Pay by voice — without handing over control.
An agentic assistant that understands your customers, retrieves what they need, and performs approved actions — including payments — against your existing backends. Text or voice, no model training, no backend rebuild.
The principle that runs through everything
The LLM is never the security boundary. A tool appearing in the model's list does not authorize it — authorization is re-checked immediately before any consequential action, and approvals are bound to the exact action, amount and resource.
Illustrative — the payment executes exactly once, only after a human approves.
From a spoken request to a safe action.
Every request runs the same path — and the model only ever sees what that customer is allowed to use.
Identify
The tenant and the authenticated customer are resolved from trusted application context — before a word reaches the model.
Session & context
A conversation session is created or resumed and the relevant context is assembled for the request.
Offer only authorized tools
The request and only the tool schemas that customer is allowed to use are sent to your chosen LLM.
Authorize before execute
The LLM proposes a tool; the policy engine validates and authorizes the call independently of the model.
Park writes for approval
Consequential actions like a payment are parked for a human — bound to the exact action and amount, with an expiry, and executed exactly once.
Respond & audit
The grounded result becomes a voice or text reply, and the full request → tool-call → approval → response trace is recorded.
An assistant that acts — on a leash.
Voice & text
A framework-independent, embeddable widget that listens and speaks — your customers ask in natural language and get a voice or text answer.
Bring your own LLM
Provider-neutral by design — plug in the model you trust, with a deterministic mock for development and testing.
MCP gateway
Connect your MCP servers and existing REST APIs (via adapters) through one capability abstraction — no backend rebuild, no model training.
Policy engine
Every tool call is authorized immediately before execution, independently of the LLM — a tool appearing in the model's list never authorizes it.
Human-in-the-loop approval
Consequential writes are parked, bound to the exact action, amount and resource with an expiry, and execute exactly once after an approver signs off.
Tool & connection registry
Admins register, test and discover tools with schema-change detection, and manage MCP connections and REST connectors — tenant-scoped and admin-gated.
Secure by construction
HMAC-signed, token-bound sessions, multi-tenant isolation and SSRF defenses throughout — everything persisted and audited.
Reliable at runtime
Rate limiting, idempotency, health probes and distributed tracing come built in, not bolted on.
Part of one platform
The voice. The guardrails. The rails.
The assistant decides what to do; AgentTrust governs whether an agent may move money; Payment Orchestration moves it. One platform, one audit trail — from the first spoken word to the ledger entry.
Give your customers an agent they can talk to.
Embed the widget, connect your backends, keep every consequential action behind approval.