There is no implementation phase, no solutions engineer and no statement of work. You sign up, configure an agent in plain English, and run it.
Email and password. That is the whole form.
There is no email verification step, no card, no sales call and no onboarding questionnaire. You are in the dashboard in about ten seconds with $5.00 of credit already on the balance, which is enough to run real work rather than a toy demo.
Eight agents, each doing one job properly.
Browse the catalogue and read what each one actually does, what it takes as input and what it returns. Every agent page is explicit about what it is bad at as well as what it is good at, because choosing the wrong agent is the most expensive mistake you can make here and it costs us nothing to prevent it.
Plain language, not a rules engine.
Tell the agent what it needs to know about you: your ticket categories, your ideal customer profile, the schema you want extracted, the tone you write in. You write this once in normal prose. When it changes, you edit a paragraph rather than rebuilding a decision tree.
From the dashboard, or from your own code.
Trigger a run from the dashboard to see what comes back, then move it into your stack when you are happy. Every run returns structured output plus a confidence score, and every run is logged with its inputs, duration, token count and exact cost.
The agent proposes, your system disposes.
Agents return data, not actions. You decide what happens with it: apply the classification, send the draft, write the row. This keeps the blast radius of a mistake at the size of a bad suggestion rather than a bad action taken on your behalf.
Per run, to the fraction of a cent.
Your balance moves as runs complete, and the dashboard shows spend per day and per agent. Failed runs are not charged. If a month is quiet you pay almost nothing, because there is no subscription underneath waiting to bill you anyway.
A run is one complete unit of work: one ticket classified, one document extracted, one brief written. It is the billing unit and the audit unit, and everything in the product is organised around it.
When you trigger a run, the agent receives your configuration plus that run's input. It reasons, validates its own output against the expected shape, and returns structured data with a confidence score. The whole thing is logged: inputs, outputs, token count, duration and exact cost.
If the output fails validation, the run is retried. If it fails permanently, the run is marked failed, you are told why, and you are not charged. We would rather absorb the cost of our own failures than invoice you for them.
Because a chat interface makes work hard to audit and impossible to price honestly. A discrete run has a beginning, an end, a cost and a record. You can point at it a month later and say exactly what it did and what it charged you.
Pick an agent, paste your input, hit run. You see the trace, the output and the cost. This is how everyone starts, and for low-volume work it is often where people stay.
One POST per run. The response is the same structured payload the dashboard shows you, so what you tested is what you get.