Join the waitlist to take part in the alpha.
Let the agents work

ENVIRONMENTS

A disposable workstation, not access to your machines.

Your agents run code in an isolated machine, already tooled for office work, put back to sleep without losing anything.

A machine, not an interpreter

This is not a cut-price sandbox where three lines get evaluated. It is a micro virtual machine, four cores and four gigabytes, with its own file system and its own shell.

No public inbound traffic reaches it.

Four folders structure the work: what comes in, what goes out, reference knowledge, and scratch. The agent does not file things at random.

Already tooled

LibreOffice, pandoc, sqlite, ripgrep, jq. Python with pandas, polars, duckdb, pypdf, python-docx, python-pptx and openpyxl. Node with the docx library. A headless browser, ready to open a page.

The agent installs nothing to convert a document to PDF, open a hundred-thousand-row spreadsheet or read a PDF page by page. It is already there, in the image, and it is the same image for everyone.

It pauses, it does not wipe

When nobody is working any more, the machine is paused with its memory, and it resumes on its own at the next call.

The agent finds its files where it left them. A session picked up the next day does not start from scratch.

An agent that delegates to others shares the same machine as its delegates. They work in the same folder, and the pause waits for the last one to finish.

Your secrets do not enter the model

Your workspace environment variables are encrypted at rest. They are injected into the machine, never handed back with their values.

The model sees the variable names, so it knows they exist and can use them in a command. It never sees their contents: values are redacted from output before it reads them, and redacted before truncation, so no cut can leak a fragment.

Scope is a setting: one variable for the whole workspace, or one variable for a single agent.

How it works

A client sends an eighty-thousand-row spreadsheet and expects a two-page note.

  1. The file lands in the machine's input folder. The agent does not load it into its context, it has no need to.

  2. The agent writes its script: a few lines of Python, with the libraries already present. Nothing to install, nothing to wait for.

  3. It runs it. The computation happens on the machine, not in the model. An aggregate over eighty thousand rows costs a second, not a context window.

  4. It produces the deliverable in the output folder, converted if needed, next to the original spreadsheet.

  5. The file comes back out: stored in your workspace, it becomes an attachment on the conversation and links to the records it concerns.