Docs navigation
Agent

Interfaces

An interface is a private dashboard for one of your agents. You describe what it should look like in a conversation with that agent, Brainbase builds and hosts it, and the people you choose sign in with their Brainbase account and use it. They never see the agent's configuration, the console, or anybody else's work.

Use one when the people who need an agent are not the people who build it. A support lead can give their team a page with the three buttons they actually use, instead of teaching them the console.

An agent's dashboards are listed in the Interfaces tab on the agent's page, and each one has its own screen: a conversation with the agent, where you describe and change the dashboard, beside a workspace where you preview it, publish it and change its settings.

How to ship a dashboard
UI
  1. 01Choose New interface, describe what the dashboard should show, and choose who can open it.
  2. 02Choose Start building. The interface's screen opens and the agent starts building in a conversation.
  3. 03Watch it in the Preview tab and ask for changes in the same conversation.
  4. 04Choose Publish, check that you are among the people who can open it, and publish the build. Or tell the agent to publish, in a message of its own.
  5. 05Open the address as a member of the audience to see it.
Which harnesses can build
Interfaces works on agents running Claude Code, Codex, Qwen Code and OpenCode. An agent on any other harness cannot build dashboards: that includes Kafka, which is the default harness for most agents, and Cursor, Factory and Qoder. The tab and the New interface page say the agent's harness cannot call the interface tools, and Build in chat and Start building stay unavailable.

Create an interface

Choose New interface in the tab, describe what the interface should show, and choose Who can open it. A Name is optional; leave it empty and one is suggested from your description. The Address is given when the interface is created and never changes: publishing again swaps what is served without changing the link you shared. Start building creates the interface on this agent, opens its screen, and sends your description as the first message.

The tab lists the agent's interfaces as cards or rows, with Search interfaces and a sort menu. Each shows its state, who can open it, its dates and its address, with Build in chat to open its screen and a menu to open the live page, copy the address, rename or archive it. An orange dot beside a name means a conversation is working on it. Archived ones stay hidden until you turn on Show archived, which is also where you Restore them. The agent's side menu lists up to five recent interfaces.

Choose who can open it

Set the audience before the first publish. Publishing an interface with no audience is refused, and a published dashboard opens only for the people in its audience, so include yourself if you want to see it. A new interface set to Only chosen people starts with nobody chosen. Until a release is live, the address answers as though it does not exist.

Change it from Who can open it in the Publish popover, or in the interface's Settings. There are two settings.

  • Everyone in the organisation: anyone in your Brainbase organization who can use this agent.
  • Only chosen people: just the people you tick. Up to 2,000; past that, choose Everyone in the organisation instead.

People in the audience also need access to the agent: an owner or admin of the organization, or a member of a project the agent is in. Anyone else can open the dashboard but cannot start or read conversations in it.

Everyone signs in with their Brainbase account first, and access is checked on every request rather than once at sign-in, so removing somebody takes effect on their next click. Sharing the link with somebody outside the audience gives them nothing.

A dashboard grants no access of its own
Every call a dashboard makes to Brainbase is made as the person using it. Letting somebody open a dashboard does not give them any other rights over the agent, and it does not give them access to the console.

Build it in chat

Choose Build in chat on an interface to open its screen. Anyone who can edit the agent can build there. The screen reopens your newest conversation for that interface; if you have none, the composer opens with a first message that names the dashboard, and the conversation starts when you send it. Tell the agent what the dashboard should do: which conversations to show, what the buttons say, what a person fills in before starting work. Ask for changes the same way.

Beside the conversation, Preview shows the dashboard, Code, Shell, Files and Browser show the agent's computer, and Settings holds the name, address, audience, versions, releases and Archive. Share and Publish sit at the top right.

An interface can have several conversations. The menu on the interface's name above the chat lists yours on it, newest first, with New conversation. Each conversation works on that one interface only: asked to touch another, it refuses and names the one it is for. Every other conversation of the agent has no dashboard tools, however it was started. Build conversations also appear in the agent's chat view and Tasks list, marked with a small i.

Only one conversation can hold an interface at a time. While one is changing the code, any other is told to wait, and the screen shows a banner saying so, with Open that conversation when it is one of yours. The hold ends when the change is committed or fails, or when the conversation holding it stops answering for long enough that the next one can take over.

Each change the agent commits becomes a version, and a version has to build before it can go live. A card under the conversation shows the version its change produced and that version's build, with Publish once the build succeeds. History, at the top of the chat, shows the most recent generation, one round of the agent checking the code out, changing it and committing: whether it is queued, running, succeeded or failed, why it failed, and which conversation holds it. Below that are the interface's Versions, Builds and Releases. A build that fails says why, and the conversation is where you carry on from there.

Preview while you build
The Preview tab shows what the open conversation is serving, at desktop or phone width, with your own data: it signs you in the same way the published dashboard does, so tables and lists show what they will show you once published. Only people who can edit the agent can open a preview. When nothing is serving, the tab starts it by itself for the person who owns the conversation: the working copy if the sandbox still has it, otherwise the last saved version, and it says which one you are looking at. While the agent is working, or when someone else opens the tab, it waits for you to choose Start preview, which is there whenever nothing is serving and the sandbox can start it. The tab is not available on every machine type. If your browser blocks sign-in inside the panel, the tab says so and offers Open in new tab. Share gives a Preview link, the interface's preview address: whoever opens it signs in with their Brainbase account, must be able to edit the agent, and sees their own data. It shows the conversation whose preview was opened last.

Publish it

Publishing deploys a build that succeeded, so it takes a moment and the page follows along. Whatever is already live keeps serving until the new release is ready, and a publish that fails leaves the old one in place. There are two ways to publish.

  • From the screen: choose Publish at the top right. The popover says whether the interface is live and shows its address, who can open it, which you can change there, and what goes live: the newest build that succeeded, or another version picked with Change. A publish that fails says why and offers Try again. Publish on a version's card or in History opens the same popover on that version.
  • From the conversation: tell the agent to publish, in a message of its own. It publishes the latest build that succeeded and replies with the address, who can open it, and that it is live.
Build and publish are two messages
The agent publishes only when the message that asks for it asks for nothing else. A message that asks for a change and a publish together gets you the build: the agent makes the change, reports the result, and waits for you to look at it and say publish. The instruction has to be your own latest message in the conversation; the agent cannot decide to publish by itself.

What a dashboard can do

People in the audience sign in at the dashboard's address and use it as themselves. A dashboard reaches a fixed, small set of actions against its own agent, and nothing else on the platform. In practice that is everything a person needs to use an agent and nothing they need to administer one.

  • List the conversations they started from this dashboard, and open one.
  • Read the messages in a conversation.
  • Start a new conversation.
  • Send a message.
  • Watch the agent work as it happens.

A conversation started from a dashboard never gains the ability to build, whoever is signed in, so a viewer cannot change the dashboard they are using. The conversations you used to build a dashboard do not appear in it either.

Roll back or archive

The Releases list, in History and in Settings, shows the most recent releases, newest first, and Roll back to this serves an earlier one again at the same address without rebuilding anything. Rollback and audience changes are yours alone; the agent has no way to do either.

Archiving an interface, from Archive in its Settings, takes it offline and out of the list, and its screen becomes read-only. Restore it from Show archived and it comes back as it was: same address, same versions, same audience, serving the release that was live when you archived it. There is nothing to republish.