several cards are open, each one resolves on its own.

When the assistant asks you a question

When it needs a decision, the assistant can put a question card in the thread, titled The assistant has a question. A card holds 1 to 4 questions, each with 2 to 4 options.

• Each question says Choose one or Choose one or more. Other lets you type your own answer.
• Move with Back and Next, then Send answer. Skip tells the assistant to continue without your answer.
• Number keys pick an option and arrow keys move between options.
• If the assistant's turn already ended, your answer starts it again. Your answers stay visible in the thread afterwards.

When nobody is in the chat (for example an agent running on a schedule), questions and approvals go to your linked chat channel instead. See Chat channels.

What a conversation made

Every workflow, application, table, interface, agent or browser session the assistant created, changed, ran or opened in a conversation, and every file it showed you, is listed once, the last one touched first. Above the message box, the latest ones appear as chips: one click opens it in the side panel, View all lists them all, and the cross hides the chips in every conversation (the list offers to show them again). A chip's ... menu (on hover, always on a touch screen) offers what a side panel tab's does: Go to page (Open conversation for an agent), Go to message, and Delete after a confirmation, for a workflow, table, interface or agent. A deleted content leaves the chips and the list. The side panel, opened with no tab, lists them too under In this conversation. In a long conversation, the list covers the messages loaded so far: loading older messages adds what they made.

The Contents button in the chat header, the stacked layers icon left of the bell, opens the same list, where Message jumps back to where each one was last used. Its Journal view shows everything the conversation ran: one group per message you sent, with the agent and tool calls it triggered, its tokens, iterations, and cost (credits on the cloud, a dollar amount on a self-hosted instance). Opened from the header, the list stays open in every conversation until you close it (your browser remembers it); opened from View all, it closes when you move to another conversation.

Sharing, search, and message actions

To share a conversation, open its menu in the conversation list and choose Share conversation. It creates a read-only public link, where the conversation reads as it does here: each reply under its agent's name and face, the steps it took (never what they sent or received), and its cards, read only. (On a workflow page, the Share button in the chat header publishes the workflow instead: see Marketplace.)

Your workspace has one quota for all shared links: conversations, applications, and chat and form endpoints. Disabled links still count. Only deleting a link frees its slot.

The Shared tab of the notification bell lists every active link, with copy, open, and revoke actions.

Search finds past conversations by content or title, and the conversation list can be filtered to all conversations, agent chats, or Studio threads.

A row of the conversation list can carry one dot. A blue dot: a reply came since you last opened the conversation. An amber dot: an agent asked you something there (a question, a permission, a service to connect) and nobody has answered yet. The amber dot takes the place of the blue one, and stays until the request is answered, even once you have read the conversation. Neither shows on the conversation you have open, nor on a colleague's conversation you have never opened.

Each assistant message has these actions:

Your own messages have Copy only.

Memory

Within a conversation, context compaction summarizes older turns so the assistant keeps the gist without carrying the whole transcript. That is separate from long-term memory: durable facts agents save while they work, which you can view, edit, pin, or delete from the Memory tab of the Agents page. See Agents.

Sending while the assistant works

You do not have to wait for an answer before sending the next message. While the assistant is working, new messages are queued instead of interrupting it. The queue holds up to 5 messages. You can Edit, Remove, or reorder a queued message, or use Send now, which stops the current answer first (like Stop, with the runs it launched) and then sends it.

The queue and your unsent draft are kept per browser tab. Text-only queued messages and drafts survive a page reload for up to 30 minutes. Queued messages with attachments are kept only until you reload.

Notifications and generation

The notification bell in the header has four tabs: Inbox (pending approvals, failed runs, expired credentials, invitations, and more), Triggers (your live automations and when they run next), Activity (recently edited items), and Shared (your shared links). See Notifications.

Daily check-in: on the Agents page, in its Settings tab, under Look after my workspace (shown unless it is set to Only when I ask), you can ask Orbi to look at what needs you once a day, at the hour you pick, every day or on weekdays only. When nothing is new, nothing runs and no credits are used. When something is new, Orbi handles it the way Look after my workspace says and writes a short report: on your workspace channel, or by email when there is no channel, and always in the bell. In a workspace you share with others, the report is never posted on the team's channel: it comes by email when your plan includes email alerts, and only you see it in the bell. Send me one now shows you one at once (that turn uses credits like a message). On the cloud it is included from a paid plan.

To generate media from a conversation, turn on Generation in the Options tab. For a dedicated space, switch to Studio, or use the Generate an image, video or sound button in the side-panel chat. On a self-hosted instance, generation stays unavailable until an administrator enables it.

Troubleshooting

Related pages