Skip to content

2 min readMemtomo

What an automation is, and why it runs in your browser

An AI-written report that re-runs on your tables every time you run it. The design, the trade-offs, and what it cannot do yet.

“Make me a weekly spend report by category” is a request for a program. Not a big one — a query, a group-by, a chart — but a program that should run again next week without anyone rewriting it. That is what an automation is.

A report the AI writes and your browser runs

  1. Make me a weekly spend report by category, with a chart

  2. Created “Weekly spend”. It re-runs on your data every time you run it — nothing leaves your browser.

    Weekly spend

    Ran in your browser

    This week

    €386

    −12% vs last week

    Groceries
    €142
    Restaurants
    €96
    Transport
    €58
    Subscriptions
    €47
    Other
    €43

The shape

An automation is a short program made of steps. Each step can query your tables, do the maths and display a headline number, a paragraph, a table or a chart. The assistant writes it from your description, after checking the real table and column names, and validates it — that it parses, that it is well-formed, that every name it uses is defined — before it is allowed to save. A broken automation never lands.

Why the browser

The obvious design is a server that runs automations. We did not build that, for three reasons.

Isolation. Code that runs on your data should run where you are. The automation executes in a sandbox inside your tab. It reads tables through your own signed-in session, with your permissions, and nothing else. There is no shared executor to escape from.

Cost. Running every user’s every report on our servers would be the single largest line in the bill, and you would pay for it. Your laptop is already on.

Latency. A report over a few thousand rows runs in well under a second locally. There is no queue.

What it costs you

The first run starts the sandbox — a few seconds and some megabytes, cached afterwards. Very large tables are aggregated by the query engine before they reach the browser; the automation sees the result, not the raw rows. And an automation cannot run while your browser is closed, which is why scheduled reports — “send me this every Monday” — are the next thing we are building: the same automation executed in the same kind of sandbox on a server, with a per-run token that can only read one project.

Embedding

An automation is referenced by a wiki link, [[automation:…]], so it can live inside a note. Open the note and you see the last run; open the automation and it runs again. Re-running on open is next. That is the whole dashboard feature: notes with live pieces in them.

Related

← All posts