The Assistant: talk to your app
Talk to your app: open, search, count, change. With approval per action the way Claude Code does it: every write is proposed as a card that a human answers.
Every Kumiko app already carries a precise description of itself: entities, fields, handlers, screens, navs, labels. ai-agent opens a chat layer on top of the running app and hands the model exactly that description. That gives you a second way to operate the app you already have.
- Questions become queries: “how many open incidents does the API have?” runs through the same query handlers a click would take, with the identity of the asking user, never a system user
- Changes become proposals: in the default mode a write can only become a card with four answers: run, check in the form, always allow, drop. “Check in the form” opens the app’s own edit screen, pre-filled with the proposed values
- The catalog belongs to the caller: it is built from the registry per request, filtered by the caller’s roles. A handler the user cannot reach is never offered to the model
- Every executed tool call is audited and lands alongside the audit entry of the handler it invoked
The screenshot comes from PublicStatus, an ordinary Kumiko app with the agent mounted next to its incident and component features, without special integration. Its UI runs in German.
ai-agent and its ai-agent-edit add-on ship in the commercial Kumiko Enterprise offering; the exposure rules they build on (agent-tools) live in the open framework. How to make your own app agent-ready is in the Make your app agent-ready guide.