Design
TriggerToolModelKnowledgeConditionAgent
Sign-in, sign-up and passkeys are live. What's new
What it is
Our approach
Every agent has a job, an autonomy level and an accountable person. You steer it the way you steer your team.
Triggers, steps and branches drawn with the mouse. A flow comes together in one sitting and runs the same way every time after that.
Like any company: new people ask first, proven ones decide alone. You raise the level — and the agent acts accordingly everywhere, immediately.
The same case takes the same path. The model judges where judgement belongs; the process sets the flow. That is what makes delegation real.
Irreversible steps wait for your decision. Everything before them keeps running while you work on something else.
Connections
Six model providers, your own systems over MCP, encrypted credentials. The inner ring runs today; the outer one arrives with the cloud.
With Ollama every model runs locally on your machine, offline.
OpenAI, Anthropic, Google and Mistral directly. Or through your own Azure instance in your region.
Enter an address and the platform reads which actions exist — classifying each one.
AES-256-GCM, its own data key per entry. The last four characters are shown.
Reading steps run straight through. Irreversible ones wait for a decision.
One level per agent. Raising it changes behaviour everywhere, and it lands in the log.
Every run in full: input, branch, decision, reasoning. The question why an agent did something has an answer.
Your model, your systems, your infrastructure. Switch providers and the flow stays as it is.
Answers
Every action of a connected system is classified: reads only, changes data, or irreversible. Every agent has an autonomy level from 1 to 5. Together they determine whether a step runs or stops. To dissolve a stop you raise the agent's level, and that is exactly what the log then shows.
Only if you set it up that way. With Ollama every model runs locally on your machine, offline. If you want an external provider you configure one — or take Azure OpenAI in your own instance and region. The platform itself runs wherever you put it.
The run pauses on the server. You can close the tab, switch devices or go home. The approval sits in the inbox of the responsible person — not behind a link you have to possess. After the decision the case continues exactly where it stopped.
No, and that is deliberate. The agent works, you decide — including runs you started yourself. Whoever deploys an agent does not want a second human for every approval. Where dual control is required it can be switched on per organisation, most sensibly for irreversible actions only.
Over MCP: you enter a server address and the platform asks which tools it offers. Each action is classified in the process, and values in the call can sharpen that classification — an amount above a threshold turns “changes data” into “irreversible”. A system is described once, then it stands.
With a process somebody does by hand today. You draw it on the canvas, connect the system it touches, and put the agent on a low autonomy level — then it asks before every step. Once it has run for a few weeks and the decisions hold up, you raise the level. That is the whole path.
A configured installation you can operate right away. It runs on your machine.