Kumiko Enterprise

Enterprise Kumiko Enterprise brings AI into the business app itself. The commercial features sit on the open framework and connect to the same registry, handlers, permissions, tenants, jobs, and audit trail that your app already uses. Talk to us about access.

Each feature card below links a generated reference page (config keys, dependencies, secrets), the same drift-free tables as the bundled features, introspected from the booted registry.

Start with the AI workflows

Choose the workflow that matches the product problem:

You want to…	Start here	What it gives you
Let users ask questions and propose changes in the running app	The Assistant	Role-filtered tools, approval cards, tenant policy, caps, voice, and audit
Turn a natural-language feature idea into typed Kumiko source	ai-generate	A generated feature file with a compile-validation loop and human review
Change an existing feature with a focused request	ai-patch	Reviewable PatternChange[] edits instead of an opaque rewrite
Extract structured data from documents	ai-extract	JSON-Schema-shaped output for invoices, contracts, receipts, and similar inputs
Route inbound messages or tickets	ai-triage	Categories, sentiment, urgency, and proposed next actions
Tune a pipeline without redeploying code	ai-pipeline	Tenant policy, drafts, activation, rollback, golden fixtures, and dry runs

The common path is provider → policy → AI operation → review or domain write. The AI package decides what the model may see and propose; your application still owns the final domain action.

AI assistantPut a governed assistant inside the app and keep users in control of writes.

Document AI workflowsExtract, classify, and route documents and messages without hiding the domain write.

AI feature builderGenerate and patch Kumiko source, then review, validate, version, and publish it.

AI operationsOperate prompts, policies, providers, provenance, budgets, and rollbacks.

Why AI inside Kumiko?

The useful part is the application boundary around the model:

• App-aware: the assistant reads the composed registry instead of a second, hand-maintained schema.
• Permission-aware: the model receives only the tools the current caller can use.
• Reviewable: writes become proposal cards by default and approved actions dispatch as the calling user.
• Tenant-aware: providers, modes, caps, prompts, and pipeline policies can vary per tenant.
• Auditable: executed assistant tool calls sit beside the application’s normal audit trail.
• Replaceable: provider adapters and application-owned implementations stay possible when the packaged composition is not the right fit.

In-app AI assistant

ai-agent puts a chat layer over a running app. It reads the same registry the app is built from, so it can search, open and, with a human approving each write, change your records. ai-agent-edit is the add-on that lets a low-risk write run unattended once the user has said “always”.

LightDark

The Assistant: talk to your appWhat the chat layer does in a running app: the proposal card, the permission rules, modes and caps, voice, and what it deliberately is not.

Make your app agent-readyThe guide: describe every handler, declare the risk, read the doc-gap check, pre-seed the tenant's approval rules.

ai-agent-basic recipeA runnable service desk the agent can search and close tickets in, but never purge.

AI feature-builder

Generate, refine, and draft with LLMs, the engine behind the visual Designer and the AI authoring flow.

ai-foundationThe LLM platform layer, per-tenant provider resolution, prompt assembly, system-prompt config.

ai-generateGenerate a whole feature from a prompt, with a compile-validation loop.

ai-patchSuggest a targeted patch to an existing feature.

ai-draftDraft a reply message from a source message and a stated purpose.

ai-triageTriage and route incoming items with an LLM.

ai-textCompletion, grammar fixes, translation, and rewrites for the AiTextField and AiTextArea widgets, with a monthly request limit per tenant.

ai-conversationPersistent, event-sourced multi-turn conversations on top of ai-foundation.

prompt-storeEvent-sourced, versioned prompt templates with edit/revert handlers.

designerVisual feature builder backed by pattern storage.

AI providers

Pluggable LLM backends. A tenant adds connections (OpenRouter is an OpenAI-compatible connection) and picks standard models per job.

ai-provider-anthropicAnthropic Claude provider.

ai-provider-openai-compatOpenAI-compatible provider for text, transcription and read-aloud (OpenAI, Azure, local).

ai-provider-mockScriptable mock provider for tests and demos.

Document AI and voice

Use the same provider boundary for documents and speech. ai-extract turns text or PDFs into schema-shaped data, while the transcription and speech capabilities of the AI providers feed voice input into the assistant and read answers aloud.

The parser providers plug into the open document-ingest-foundation and turn uploads into text before any AI step runs. Mail from Microsoft 365 comes in through the open inbound-mail-foundation with the Graph provider.

ai-extractExtract structured data from invoices, contracts, receipts, and other documents.

Voice in the assistantPush-to-talk, live transcription, read-aloud and hands-free, with their budgets.

document-ingest-provider-liteparseParse PDFs and images, with OCR for scans.

document-ingest-provider-officeRead the text of DOCX and DOC files inside the worker.

inbound-provider-m365-graphRead Microsoft 365 mailboxes through the Graph API.

Build your own version

Enterprise is an implementation layer, not a requirement of the framework. You can build an application-specific alternative with the public feature, handler, job, screen, provider, and renderer APIs. For example, an app can use its own LLM client, prompt storage, approval UI, extraction handler, or worker.

The Enterprise packages save that work and provide a maintained composition, but they are not automatically drop-in compatible with a custom implementation. An app-owned version remains responsible for provider integration, access rules, validation, tests, prompt/data handling, security review, and production operations. See the open framework concepts when you want to own that layer yourself.

Pattern storage & publishing

Store, version, and publish the feature patterns the Designer and AI builder produce.

pattern-storageVersioned storage for feature patterns.

publishPublish stored patterns into a running app.

Advanced renderers

Richer notification rendering than the open renderer-simple.

renderer-mail-htmlHTML email rendering via the template resolver.

renderer-puppeteer-clientPDF / image rendering via a Puppeteer service.

Domain features

creditEvent-sourced amortization credits. The feature stores the inputs and computes the monthly schedule on read.

GDPR coverage

Features that keep personal data ship a separate *-user-data feature. It registers the export and erasure hooks of the open user-data-rights feature for that feature’s entities. Mount it next to the feature when your app runs the user-data-rights pipeline: ai-agent, ai-agent-attachments, ai-conversation, credit, pattern-storage, and prompt-store.