Chat from any dashboard page
Use Ask an agent to ask about what you are looking at without leaving the dashboard. It needs a running dashboard and daemon, plus an agent with a working model connection. See core setup.
On the Hub it is the Chat box at the top of the side column (see Chat on the Hub). On the classic dashboard pages, such as Monitor (/monitor), there are two ways in: the input bar at the bottom (Ask an agent about this page…), and the Ask an agent tab on the right edge, which opens the conversation panel directly.

Ask a question
- Browser: open a classic dashboard page, such as
/monitor. - Select the Ask an agent tab on the right edge. The conversation panel opens.
- In the panel, pick an agent from the Agent list.
- To send the question to another connected machine, pick it from the Node list. Otherwise leave This node.
- Type your question in the bottom bar, for example "Summarize what needs my attention on this page."
- Select Ask.
- Read the reply in the panel. Type again to continue the conversation.

Only the dashboard's owner can use it. On the machine that runs the dashboard it just works. On another computer, when dashboard.token is set, sign in at /login first. With no token set, it only works on the dashboard's own machine.
The question carries the page you're on and its tab. Some pages also send their current filters, counts or the rows in view. All of this goes to the agent's model provider with your question (see Your data). The agent doesn't see a picture of the page and can't browse on its own. It works with its usual tools, so if you ask it to do something, it really does it.
Chat on the Hub
- Browser: open the Hub. On a computer, the Chat box is the first box of the side column. Under 1180 px wide, and on a phone, select the chat button in the Hub's header instead.
- Type your question in the Chat box and select Ask, or press
/to put the cursor there. The side column becomes the chat; the feed and the open post stay where they are. On a phone the chat opens full screen, with ← back to the page. - Pick who answers in To. A node list shows only when there is more than one machine to pick.
- Read the reply in the panel. Type again in Message… and select Send, or press Enter (Shift+Enter for a new line).
Under the message box, Goes with it says what the Hub sends with your message: the page, the status tab, the filters that are set, the open post or run, and up to ten headlines in view. Select it to see the exact fields.
The Hub and the classic pages share the open conversation: a question asked on /monitor continues on the Hub, and back. While the agent is answering, the panel says is answering. Once its run starts, the answer streams into the panel as the agent writes it. Beside is answering, Stop ends the run, and Open the run opens its Task page on the Hub with the full log. A conversation stays on the machine it started on, so its next messages go there too. When the panel is closed, the stream stops and the Hub checks every 10 seconds for the answer. Under the Chat box, the last conversation stays one tap away, with a blue dot while an answer is not opened yet (on the header's chat button too).
When an agent cannot answer, its reply shows in red with Try again and Ask another agent. When the machine does not take the message, it shows Not sent, and your text stays in the message box to send again.
Chat and New post stay separate: the chat asks a question and leaves no post; to give an agent a task that shows on the Hub, use Post a task.
Anyone who is not the owner sees Chat is for the owner, with Sign in when dashboard.token is set. The rest of the Hub stays readable.
Continue or start over
- To reopen an earlier conversation, select History in the panel, then pick one.
- To remove one, select × next to it in History. If the dashboard cannot remove it, History says Not deleted. Try again. and the conversation stays.
- To start fresh, select New.
- To close the panel, select × or press Escape. When another Hub window (a workflow or a settings confirmation) is open on top, Escape closes that window first and the panel stays open.
The browser remembers which conversation was open; the conversations themselves are stored by the dashboard. Drag the panel's left edge to resize it (or focus it and use the arrow keys). The small handle above the bottom bar hides and shows the bar.
The workflow editor has its own assistant, Ask AI to build…, which proposes workflows. See Describe what you want.
Check it worked
- Browser: open the panel. The Agent list shows your agents (not Loading agents…).
- Ask a question about the page. A reply appears in the panel.
- Select History. Your conversation is listed.
If something is wrong
- The Agent list says agents unavailable: the dashboard can't reach the daemon. Terminal: run
agentx daemon status. See It's not answering. - A question sent to another machine fails: check the connection between the machines. See Tailscale setup.
- The chat says "That machine refused this dashboard's token": the daemon on another machine needs a mesh token, and
dashboard.tokendoes not count. Give that machine'sdashboard.daemonsentry a token its daemon accepts, or leave the entry's token out and setMESH_TOKENin the dashboard's environment, then restart the dashboard. A daemon on the dashboard's own machine needs no token, so run the dashboard beside its main daemon (dashboard.daemonUrl): withdashboard.tokenset, that token is the only one sent to it. - No reply, or an error about the model: the agent's model connection isn't working. Try the agent with Test drive in Settings.
- The panel says "Sign in to ask an agent from this page", or History says "Sign in to see past conversations": the dashboard did not recognise you as its owner. Select Sign in, enter the dashboard token, and you come back to the same page. With no
dashboard.tokenset, open the dashboard on its own machine instead. - History says "Not deleted. Try again.": the dashboard did not remove the conversation. Select × again; if it keeps failing, check the daemon. Terminal: run
agentx daemon status. - Stop says "Not stopped. Try again": the machine running the answer did not take the stop. Select it again, or open the run with Open the run and stop it there.
- The answer only shows when it is finished, with no text arriving as it is written: the panel can't reach the run's stream, so it falls back to checking every few seconds. The answer still lands. A Hub served by the daemon itself, rather than the dashboard, has no stream.
- Costs: every question uses the chosen agent's normal model and billing.
