Chat / Cowork
Chat/Cowork is the studio's chat surface: the place where you talk to your own workflows. Use it to test an agent before customers do, to ask questions of your knowledge base, and to work alongside the agent as a coworker, telling it to look up customers, book appointments, or send quotes.
Open it from the sidebar: Workspace, then Chat/Cowork. Chat always operates inside your active project (chosen with the project selector in the top bar), and your chat history in a project is private to your account.
The chat screen
The screen has two parts:
- The conversation area with the message thread and, at the bottom, the composer: a text box, an attach button (plus icon), a Show history toggle, and Send. Press Enter to send; Shift + Enter inserts a new line.
- The history panel on the right (open it with Show history). It holds the New chat button, the collapsible Session context card, and the list of your saved chats. While the panel is closed, a round plus button floats at the top right of the conversation as a shortcut to start a new chat.
On an empty chat you see the greeting "How can we help today?" and four starter prompt cards. Clicking a card fills the composer with that prompt so you can edit or send it. If the selected workflow defines its own starter prompts (set in the workflow's Edit Workflow dialog, which has four Starter Prompts fields), those replace the default cards.
Starting a chat
Click New chat (or the floating plus button). The "Set chat context" dialog opens with four choices; all are optional:
| Field | What it does |
|---|---|
| Workflow | Which workflow answers in this session. Leave it unset to use the project default. |
| Entry point | Where in the workflow the conversation enters. The default is the workflow's start node; the list also offers the workflow's other conversational entry states (message receivers, HTTPS listeners, webhooks, schedule triggers, and similar). |
| Knowledge bases | Which knowledge bases retrieval may search. Pick one or more from the list; each appears as a removable chip. Unset means the project's default selection. |
| Private indexed folders | Private root folders to include in retrieval, also added as chips. Public indexed root folders are always included automatically; this field only adds private ones. |
Click Start chat to open the session, or Cancel to keep your current one.
These four choices are the session's binding. The binding is saved with the session on the server, so reopening a chat days later restores the same workflow, entry point, knowledge bases, and folders. Different sessions can have different bindings, which makes it easy to keep one chat per workflow or per customer folder.
The session context panel
At the top of the history panel, the Session context card (expand it with the arrow) shows what the current session is bound to:
- Workflow: the bound workflow's name, or "Auto" when the session has none and chat is choosing for you.
- Entry: the entry point, or "Workflow start".
- KBs: the selected knowledge bases, or "Default".
- Folders: "Auto public" plus any private folders you added.
Below that, a Developer IDs section shows the raw project ID, workflow ID, knowledge base IDs, and folder IDs with copy buttons. These are handy when you move from chat to the API, which takes the same identifiers.
Session history
The history panel lists all of your saved chats in the project, newest first, each with its title and date. A session's title comes from its first message.
- Switch chats by clicking an entry.
- Delete a chat with the trash icon on its row. You are asked to confirm ("Delete this chat history permanently?"), and deletion is permanent. Deleting your last session starts a fresh empty one.
Sessions are stored on the server per project and per user: you see only your own chats, and switching projects switches the list.
Attachments
The plus button in the composer attaches a small text file to your next message. Supported types are .md, .txt, .json, and .csv, up to 400 KB. The file's content (up to the first 8,000 characters, marked "(truncated)" beyond that) is inserted into the conversation as a message, so the agent can read and discuss it.
This is for handing the agent a one-off file in conversation. To add documents the agent should permanently know and cite, upload them to the Knowledge Base instead.
How replies are produced
When you send a message, chat runs the bound workflow (see Workflows):
- Chat resolves the effective binding: your session's workflow, entry point, knowledge bases, and folders.
- The workflow executes with your conversation as input. Its agent reasons, searches the selected knowledge, and calls its tools as needed.
- The reply comes back and streams into the thread. Replies that contain tables or charts appear all at once behind a short "Composing" indicator instead of typing out.
Because the agent knows who it is talking to (your name and email are passed along, the same way a channel message carries the sender's contact details), action tools can default to you as the customer when you do not name one.
Real actions from chat
The agent's abilities depend on the tools its workflow grants it (see Agent tools). With the right tools and connectors, chat commands have real effects:
| You say | What happens | Requires |
|---|---|---|
| "Book a 30 minute call with jane@acme.com tomorrow at 2pm" | A real event is created on your connected Google Calendar and the customer is invited. | Book appointments tool, Google Calendar connected. If no calendar is connected, the agent says so instead of pretending. |
| "Send a quote for $150 for a consultation" | A Square quote is created and the hosted payment link is shared in the chat. Asking for an invoice instead makes Square email it to the customer. | Quotes & invoices tool, Square connected. |
| "Has that invoice been paid?" | The agent checks the Square invoice status. | Invoice status tool, Square connected. |
| "What does our refund policy say?" | The agent searches the selected knowledge bases and answers with citations. | Knowledge base search tool, indexed documents. |
| "Note that Acme wants a follow-up next month" | The agent files a markdown note into the knowledge base (the workflow's vault or chat memory folder). | Knowledge notes tool. |
| "Look up the customer jane@acme.com" | The agent finds (or, on request, creates) a customer record in Customer Management. | Customer lookup or Customer creation tool. |
These are not simulations. A booking creates a real calendar event and emails a real invitation; an invoice is a real Square invoice that can collect payment. Test with your own email address first.
Sources and tool traces
Assistant replies carry their evidence:
- A Sources (n) pill under a reply opens the Sources dialog, listing the documents and images the answer drew on (with an Open button for each file) and the connector tools it called, labeled "Tool call" with the tool and connector name.
- When a reply used tools, a collapsible Used n tools line under the reply lists each call inline.
When the reply cites specific sources, only those are listed; a reply produced by tool calls attributes the tools rather than every retrieved document.
Charts and tables
Workflows can return structured artifacts along with their text. Chat renders them as line or bar charts (with legends and axes) and as data tables under the reply. Markdown tables in the reply text render as formatted tables as well.
When something goes wrong
- The workflow run fails. The failure is shown as the assistant's reply: you see the workflow's error text in the thread (or "Workflow execution failed before producing a response." when there is none). This is intentional; chat is where you debug workflows, so errors are surfaced, not hidden.
- The workflow pauses for human approval. The reply says the workflow is waiting for human approval; it resumes once a reviewer acts.
- The request itself fails (network problems, or your organization is out of credits). The error message appears in red next to the Send button, and your message stays in the composer history so you can retry. Credit errors mean it is time to visit Billing & Credits.
- The reply is empty. You will see "Workflow completed without a response message.", which usually means the workflow ended without reaching a reply step; review its graph on the Workflow page.
Default workflow behavior
You do not have to pick a workflow for every chat:
- Each project can mark one workflow as its default (the "Default workflow for this project" checkbox in the workflow's Edit Workflow dialog; only one workflow in a project can hold it).
- When you have not selected a workflow, chat preselects the project's default workflow, or the first workflow in the project if none is marked default.
- Sessions without a saved binding show "Auto" as the workflow in the session context panel.
This makes the default workflow the project's front door: teammates can open Chat/Cowork and just start typing. See How Faberiq is organized for how defaults fit the project model.
Tips
- Keep one session per task or customer. Bindings are per session, so a chat bound to a private customer folder stays scoped to it every time you return.
- If answers ignore your documents, check the session context panel: the right knowledge bases must be selected, and documents must be
indexedin the Knowledge Base. - Before pointing customers at a channel, run through the whole conversation here first; the studio chat exercises the same workflow, tools, and knowledge the channel will use.