What adopters actually live through: from launching agents blind and at uncontrolled cost, to watching them work by department, approving only what has effects, with budget and audit — and having that control compound as capital.
highlighted step = the moment of governance (human decision)
One command. Zero dependencies (Python only). You have it running in 2 minutes.
With /setup or a template, define your specialists and departments.
Bring up the broker and plug in any LLM.
Each agent is a live card, grouped by department (lanes), with its tokens and cost in €.
Every operation with effects stops at Needs Input and waits for your Approve or Deny. The decision is made by a policy in code, not by the model.
Budgets in €: local barely counts; the expensive model that goes over falls into «⚡ No budget».
Every decision lands in a verifiable chain, tamper-proof.
The residue becomes an asset: captured human judgment, data for evals, and cost per task trending down — your moat grows.