Jinba Flow
Jinba App
Jinba ToolboxRun status
The AI’s status is shown in the message area in real time.
While the status reads Thinking…, a one-line “Tips:” hint about using Jinba App may appear beneath it. It changes a few times over a long wait and disappears once the answer starts; hovering pauses it, and the ✕ (“Hide tips”) hides tips for the session — turn them off for good under “Tips while you wait” in Personalize settings.
Each step in the activity timeline is named in plain language rather than by its tool name. A step whose label is still being prepared shows “preparing” beside the name; a tool that offers no readable name at all reads “Working…” while it runs and “Ran a step” once it is done. Raw tool names and commands stay in the Console log.
Interrupting a run
While the AI is running, an “Interrupt” button is shown. Click it to stop processing at any time. An interrupted chat records “AI run was interrupted”, and you can send follow-up instructions from there.The workspace panel
Use the “Workspace” toggle in the chat screen to open a side panel showing what the AI has been working on. The panel shows one thing at a time. The preview takes a pane of its own beside the chat, and the tab rail on the right edge opens one list at a time next to it.- Drag a pane’s edge to resize it. “Expand to full width” gives the preview the whole chat width and “Restore size” hands it back — if there is no room to bring the chat back, the control tells you to widen the window or close a pane instead of leaving you stuck
- The rail remembers which list you had open, so reopening the panel returns you to it
- On a narrow window (roughly 970 px and below) the panel keeps its earlier form: a single drawer with Results, Files, Code, Terminal, Plan, Preview, and Console log tabs, which can also be switched to full screen
Browser view (a live view of the AI’s browser operations) is coming soon.
The agent’s configuration, from the chat
A chat bound to an agent carries a “Configuration” tab in the rail, alongside Files and the Console log. It opens the same configuration panel as the agent’s own page — model, skills, knowledge, connectors — so you can check or adjust a setting without leaving the conversation.- The rail keeps one pane open at a time: opening Configuration closes Files or the Console log, and the other way round
- Clicking a skill or knowledge chip inside the panel drills into that item’s details in place
- A note sits at the top of the panel: “Config changes take effect in new chats, not this conversation (knowledge is picked up immediately).”
- An ad-hoc chat has no agent to configure, so it does not get the tab
The preview toolbar
Everything that opens in the preview — a deliverable, a knowledge file, a flow’s output file — carries the same toolbar. Each format contributes only the controls its viewer supports, so a plain text file shows far fewer of them than a PDF.
PDFs open in the app’s own viewer, with zoom, paging, in-document search, and text selection. Mermaid files (
.mmd / .mermaid) render as diagrams rather than as their source — the same viewer chat messages use for diagrams, so zoom, pan, and export work on them too. draw.io files (.drawio, and an .xml file whose content is a draw.io diagram) render as diagrams as well: a file with several pages is paged from the toolbar, Ctrl (⌘ on Mac) + scroll or a trackpad pinch zooms up to 400%, and the download menu exports the page on screen as SVG. PowerPoint (.pptx) and Word (.docx) previews also zoom with a trackpad pinch or Ctrl (⌘ on Mac) + scroll, from 10% to 400% — plain scrolling keeps behaving as before (in a pptx, it moves through the slides). In an Excel (.xlsx) preview, “Search in document” (or Ctrl / ⌘ + F) runs across every sheet in the workbook, not just the one on screen — jumping to a hit switches to that sheet and selects the cell; a large workbook shows “Searching…” while it scans and can report a “Partial search” until it finishes. Legacy Word and Excel files (.doc / .xls) open in the same viewers as their modern counterparts; PowerPoint 97-2003 (.ppt) is still download-only.
A .csv or .tsv file is shown as a table with a sticky header row and row numbers — the delimiter follows the file extension, and the first row is always read as the header. Only the first 2,000 rows are drawn (fewer for a very wide file), and a notice such as “Showing the first 2,000 of 5,400 rows.” tells you when the file goes on. The table has no zoom, paging, or search controls of its own; “Select & ask” (below) is the one toolbar control it adds.
Select & ask
“Select & ask” in the toolbar lets you point the AI at one part of what you are looking at instead of describing it. Press it (it reads “Selecting…” while armed) and a band under the toolbar tells you what to do in that format — “Drag around the part you want the AI to look at” on an HTML page, for example. Make a selection, and a context card lands above the composer — a chip reading “Selected: …” — to be sent ahead of your next message. Whenever the area can be captured, the card also wears a picture of what you selected as its background, and that picture is attached to your message alongside the quoted text; if it cannot be captured, the text goes on its own. Type what you want done with that part and send. Remove a card with its ✕ (“Remove pasted text”). The mode stays armed after a selection, so you can point at several places in a row; leave it with the ✕ at the end of the band (“Exit Select & ask”) or by pressing the toolbar button again. While it is armed, the toolbar’s download, save, and expand controls are dimmed and cannot be used.- HTML — drag a rectangle over the rendered page (an HTML file you uploaded or attached works the same way as one the AI wrote). The text blocks the rectangle touches (headings, paragraphs, table rows) are named in the card as “[Selected area: index.html: …]”. A plain click selects nothing
- PDF — drag over text to quote it; the AI gets the text, stamped with the page (“[Selected area: report.pdf p.3: …]”), and a picture of the passage. A drag that starts on blank paper draws a rectangle instead, and that area is sent as a picture with its page number, without the text — use it for a chart or a scanned page
- PowerPoint — click a slide to quote all of it, or drag over part of it to quote just that part; the card names the slide (“slide 4”)
- Word — select text as you would in any document. The paragraphs your selection touches are quoted from their start (a table row as one entry; a very long one is cut short), up to 40 of them, and the card is captioned with the heading they sit under when there is one
- Excel — drag over cells. The values are quoted as a small table with column letters and row numbers, stamped with the range (“[Selected area: sales.xlsx Sheet1!B2:D10: …]”). At most 40 rows by 20 columns are read, from the top-left of the selection, and long values are cut short
- CSV / TSV — select rows. Each row you touch is quoted from its first cell (long cells and long rows are cut short), with the header row ahead of them and the row span as the stamp (“rows 3–7”), up to 40 rows
- draw.io — click a shape, or drag around several. Their labels are quoted (a connector as “A → B”), stamped with the page name when the file has more than one page
Receiving generated files
When the AI creates files, they are listed in the message under a “Deliverables” heading, one row per file: the format icon, the file name (a well-known extension is dropped from the name — the full name shows on hover), and a second line reading kind · format · size, such as “Document · PDF · 2.7 KB”. Click anywhere on the row to open the file in the preview; the bookmark saves it to Knowledge or Notion (below) and “Download” saves it to your computer. Word (.docx / .doc), Excel (.xlsx / .xls), PowerPoint (.pptx), and PDF files render as previews; a format with no viewer (.ppt, archives, and the like) opens the preview on a download link instead. Images appear as thumbnails rather than rows — click one to open it full size.
