Chat Tour

Say what you want, and the studio does it. Chat is how you direct Circuitry in plain language — a conversation that can see your documents, act on your selection, and drive coding agents across your whole project. Every other surface in the studio is something you shape with your hands; chat is where you shape things with intent. Pair it with a code file and the agent sees what you're looking at. Point it at a diagram and it builds what you sketched. Hand it to a coding-agent CLI and it becomes mission control for work happening on your machine.

One surface, three ways to talk

The icon at the left of the chat's header tells you which conversation you're in:

• Agent mode — the full studio agent, with tools. The icon glows blue when the agent's tool connection is live, emerald otherwise.
• Plain chat — a straight conversation with the model, no tools.
• CLI mode — the tab becomes an interactive terminal running a coding agent (Claude Code, Gemini CLI, or Codex CLI) with your project as its workspace.

Chats linked to a chat node in a workflow show a Synced badge here — the conversation and the node share one history.

A tour of the controls

Header toolbar

• New chat — start a fresh session. The new chat opens in whatever mode you're already in: start one from a CLI chat and you get another CLI chat, ready in the same project.
• Model picker — a dropdown showing the current model. In agent mode, each row lists the model name with its provider badge and a check on the active one. In CLI mode, the same dropdown switches between installed coding-agent CLIs — Claude Code, Gemini CLI, and Codex CLI — with an amber "not installed" flag on any that aren't available.
• Switch to CLI / Switch to Agent — flip the tab between agent chat and an interactive coding-agent terminal. Appears when you have a project folder open and a Circuitry Server connection.
• Clear chat — wipe the conversation (disabled when there's nothing to clear, or in CLI mode).

The composer

The message box at the bottom, plus a control row underneath it:

• Send — send the typed message; ⏎ sends too. Disabled while a turn is running or when there's nothing to send.
• Dictate — speak your message instead of typing it. See Talking instead of typing below.
• Stop — while the agent is working, a small square stop button appears next to the Working… indicator; click it to abort the turn in flight.
• Permission mode chip — on agent chats, a chip at the left of the control row cycles through four modes with ⇧⇥ (or a click):
  • ▸ auto — the agent runs tools freely.
  • ? ask — approve each change before it happens.
  • ✓ accept — also auto-accept code edits.
  • ☰ plan — read-only; the agent plans but touches nothing.
• effort — on reasoning-capable models, a per-chat reasoning-effort selector (auto, off, minimal, low, medium, high, max). You can also type /reasoning <level> in the chat.
• model — in CLI mode, a per-chat choice of which model the coding-agent CLI runs with (CLI default, Sonnet, Opus, Haiku, or Fable).
• Remove attachment — when images or files are staged, thumbnails appear in a row above the composer; the ✕ on each one removes it before sending.

Several chats at once

Chats stack in the chat panel as tabs, so a long-running coding job can keep working while you talk to a second agent about something else. Open chats is the list of them: every chat you have open, with its colour, its name, and a working / needs-you indicator, so you can jump between them without reading tab titles.

• starts another chat in the current mode.
• Right-click a row (or press and hold on a tablet) to rename it or give it a colour; double-click goes straight to renaming. A name you set sticks — automatic titling never overwrites it.
• on a row closes that chat.

Where the list lives

The ⚠ unknown icon “MoreVertical” menu in the list's header decides where Open chats sits:

• In the side panel — the default. The list appears in the left panel whenever you're working in the chat.
• Dock left of chat / Dock right of chat — the list becomes a column attached to the chat panel itself. On a wide screen the left panel can be a long way from a chat docked at the right edge; this keeps the two together.

Docking takes its width from the document area, so the conversation stays exactly as wide as it was. Once docked, the divider between the list and the conversation resizes the list, while the outer divider still moves the whole chat panel against your document.

Keyboard

Shortcuts use ⌘ on Mac (Ctrl on Windows/Linux).

Cmd + and Cmd − zoom the chat text in and out while the chat panel has focus.

⌘T works while the chat has focus. On the web it can be claimed before Circuitry ever sees it — if nothing happens, put New chat on a different key in Settings → Keyboard Shortcuts.

Talking instead of typing

When a message is easier to say than to type, dictate it. There are three ways to start:

• Hold Space — hold the spacebar for a moment and the microphone opens; let go and it closes. A quick tap still types a space as normal, so ordinary typing is unaffected. This is the fastest way to say one thing and send it.
• the microphone button — next to Send. Click to start, click again to stop.
• ⌘⇧Space — turns dictation on and leaves it on, so you can talk without holding a key down. Esc stops it.

Your words appear in the message box as you speak, firming up as the recogniser settles on them. Nothing is sent automatically — read it, fix anything it misheard, then send. You can type in the middle of dictating too; your typing wins, and speech carries on from wherever you left the cursor.

While it's listening you'll see a small bar standing where the next word will land, moving with your voice. That's there to answer the question you'd otherwise have to guess at: is it actually hearing me? If the bar doesn't move when you speak, the microphone isn't reaching the studio — check which input your system has selected. The bar steps aside while words are arriving and returns during the pauses.

Dictation uses whichever microphone your operating system is set to use, so switch inputs in your system's sound settings rather than here.

Things you can say

Some phrases are treated as commands rather than words. A phrase only counts as a command when you say it on its own — "the space between them" types those words, while "space" said alone types a space. That rule is what keeps ordinary English dictatable: "over a period of time" and "a dash of salt" are prose, not punctuation.

On a Mac, punctuation is usually added for you as you speak, so you rarely need the punctuation commands — they're there for when it isn't, and for dictating in a web browser.

Nothing you can say will send the message. There is deliberately no voice command for that: dictation puts words in the box and stops there, so a misheard phrase can never fire a half-finished message at the agent. Sending is always yours, with the button or ⏎.

Spacing is looked after too. Dictated words join what's already there the way you'd type them — a space between words, none before a full stop, and none inside a hyphenated compound like "self-aware".

It turns itself off. If you leave it listening and stop speaking, dictation ends after a short pause, and any session ends after three minutes regardless. Holding Space is exempt — while you're holding the key, pauses are yours to take. This is deliberate: a microphone left open helps nobody, and you'll be told in the message area when it switches off.

Where it works. Dictation is available in a web browser — Chrome, Edge and Safari — and in the desktop app on Mac, Windows and Linux. iPad support is on its way. Where it isn't available the microphone button is greyed out and says why.

On a Mac your speech is turned into text on your own machine: the audio never leaves it, and it works with no connection. On Windows and Linux the desktop app opens a small listening window in Chrome or Edge — you'll see it appear with its own level meter, and it closes when you stop. One of those browsers needs to be installed.

The first time you dictate in the desktop app, macOS asks for permission to use the microphone and to recognise speech. Both are needed. If you decline, the microphone button keeps working but will explain what's missing — you can change your mind in System Settings under Privacy & Security.

Works with the rest of the studio

• Agent Chat in Split-Screen — pair a chat with a code file or workflow; the agent sees your file, selection, and containing function, and its edits land as reviewable diff hunks.
• The Coding Agent — what the studio agent can do across your project.
• Circuitry as an Agentic IDE — the whole loop: plan visually, direct in chat, inspect in a real IDE.
• The Smart Whiteboard — when intent is easier to draw than to type, sketch it and let the agent read the board.
• Workflows Overview — chat nodes put a conversation inside a running flow.