Watch it done once, then do it yourself

Steps

1. Download the installer and check it against the SHA-256 shown on the download page.
2. Run it on the Windows PC that will do the work. Harness runs there, not in a cloud.
3. Open Harness and paste the licence key from your purchase email.
4. Activation ties that licence to this one machine. A second PC needs its own licence.

If activation fails, check that the PC can reach the internet for the licence check, then try again.

Steps

1. Open settings and choose your provider — Claude or GPT.
2. Paste your own account credentials or API key. You pay that provider directly, at their price; we add no markup.
3. Send a test task to confirm the connection works.

When a task needs the model, the relevant task content goes to that provider under their terms. That is the one place task content leaves your machine, and you decide when.

Steps

1. Start from default-deny: out of the box the worker can do nothing.
2. Turn on one capability at a time — a folder, a mailbox, a system — for the routine you are building.
3. Mark the sensitive ones (sending, spending, deleting, signing) as requiring approval.
4. Review the list before you run anything. If a capability is off, there is nothing to jailbreak into.

Credentials are held outside the model context — the AI signs in through a broker by name and never sees the passwords or keys themselves.

Steps

1. Pick one routine your team already repeats and that has a right way of being done.
2. Describe it in plain language: the trigger, the sources it may read, and what the finished result should look like.
3. Harness restates the scope and goal. Correct it before anything runs.
4. Run it once, end to end, and check the result against how you would have done it by hand.

Steps

1. When a routine reaches a sensitive action, it stops and queues it for approval — even if it is running overnight.
2. Open the approval queue, see exactly what is about to happen, then approve or reject it.
3. Afterwards, open the run record: what was read, which tools were used, what was queued and what was done.

The record is written in plain language on your machine, so an auditor or a colleague can follow it months later.

Steps

1. Give the routine a schedule — every Friday, month-end, or whatever the job needs.
2. Use the operational view to see what is queued, running, waiting for approval and finished.
3. If a step fails or a source is briefly unavailable, Harness retries; if the machine restarts, scheduled routines resume.

A routine should never quietly stop without you knowing — that is what this view is for.