LEFT RUNNING

How it works

There are three moving parts: a board it reads, a key you press, and two decisions it always gives back to you.

1. The board is the instruction

You do not write prompts. Left Running keeps a task board — yours to read and reorder — and takes the top item that it can actually finish. When the board runs dry it does not stop; it goes looking for what should be on it: a broken page, an unshipped asset, the biggest gap between where a number is and where you said it should be.

Everything it does lands back on that board as a finished card with the proof attached: the commit, the live URL, the number before and after. If you read nothing else, you can read the board.

2. One key decides how long

p
Run. Leg after leg until you say stop. This is the overnight key.
d
One leg. The single best next action, verified, then it waits for you.
q
Land. Finish what's open, ship it, confirm it's live, stop cleanly. Nothing left half-done.
O
Operator. Your queue — only the things it is not allowed to do alone.

The reason duration is a keypress and not a guess is deliberate: the machine should never decide on its own whether you are watching. d means you're at the desk. p means you've gone to bed.

3. Two gates it will not cross

It has broad permission to do reversible work — write code, commit, deploy, edit configuration, read your analytics — because all of that can be undone in a minute. It stops, every time, at two things that cannot:

Alongside those it will not type your passwords or two-factor codes, will not make the final create-click on a new account, and prefers the reversible path when both exist. When it hits one of these it does not stall — it prepares the whole thing down to the last click, files it as a card for you, and carries on with the next piece of work.

What a night actually looks like

A leg is one verifiable outcome. It repeats until you stop it:

  1. Pick. Highest-value item it can finish, weighted toward the parts of the business that already earn.
  2. Do. Write the code, run the build, deploy.
  3. Verify on the real thing. Fetch the live page and confirm the change is actually being served — not the local copy, not the exit code.
  4. Report. One line: what changed, where, and the check that proves it.
  5. Learn. Log what it expected against what happened, so the next pick is better than the last.

In the morning you read a list of one-line results and a board that is more true than it was. The work that needed you is waiting under O.

What it runs on

Left Running runs on your own Mac, against your own repositories and hosting accounts. Nothing executes on our servers, and we never hold your deployment credentials. You'll need a Mac you can leave awake and a Claude Code subscription of your own — the reasoning is billed by Anthropic to you, not resold by us. We're explicit about that on the pricing page.