AI assistant for business apps
The assistant is for applications where users need to ask about live records and sometimes change them. It uses the app’s own registry and handler permissions, then keeps a person in the loop for writes.
The product flow
Section titled “The product flow”- The user asks a question in the app, by typing, dictating, or attaching a file or photo.
- The server builds a tool catalog from the composed registry and the caller’s access rules.
- The model searches or reads through ordinary Kumiko query handlers.
- A write becomes a plan in plain words: the steps, their proposed values, and one approval.
- The user applies it, asks for a change, opens the normal form for review, allows it to run without asking from now on (never for a high-risk write), or cancels it.
The Assistant walkthrough shows this flow against a real service-desk example. The agent-ready guide shows the handler descriptions and risk declarations the app owns.
Why this belongs inside the app
Section titled “Why this belongs inside the app”| Requirement | Kumiko Enterprise behavior |
|---|---|
| Tool visibility | The catalog is filtered for the current caller. A handler without a description stays invisible. |
| Write safety | Approval mode proposes writes. The default does not dispatch them directly. |
| Identity | Approved handlers run as the calling user, with normal access and validation. |
| Tenant policy | Mode, caps, provider, and approval rules can be tenant-scoped. |
| Audit | Executed assistant tool calls sit beside the application’s audit trail. |
| Data lifecycle | Conversation turns use the same personal-data and erasure model as other user-owned data. |
| Attachments | Files are checked against their bytes, bound to the uploader’s own conversation, and treated as that user’s personal data. |
| Conversations | Each user sees and reopens only their own conversations, admins included. |
| Voice | Dictation, live transcription and read-aloud run through the tenant’s providers, never store audio, and count against monthly budgets in seconds and characters. |
Start with the runnable example
Section titled “Start with the runnable example”AI agent basic is a small service desk. The agent can search and close tickets with approval, while a high-risk purge operation can never become an unattended action.
Build your own assistant
Section titled “Build your own assistant”The runtime is optional. You can build an app-owned chat surface around
agent-tools, your own model client, and ordinary Kumiko handlers. Enterprise
provides the maintained tool loop, approval model, caps, voice path, and
persistence. A custom assistant remains responsible for its own conversation
storage, provider errors, UI, policy, and security review.