What Else is

Else is something you delegate to. You describe a piece of work; it makes a plan, does the work, and hands back a result. It stops to check with you before steps that leave the app, and it tells you what it actually did.

That is the whole idea. The age of the operator — the person who drives a tool click by click — is the thing Else is built to end.

What makes it different from a chatbot

A chatbot answers. Else acts.

While a run is in progress, Else shows you the stages it is actually working through — beginning with Planning the run, and adding stages as it goes: Working in a secure browser when it needs to read the live web, Working in the secure workspace when it needs to open files or produce them.

It doesn’t just describe what you should do next. Within one run it can search and cite current sources, open pages and follow links, pull text out of a PDF, read and write Word documents, spreadsheets and slide decks, produce a design or a piece of generative artwork, and write files you can download. Those are the built-in Add-ons — 13 of them, all on. Nothing to install, nothing to switch on per account.

It can also work inside the accounts you already use, once you connect them under Apps — six are first-class (Gmail, Google Drive, Notion, Slack, GitHub and Airtable), and a large directory of other services can be connected by pasting a key. Connect one, then choose exactly what it’s allowed to do.

What it is not

It is not a tool you operate step by step. You do not walk it through a sequence. You state the outcome and the constraints; it works out the steps.

It does not decide for you whether something is a quick exchange or a handed-off job. Typing a sentence and pressing Send always starts a chat. Handing work off to run without you is something you choose, explicitly — see Your first task.

It is not a black box you have to trust blindly. Every run carries a Receipt: what was charged, how long it took, and which engine choice you submitted. Nothing about the outcome is inferred for you.

You stay in control at three points

Before it starts. Each run has its own controls — the panel is headed This run — Just for what you’re about to send, and only what you pick there is available to that run. That includes which of the 13 built-in Add-ons and which connected Apps it may touch, and a spend limit field — $0.15 unless you change it, up to $5.00. Read that last one carefully: on a chat the number you type is recorded and shown back to you, but nothing acts on it. What stops a chat is a fixed per-turn ceiling you cannot see or change. The limit you set is read only on work you hand off, and only between steps. Spending limits is exact about both.

While it runs. Plan before doing is on by default: it shows you the steps first, then gets going. So is the strictest approval setting, Before anything leaves this app — sending, publishing, buying or changing things all wait for you. You can loosen that to Before sending or publishing, letting other reversible changes go ahead on their own, or to Don’t check with me. Sending email is a hard stop regardless of the setting. So is any action in a connected App that isn’t on the approved list — an unknown action asks rather than guesses.

Waiting for you costs nothing. A paused run is genuinely paused: no work is running, nothing is being charged, and there is no expiry. It waits until you answer.

Any time. You can stop a chat mid-run, or stop a handed-off task and ask the same question in a chat instead. Files it already finished are kept. Money already spent stays spent — there is no refund, and we would rather say so than imply otherwise.

How it handles the truth

The honesty behaviour is built in, not a promise about intentions.

• If it has no tools in a given turn, it says so instead of claiming it browsed the web or read your files.
• It never confuses “this tool exists” with “you have connected this account.”
• If it cannot generate a real image, it will not pass a hand-built layout off as one — the caveat is attached mechanically, not left to judgement.
• A scanned PDF with no text layer comes back empty with a warning. It will not invent document content it did not read.
• It will not ask you to raise its own spend limit.
• It will not reach internal or private network addresses.

Real runs do flag their own failures. A completed research task recently closed with “Two things to flag honestly…” and then listed what it had not managed to do. That is the behaviour, not the exception.

Where to go next

Your first task walks through one real delegation end to end — from the sentence you type to the file you download.