Ask the agent

Every area — Overview, Components, Schematic, PCB, Enclosure — has a collapsible chat panel at the foot of the screen. Collapsed by default, so it never costs the area its width; the app remembers whether you left it open, per area.

It is five agents, not one

The tab you are on picks the agent, deterministically, before any model is called. There is no router reading your words to guess where to send them — the area key is real, typed state the app already has. Typing a question into the Schematic chat can only ever produce a schematic answer.

Each agent is grounded in different real state, and can call different tools:

Area	Grounded in	Can call
Overview	Project intent, your last check results, export history, the parts this project references	Search, library lookups
Components	The selected part’s design guidance, connection guidance, pins, package, datasheet	Search, datasheet reading, library lookups
Schematic	Parts your project uses, plus real ERC findings	Search, datasheet reading, ERC results, part search
PCB	Real DRC findings, component heights	Search, DRC results, component-height lookups
Enclosure	Board outline, mounting holes, component heights	Search, component-height lookups

On Components, the chat is scoped to the part, not the project — a Part is a real object you own (see Find a part), so its history follows it everywhere, into every project it is used in. Every other chat is scoped to the project.

Every answer carries its sources, or says it does not

The same contract design guidance already makes: an answer renders clickable source chips beneath it. A datasheet chip opens your cached PDF at the exact page. Content the model is offering as general engineering practice — not something your specific datasheet says — is visually marked as such, never blended in with a cited fact.

Nothing here is described as “verified.” The chip is one click from the real page precisely so your own eye can be the last check, in about five seconds — the same discipline design guidance uses.

Answers arrive whole, not token by token

There is no streaming. While the agent is working — reading a datasheet page, checking a search result — you see what it is doing, not a typing-indicator animation implying a token stream that is not actually happening.

Save as note

Found an answer worth keeping? Save as note promotes it to a durable, cited record on the part or the project — the same kind of real object the library already is, not something you have to scroll back through a transcript to find again. The next conversation starts from it instead of re-deriving the same answer.

Project intent

From the Overview tab, you can state what you are building — “a macropad from scratch,” “a battery-powered soil sensor” — in a plain-text field, editable any time. Entirely optional; every project works exactly the same with it empty.

It is injected verbatim into every agent’s context as your stated goal, never treated as a verified fact about the design. It is what lets an answer be about your board instead of a generic one.

What replaced the old command box

An earlier version of Overview’s chat recognized two typed commands — generate <part> and inject — alongside plain questions. That design is gone; see What it is, and what it is not for why. The capabilities moved to where they actually belong:

• Generating a component directly from a part number, for when search cannot find it, is now a real fallback next to search on the Components tab — it lands you in the same review-before-save flow as a real search result, rather than a raw, unreviewable JSON block in a chat message.
• Injecting a component into the board KiCad has open is now a real action on a saved part’s own Part Detail page, behind the same explicit confirmation step as before — nothing writes to your board without you reading exactly what it is about to do first.