Skip to content
Steptree

Where agents run

In the cloud when they can. On a real desktop when they must.

Some jobs genuinely need a screen. Steptree reads the procedure, tells you which one it needs, and refuses to deploy it to the wrong half.
Start freeWhat an agent can do

The two ways it runs

With nobody there, or with you watching

The curve of the Earth at night, city lights scattered across the dark side.

Cloud runtime

The agent runs on our servers, whether or not anybody is at a desk.

  • Best for

    Anything your software can be reached over the internet for.

  • Starts when

    A schedule comes round, an email or a form arrives, a record changes in a connected system, or somebody asks in Slack.

  • Your part

    Nothing. You read what it did in the morning.

A hand pointing at something on a laptop screen while another hand rests on the keyboard.

Desktop runtime

The agent gets the keyboard and mouse on your own computer.

  • Best for

    Jobs that need a real screen: your in-house system, a file on your own machine, software nobody is ever going to replace.

  • Starts when

    You ask it to in the chat window, or a cloud agent hands it a step it cannot do alone.

  • Your part

    You watch. It asks before each action, and you can stop it mid-step.

You do not have to work out which one a job needs. Steptree reads the recording and tells you — and because the agent runs from the same approved document either way, that choice is never a rewrite. If the job depends on a real screen, it refuses to publish to the cloud and points at the exact step that needs a machine.

On your own machine

What it is like to hand over the keyboard

Handing over the keyboard is the part that needs the most safeguards, and this section covers them. You approve, you watch, and you can stop it mid-step.
It gets the keyboard, the mouse and the screen
Not a browser script and not an API wrapper. The agent works your applications the way a person does. That is why it reaches software that was never on the web to begin with, including the in-house system nobody is ever going to replace.
An approval card in front of the action
Each action is described before it happens: what it is about to click, type or run. You approve, or you don't. There is an auto-accept mode for the safe ones, and it is deliberately narrow. Reading a file full of credentials still comes back to you, because that is the one read nobody can undo.
It cannot act on a stale look at the screen
If the window moved since the agent last looked, the click is refused, and the refusal names the window that is in front now. Guessing here is how automation types a password into the wrong box.
You can stop it mid-step
Stop means stop: the request is aborted, the machine is handed back, and the run is recorded as stopped rather than quietly finished.
Several conversations, one machine
Only one run can hold the screen at a time, so a second run reaching for the keyboard waits and says so, instead of fighting over the mouse. Work that does not need the screen carries on beside it: a build, a file, a spreadsheet.
It can show you the file, not just describe it
A spreadsheet as a real grid, a document as pages, a deck as slides, a PDF, an image, all beside the chat. The agent can point at the exact row or heading it means. It can also browse a real website in the same pane, signed in as you, without taking over your screen.

Steptree refuses instead of deploying blind

Anyone can build a cloud agent and let it fail on the first screen it cannot see. Steptree analyses the approved procedure, and when a step needs a desktop it refuses the cloud deployment and names the skill and the step. Knowing in advance what it will refuse is what lets you plan around it.

Questions

About the runtimes

Our core system is ancient and has no API. Does that rule us out?
That is the case Steptree is built for. In the desktop runtime an agent gets a real keyboard, mouse and screen. So it works with software that has no other way in, including systems nobody has touched in twenty years. Steptree also tells you up front when a procedure needs a real screen like that, instead of publishing something to the cloud that could never have worked.
How do I know which runtime a process needs?
You do not have to work it out. Steptree reads the recording, recognises the applications involved, and states whether the work can be done in the cloud or needs a desktop — before you deploy anything. A cloud publish that needs a screen is refused with the step quoted.
Does the desktop agent need me sitting there?
It needs the machine, and by default it needs you to approve actions. That is the trade: a desktop run is supervised work on your own computer, and a cloud run is unattended work on ours. A cloud agent can also hand a single step to a desktop and carry on afterwards.
What does the cloud runtime actually run on?
Steptree's own infrastructure, multi-tenant, with no database credential of its own — the runner asks the app for the agent when a session starts. Each customer's data stays separated inside the database rather than by the runner remembering which tenant it is serving.
Can we keep everything inside our own network?
No. On Team and above, Steptree AI's chat in the Steptree App can use a model on your own hardware. That conversation then goes to your machines rather than to a hosted model. The app still signs in to Steptree, recordings are uploaded, and cloud agents run on our servers. There is no choice of data region today.
What happens if the desktop machine is asleep?
A desktop run needs the machine awake and the app running — it is your computer, not Steptree's. Anything that has to run unattended belongs in the cloud instead, which is why Steptree tells you which half a procedure falls into before you rely on it.

Next

Where to go from here

  • The two apps

    The Steptree App, the Recording App and the phone — what each is for.

  • What agents can reach

    Channels, connected systems, and the applications that can only be reached through a screen.

  • Security

    What is never recorded, how workspaces are separated, and what a local model covers.

Do one real job once, and read the SOP Steptree writes

Record it on Windows or Mac. Steptree writes the SOP with the goal, the steps and the judgement calls. Approve it, share it with your team, and connect your AI. Free to start, no card.

Start freeCompare plans
Someone at a wall of sticky notes talking a seated team through a process.