An HTML file the AI creates runs live in the preview — buttons, page navigation, and other scripts work as they would in a browser. It opens behind a “Preview” / “Code” toggle in the panel’s header, follows the AI’s edits while the reply is still being written, and the closed panel opens onto it by itself the first time the AI writes such a file; a page the sandbox cannot run (module scripts, TypeScript sources) opens on the Code tab instead. This live view has the same “Expand preview” control as other previews, available as soon as there is a page on screen — including while the AI is still writing it; the control is dimmed only while the file’s bytes are not on hand (still loading, failed to load, or too large to show), when it reads “Expand preview (available once the file is shown)”. The expanded dialog shows the same page as the panel and follows the AI’s edits, keeps the Preview / Code switch and Select & ask, and has its own tab choice — switching the dialog to Code leaves the panel on Preview. Every other HTML file opened in the preview — one you uploaded or attached, a Knowledge file, a reference the AI read, a file on a flow-run card — runs its scripts too, in the same sandbox: the page has no access to your cookies or app data and cannot fetch data over the network while it runs, so a dashboard that loads its data at runtime stays empty, while the script libraries, stylesheets, fonts, and images it links to still load. Only the Preview / Code toggle is reserved for HTML the AI wrote in this chat; Select & ask works on any HTML file in the preview.
Files the AI reads from Google Drive
When the AI reads a file from Google Drive during a chat, a copy is kept with the chat so you can see exactly what it read. These files appear in the message under a “References” heading — separate from Deliverables — and in the Files pane under the “References” filter, and they open in the preview like any other file.- Google-native formats are converted for preview: Docs open as Word (.docx), Sheets as Excel (.xlsx), Slides as PowerPoint (.pptx), and Drawings as PDF. Regular files such as PDFs, images, and Office documents are kept as they are. Files on shared drives are covered too
- The copy is made with your own Google Drive access, and files over 25 MB are not copied — the AI’s answer itself is unaffected either way
- A reference can be saved to Knowledge from its row, just like a deliverable. Reading a file never opens the workspace panel on its own, and references are not offered as file inputs to flow runs
Skills the AI referenced
When a reply drew on one of the agent’s skills, a “Referenced skills” row appears at the foot of the reply, next to the “Referenced documents” row that lists the knowledge it read, with one chip per skill. Click a chip to open that skill — in an agent chat it opens in the agent’s configuration panel, elsewhere as a read-only skill view. The step in the activity timeline that loaded the skill is named after it too: “Referenced the skill ”…"". The row records that the skill was read, not that every instruction in it was followed, and the built-in document engines (PowerPoint, Word, Excel, PDF) never appear in it.Knowledge files the AI points you at
When the AI refers to a file in the agent’s knowledge, it renders it in the reply as a clickable card labelled “Knowledge,” carrying the item’s title and type icon. Click the card to open the original in the preview; items with no file behind them, such as a URL or a Notion page, open their knowledge details instead. Ask for something explicitly — “open the price list” — and the preview opens on its own. Only knowledge the agent is allowed to read becomes a card, so a card never points at material you cannot open.Pages the AI links to
When the AI points you at a screen — the Connectors page, an agent, a skill — it writes a clickable link to that page in this workspace, so following it lands you in the right place. Screens your workspace cannot open are never linked.Saving generated files to Knowledge or Notion
Deliverables can be saved to a permanent home without downloading and re-uploading. Click the bookmark icon on a deliverable card (or “Save…” in the preview toolbar — use this for images, which have no card bookmark) and choose a destination:- Knowledge — save the file as a knowledge source, under “Workspace knowledge” or “Agent knowledge.” The agent option carries its own picker offering every agent you can edit — the chat’s own agent is preselected, but you can target any of them, and the option appears in agent-less chats too. For workspace knowledge, pick the destination with the built-in folder browser (the folder you are currently viewing is where the file is saved)
- Notion — export the file to a Notion page you search for and select (requires the Notion personal connector; if it is not connected, the menu shows a Connect link)
When a run takes a long time
If a tool run through a Connector is slow to respond, a hint message is shown. If a jinba-flow integration leaves no run history on the jinba-flow side, check that the MCP URL is reachable from outside (Anthropic’s cloud) —localhost and internal-only hostnames cannot be reached.