Chat

Conversation

Answers render as markdown, and the last answer ends with three suggested replies. Clicking one sends it right away.
You can stop, regenerate, edit your message and resend, or copy (editing works on your last message; regenerate and copy on the last answer). In a project with a team, “Post to team” next to the last answer puts it in the team input (you press send yourself).
The chat list on the left filters by a search box and chips (All · Chat · Flow · Drawing · Slides). Side chats of workflows, diagrams and slide decks (pptx) appear in this list too; browser and vertical chats appear only under “All”. Rename and delete here as well.
The AI searches and reads project files directly and cites sources (path#section). Clicking a source opens the folder that holds the file plus a preview in the right files drawer; “Open in Files” in the preview continues in the Files tab.
It also searches the web and reads pages. If a human check (captcha) appears during a web search, solve it in the web search box shown in the chat (or, if that box is not in view, the Browser menu’s “AI” tab) and it continues. Reading a page does not go through the browser, so it does not see pages behind your sign-ins.
Ask it to change a project md file (“add 10.30 settlement to the schedule section of the minutes”) and it edits only that spot, not the whole file. Each edit can be undone on its own from the Files history.
The AI opens documents itself. It takes a picture of a drawing layout, a pdf/hwp page, a slide, a sheet, an html page or a photo to look at, and when needed crops a part, draws numbered boxes (1 · 2 · 3 …) and saves it as an evidence image (PNG) for a report, explained by a numbered table under the image. Each saved image can be undone from the Files history.
It hands work to subagents. Making or editing a deck (pptx), a spreadsheet or an html page, writing a long md, or capturing part of a file or a web page goes to a dedicated agent (pptx · spreadsheet · html · md · capture · figure). The agent creates, looks at the result as a picture, fixes it and reports; progress shows in the chat as “🤖 pptx · …” lines. The pptx agent builds and edits decks directly with the pptx engine and asks the figure agent for pictures such as flowcharts; turning html slides into a deck is the html agent’s job (the html source stays). Each created or edited file can be undone from the Files history.
Say “mark the comparison table done in the October quotes plan” or “add legal review to my to-dos” and it changes just that one to-do in the plan (md) in the To-do tab. Say a date (“by tomorrow”, “by Nov 10”) and that to-do gets a due date (📅), so it also shows on the calendar’s plan — the AI never creates calendar events directly.
Files it creates, edits or saves stay as paths in the answer and in the progress lines (“📄 Document created: …”, “🤖 capture agent done — …”); click one to see it in the drawer like a source. The list icon at the right of the conversation header opens this conversation’s folder in the drawer.
While the AI moves the app screen or works on a web page, a pill shows in the header on every tab (“AI is moving the app screen · conversation · Stop”), and moving the app screen also draws a border around the window. Once the actual moving is done the pill turns into “AI working” and the border goes away — you can keep using the app meanwhile. Click the pill to go to that conversation (or the Browser menu’s “AI” tab for web pages); “Stop” stops that answer. Before the AI first moves the app screen, each conversation asks once with an approval card (⌘↵ approves).
Files the chat creates (notes, documents, decks, spreadsheets, evidence pictures) are saved in one folder per conversation (like notes/chat/CHAT-00012/). In the Files tab they gather under the fixed “Chat” folder after your own folders, in a folder named after the conversation title (you can move them into your folders). When the chat edits one of your existing documents, that document stays where it is.
Say “turn this into a document” and each conversation gets one note (like NOTE-00001.md) in its conversation folder that keeps being updated.
Complete things with a teammate’s data (secure share). In a project with a team, say “get name, birth date and career from Minsu’s résumé and fill the grant application” and the AI asks that teammate for just those items. When they approve in their To-do tab, their AI looks only in their shareable files, organizes the values and sends them to this device; the values go only into the result (the filled form or written text). Chat, the AI and logs hold no values — only states like sent · approved · arrived · completed and the result path. The AI doesn’t read the result back. If nothing arrives within 2 hours of approval, ask again.
Say “remember this” (while “Direct memories” is on under Settings → AI memory) and the fact is kept in the project’s memory and brought into later chats when relevant. Click “🧠 Used n memories” above an answer to see the memories it used and edit or forget them on the spot. The full list is Settings → AI memory → “What the AI remembers”; switch memory off there too. In a project with a team, say “put ‘final = the version the client replied to’ in the team glossary” and it goes into the team glossary everyone sees (team/전체/용어집.md).
Unsent input is kept per conversation, and answers keep generating when you switch tabs. If an answer finishes while you are on another screen, it shows up in the header notification bell.
Answers need you to sign in with your account under Settings → Account.
On the empty screen with no chat picked, turn on the Collection switch under “Start a new chat” and a moving painting in dots (The Starry Night at first) takes the screen — “Start a new chat” hides, and the arrows on the pill floating at the top step through the other pieces (twelve in all). Your pick stays for next time, and turning the switch off brings the button back. Tap or rub it and color spreads from that spot until it fades at the frame and at strong color edges. The piece comes from the server, so it needs the internet.

Attaching pictures and files

Pictures by paste or drag and drop, up to 10MB.
Type @ and part of a name in the input to find a file instantly. Pick with ↑↓ and Enter and it is placed in the tray above the input. Dragging from the files drawer does the same. Diagrams can be found and attached by title too; clicking a diagram chip opens it in the Diagrams tab.
When you send, the app reads the attached files and sends them along — text up to 12,000 characters, large files as their summary, pictures as they are. For an md, the pictures it embeds go along too, up to the first 4 (the same when the AI reads a document itself). A large file without a summary shows “No summary” and “Create summary” on its chip, so you know before sending.
“Copy path” in the files drawer gives @path, so pasting it works the same way.

/help

/help — shows an app intro and the menu list (enabled menus only, one line each) in the current display language right away (no AI needed).
/help question — the AI answers based on these help documents.