Multi-tenant AI orchestration
Turn every conversation into an understood intent, a real answer, and a controlled action.
runs the part between your customers and your systems: retrieval over your own knowledge, actions through tools you register, escalation to a person when it matters, and a measured outcome for every one.
- Deploy as
- A website widget, a social inbox, or inside your own product.
- Answers from
- Your site, your documents, your FAQs — ingested, sanitized and scoped to you.
- Acts through
- Tools you register against your own API, with your system authorizing every call.
A conversation becomes an outcome
Not a chatbot. A message arrives on any channel and follows one path: understood, answered from your own knowledge, acted on against your own systems, followed up, escalated to a person when it should be, and measured.
01
Understand
Intent, not keywords. The prompt is assembled from fixed slots in a fixed order, and content a visitor or a crawled page can influence is marked as data rather than instruction.
02
Retrieve
Answers come from your own site, documents and FAQs, ingested, sanitized and scoped to you. Retrieval carries an explicit tenant predicate, so recall does not quietly degrade as the corpus grows.
03
Act
Registered tools call your own API through a connector. The model emits a request; ten gates decide whether it becomes an HTTP call, and your system authorizes it again on arrival.
Built for running other people’s businesses
Your data stays yours
The platform never learns your database schema and never issues SQL against it. Your application remains the system of record; everything reaches it through a tool you registered, over HTTP, with your own authorization on every call.
Isolation, twice
A repository guard in the application and row-level security underneath it. Either alone eventually fails; together, a forgotten predicate is an empty result set rather than a breach.
An audit trail that is true
Audit writes share the transaction that made the change. A rolled-back mutation leaves no audit row, and a committed one always leaves exactly one — a best-effort trail is worse than none, because it looks like one.
Cost you can see
Every model call is priced at write time against a versioned price book, in integer micros, pinned to the exact price row used. A model with no price records zero and appears in a report, never a silent estimate.
White label to the ground
Name, logo, colour and the assistant’s own title are tenant data. Your customers can see nothing of the platform underneath.
One conversation, many channels
Website chat, Instagram, Messenger, WhatsApp, email and in-app assistants resolve to the same conversation model, the same escalation lifecycle and the same inbox.
Start with a trial workspace
No card. The trial is a conversation allowance, and you can see what it has spent at any point.