Files

A file browser for agent workspaces. Read, edit, download and delete files, including the identity and memory markdown files that shape the agent's behaviour. Deleting through the agent's own tools moves files to a recoverable trash/ folder rather than destroying them.

The tree lists every agent you have access to, each as a collapsed row with a workspace folder under it, so you browse the whole team from one place without switching the agent picked elsewhere in the dashboard. Nothing is loaded until you open it: expand an agent, then its workspace, then the folders you want — each one lists on demand. Right-click an agent or its workspace row to create a file or folder there or refresh it; the toolbar's New file / New folder buttons land wherever the row you last clicked lives. The footer names the agent whose files you are looking at. Links from other tabs (a Cortex item's files, a skill's folder) open straight to the file under its agent.

Files an agent produces. Every agent keeps one layout: reports/ holds everything it makes for you (charts, images, PDFs, spreadsheets, documents, exports — one canonical file per deliverable, revised in place), scratch/ holds its scripts and intermediate files, and uploads/ holds what you sent it in chat or on a messenger. When an agent names a file in its reply — its own Saved: line, a path in backticks, or a full path — the chat shows it: images, audio and video render inline, and every other file appears as a card with Download and Open in Files, which jumps to that file in this tab. The same works when you reopen an old conversation. On a messaging channel the agent's Saved: line attaches the file to the chat instead. Members see and download the files of the agents they have access to; editing and deleting stay with admins.

Only the agent's finished answer attaches files, and only the files it names there. A path that came up while it was working — a file it opened to read, a scratch file, a path quoted out of some document — is not attached, so you get the deliverables and not everything the agent touched on the way. If it names the same file more than once in one answer, you still get one card. Agents are told to list every file they made or changed for you, with its full path, so if one leaves something out, say so — and if you would rather it not attach a particular file, tell it that too.

What an agent can do with documents. Every box now carries the tools an office document actually needs, so an agent does not have to explain that it cannot: it can turn a Word file, a spreadsheet or a slide deck into a PDF (and open a legacy .doc), read a scanned PDF or photographed page with OCR, merge, split, rotate and unlock PDFs, and check its own work by looking at the pages it produced. Two limits worth knowing: OCR reads English — for another language the agent reads the page with its vision instead, which it will say — and a converted document uses substitute fonts for any typeface the box does not have, so the common ones (Arial, Times New Roman, Courier New) keep their layout and an unusual one may shift. Ask for the result you want ("turn this into a PDF", "what does this scan say") rather than naming a tool.

If an agent runs on your own machine rather than a hosted box, whoever installed Olano could have declined these packages — the agent checks before it relies on one, and tells you what it did instead.

The tree is fully keyboard-driven: ↑/↓ move, → opens a folder or steps into it, ← closes one or steps back to its parent, Home/End jump, Enter opens, F2 renames, ⌘⌫ / Ctrl+⌫ deletes, and typing a few letters jumps to the next row that starts with them. Very large folders show their first 2,000 entries and say how many more there are — use the filter box to narrow. A very large text file previews its beginning, says so, and offers the download for the rest; editing is switched off for it so a save can never lose the part you were not shown.

The Filter box searches by name across every workspace you can access, not just the folders you have opened: type memor and each agent's MEMORY.md appears under its agent row even with everything collapsed. Matching is case-insensitive and looks at the file or folder name and at its path, so memories/2026 finds a daily log by its folder. It starts from two characters, waits for you to pause typing, and searches everything at once — press Enter to search without the pause, Escape to clear. Version-control, dependency and cache folders are skipped, and dot-files only take part while the eye button shows them. Click a match to open it and reveal it in the tree; when there are more matches than it shows, it says so — keep typing to narrow.

Markdown and HTML files open in a rendered preview, with a Preview / Raw toggle in the preview header to switch to the source. An HTML page renders live in an isolated frame: stylesheets, scripts and images it references by relative path are pulled from the same workspace folder, so a small prototype site looks the way a browser would show it. The frame is sandboxed — the page runs cut off from your dashboard session and cannot read or change anything outside its own preview.

On a phone the tree and the preview are not side by side — whichever you are using gets the whole screen. The tree is what you browse; opening a file hands the screen to the preview, and the All files button at the left of its header hands it back. The Preview / Raw toggle and the file actions sit on their own row under the file's name, so the name stays readable however long it is.

Files whose name starts with a dot — .env, .lionclaw/, .git/ — are left out of the tree by default, so a workspace opens on the files you work in. The eye button in the tree toolbar shows them; click it again to hide them. The button lights up while they are showing, and your choice is remembered the next time you open the dashboard. It only changes what the tree lists: a hidden file you open, edit or delete behaves like any other file, and a link that points at one turns the toggle on for you.