Chat

Chat lets you ask questions and make governed changes from the architecture you already mapped.

One chat, everywhere

There is one conversation in the right rail. A Chat button appears on the canvas, hub, aspects, intents, skills, and Pages. Opening it from any screen resumes the same thread.

The screen behind the dock stays usable while chat runs.

What this conversation is about

Every conversation has a scope chip. Scope tells chat what to focus on and limits what it may change.

Click the chip to change scope. Widen it for bigger questions; narrow it for one element or board.

Expand

1. 1

   Click the scope chip

   Every conversation has one, and clicking it changes what the conversation is about.

2. 2

   Widen or narrow the scope

   From one selected element up to the whole workspace.

3. 3

   A facet scope

   Open chat from a facet's screen and the conversation scopes to it.

What chat can do

Chat reads nodes, edges, aspects, bound sources, and past insights. It picks tools from the same registry exposed to external agents over MCP access, with the same policy applied.

In practice that means it can:

• Answer from the graph. "What breaks if the auth service goes down?" "Which services have no owning team?"
• Draw. Ask for a diagram and chat can reply with a context board built to answer the question. It appears as a live preview in the conversation without changing the board you're viewing.
• Edit the board you're on, within the scope's fence.
• Draft intents. Every intent chat creates asks for your approval first and is created as a draft for review. Chat proposes; you decide.
• Start an analysis, below.
• Reach your own tools, if you've registered MCP servers for chat to call.

An answer can be prose, a context board, or a batch of insights. Scope decides which tools are available on that turn. The plugins' /discover draws context boards too — persistent ones that stay in the hub's Context boards card until the next run replaces them.

Chat can also read reference sources, such as an ADR, PRD, or runbook, on demand.

Changes are never silent. A board edit or new intent asks in the conversation or is staged for review, and chat can only change boards you could change yourself.

Turning it on

Chat starts on your organization's trial credits. The first message you send offers to turn them on. When they run out, connect one org-level model provider: Settings → Connectors, then authorize a provider such as OpenRouter. From then on chat runs on your own account and models. Plugins need no provider either way.

Quick starts

A new thread opens with a few quick starts. Each one only pre-fills the message, so you can edit it before sending or ignore them and type your own question:

• Assess this board runs a structured review and records the surviving insights as a batch.
• Draw this as a context board answers a question with a board built for it.
• Draft an intent from a document turns a PRD, RFC or decision into anchored work.
• Set up this workspace appears only while nothing has been modelled yet. It interviews you about your systems, draws the landscape, and creates each app's empty board. The boards are empty on purpose; content comes from each repo's first push. Once the workspace has content, this quick start disappears.

There is no response format to choose. Chat reads what you ask for and picks the workflow: an answer, a drawn board, a review, or drafted intents. When a turn creates or draws a board, a card in the conversation links to it.

Starting an analysis from chat

Chat is the one place inside the portal where an analysis starts. Ask for one and chat matches an insight skill, or runs the one you name, over the conversation scope. Each analysis produces a batch of insights.

There is no Run control on the canvas. The Insights Bar lists insights as they arrive but does not start an analysis. The other trigger lives outside the portal: /insights in the code or connect plugin.

What chat isn't

Two things nearby get mistaken for chat, and both are worth separating.

It is not the write-session pill. The amber "{n} uncommitted changes" pill in the Topbar shows the workspace-wide working copy. It includes every uncommitted change across every board. Commit and discard live there, not in the chat dock. MCP Access covers it.

It is not where large architecture work belongs. Each chat turn is bounded so the conversation stays responsive. Restructuring a whole board tree, adopting a decision across a dozen apps, or working from a folder of local design docs belongs in the architect workbench. It reaches the same boards through the same governed tools, with no turn limit.

What's next