Data & files

Your work needs data going in and produces files coming out. Both are managed from the Files tab in the sidebar, which shows four sections: Notebooks, Deliverables, Datasets, and Workspace.

Those last two are not the same place, and the difference matters more than it sounds:

• Datasets is project storage. It's where an upload lands. It is not the sandbox filesystem, so no cell can open a file from it directly.
• Workspace is the sandbox filesystem (/home/user), the kernel's working directory. This is the only filesystem your code can see.

A file has to be copied from storage onto the sandbox disk before pd.read_csv can read it. That copy is called staging, and most of the time the agent does it for you.

Uploading data

There are two ways in, and they both put the file in the same place.

Attach it to a chat message. Drop a file onto the composer, or use the attach button. The file goes to project storage, and then Clusy stages it onto the sandbox disk automatically, before the agent's first token. The agent is told exactly where it landed, so this just works:

Load transactions.parquet and show me the date range and row count.

It reads it by bare name, from the kernel's working directory. No path, no hunting.

Upload it in the Files tab. The file goes to the same project storage and appears under Datasets, but it is not copied into Workspace automatically. The agent copies it into Workspace when you ask it to use the file. If you upload a dataset and then write a cell that opens it by name yourself, you'll get a FileNotFoundError and wonder why. Ask the agent to use the file, or attach it to a message instead.

Attaching and uploading both land in the same project storage, which no cell can open. Attaching to a message also copies the file onto the sandbox disk — where pd.read_csv finds it by bare name — while a file that was only uploaded waits in storage until the agent stages it.

Use the Files tab for bulk uploads and whole folders. Use an attachment when you're about to talk about the file anyway.

Either way, uploading beats pasting a sample into chat for anything sizeable. The code reads the real file, and you only move the data once.

Dropping an .ipynb file into the chat's file manager is a special case: instead of landing as a plain file, it's imported as a new notebook, opened in the canvas and ready to run.

Find a dataset

You don't have to bring your own. Ask for the data you need and the agent goes looking:

Find me a public dataset of daily bike-share trips with weather, and load it.

It searches the Hugging Face Hub and Kaggle's public datasets, ranks the candidates by how well they fit what you asked for, and tells you what it found. Pick one and it writes the cell that pulls it into the workspace, then loads it and shows you the shape. Neither hub needs to be connected for this: it's public data, read anonymously. Connecting Hugging Face is what lets it reach your own private repos.

This is a search, not a magic trick: it finds real public datasets and it will tell you when nothing suitable exists. If you want a specific one, name it.

Browsing the workspace

The Workspace section of the Files tab shows what's actually on the sandbox disk: staged datasets, files your code or the agent wrote, and intermediate artifacts. You can browse the tree and download anything. When the project is connected to GitHub, this view also reflects each file's git status.

Clicking a file previews it in the canvas: tabular files like CSVs open as a scrollable spreadsheet-style table, and documents open in a preview tab, so you can check a file without downloading it or writing a cell to read it. Files under Datasets preview the same way.

Workspace files are saved when the sandbox shuts down and restored when it comes back, so a file you wrote yesterday is still there today. See Runtimes for what does and doesn't survive a restart.

What's in memory

There's no variable inspector. To find out what the kernel is holding, ask:

What's in memory right now, and how big is df?

The agent runs a cell and tells you. Same for a quick shape check, a dtypes dump, or a peek at a fitted model's parameters.

Getting output back out

Anything the notebook produces (a cleaned dataset, a model file, a chart, a report) is a file you can download from the Files tab. The Deliverables section collects the things the agent flagged as finished work, so you don't have to remember which cell wrote which file.

From there it can go further: commit it to GitHub, or publish a notebook to Hugging Face or Kaggle. You can also download the whole notebook as .ipynb, output included.

Uploads and workspace files do not count toward the plan's saved-state volume. That volume is only for kernel checkpoints—the variables used to resume sessions—and is managed under Settings → Storage.

Files still have operational boundaries. Files-tab uploads have no additional Clusy per-file cap; chat and Home attachments accept up to 100 MiB per file. A workspace boundary captures up to 1 GiB per file, 5,000 files, and 10 GiB total per project, with files over 100 MiB restored on demand. See Limits.