Work that runs itself
and never loses its place.
Describe what should happen and the AI builds it. It remembers every step it has finished, so a crash, a retry, or a week spent waiting on someone's approval never starts the work over or sends the same email twice.
Four ways to start a run
An automation is a 'when this happens, do that'. Point it at a time, an incoming request, something happening in your workspace, or an app you already use, and it runs on its own, every time.
On a schedule
Run every hour, every morning, or on whatever schedule you set. The schedule fires once, even if two of our machines see it at the same moment, and there is no server for you to keep alive.
On a webhook
Give an outside system its own address, so a payment, a form, or a finished build can start a run. Every call is checked to be sure it really came from them, and that it isn't a replay of an older one.
On something in your workspace
When something happens in one of your apps, every automation waiting on it starts. This is how your apps, agents, and automations work together.
On a connected app
Start on something real happening in an app you use: a new email, a calendar invite, a new customer, across 1,050+ apps. Your keys never go anywhere near the code.
What a run looks like.
It can't lose the thread
This isn't a setting you switch on, and it isn't one you can forget. It's how every automation runs: the work remembers each step it has finished, so it always picks up exactly where it stopped.
It remembers what it finished
Once a step is done, it's done. A retry, a new release, or a crash picks up at the first unfinished step, so a charge or an email never goes out twice.
Retries that don't double up
When something fails for a moment, it waits and tries again on its own. Only the unfinished part is redone, so anything that already worked never happens a second time.
Wait for hours or days
Sleep until a date, or stop mid-run to wait for something to happen, for up to a week. Waiting costs nothing, and when it arrives the run carries on from the exact step it paused on.
Wait for your approval
Put a person in front of anything sensitive. The run pauses and tells the owner by email or right in the channel. Approve or reject and it carries on, or it times out on its own.
Anyone can tell what it does
An automation shouldn't be a black box only the person who built it understands. Every one comes with a plain summary, a picture of what it does, and a simple form to run it. You never have to read the code to operate it.
A plain summary
A readable description of what the automation actually does, so anyone can tell what it touches before they run it.
See the whole flow
What starts it, which apps it uses, and what it calls, laid out as a diagram you can read like a story.
Run it from a form
Run or test it from a form built for it automatically, with the right fields, already checked for you.
Nothing runs you can't see or price
Every run is recorded from start to finish, with what it did and what it really cost, so you always know what happened.
Read any run
Open any run to see its steps and what went in and out, filter by status, trigger, or date, and run it again with the input it used.
See what each run cost
The real cost of a run is shown on its row: the models it used, the apps it called, and a small fixed fee for each step it remembers. Set a budget for a run and it stops before it goes over.
Stop or fix any of them
Pause an automation instantly, turn off a single schedule, retry or cancel a run, act on many at once, and see every automation's health and waiting approvals in one place.
Safe to grant real power
Automations act through your tools and can be called by other systems. The guardrails are already on, not something you have to remember to add.
Secrets stay off the model
Your keys are added on our side, at the moment a run needs them. They never appear in the workflow's inputs, outputs, or logs, and never reach the model. Every call out runs under a policy you set.
Screened before it ships
Every release is checked for risky combinations of powers. The risky ones need you to acknowledge them, and anything unsupported is stopped before it goes live, instead of surprising you later.
Share it with your AI tools
Mark an automation as shared and it becomes a tool that Claude, ChatGPT, and Cursor can call, on one address for your whole workspace. Every key you hand out has its own limits and an off switch.
They work together
One automation can call an app, an agent, or another automation. Before any of it ships, the whole chain is checked, so it's as safe as its parts.
Automate the busywork
Describe the workflow, connect the tools, and let it run. It keeps its place, asks when it matters, and records everything.