Layer 1
Assist SDK
Embeds Clara into customer sites with contextual chat, live voice controls, customer-friendly states, and tenant-aware runtime config.
domain allowlistssnippet installvoice statusworkflow context
Architecture
The depth that makes Clara defensible.
A serious customer agent must separate customer UX, tenant control, retrieval, workflow actions, telemetry, and live voice operations.
Layer 2
Clara API
The tenant-aware platform boundary for channel requests, runtime configuration, workflow execution, connectors, and administration.
tenant routingworkflow runtimeprovider configaudit surfaces
Layer 3
Unified workflow runtime
One shared runtime owns every customer turn across channels, selects an enabled workflow, persists execution state, and invokes only allowed tools and connectors.
workflow registrydurable sessionsbounded toolscustomer support fallback
Layer 4
Voice workflow adapter
LiveKit handles real-time media and turn-taking while the same Clara workflow runtime used by web chat owns business decisions and responses.
LiveKit Cloudshared runtimeinterruption handlingsession reports
Runtime behavior
How Clara keeps decisions durable and actions bounded.
The architecture separates conversational reasoning from authoritative workflow state and connector execution.
A durable record for every workflow turn
The shared runtime persists workflow sessions, prepared turns, selected workflows, and tool operations in PostgreSQL. Another Clara API instance can continue the session from authoritative state instead of reconstructing transactional progress from conversation text.
Bounded tools between AI and business systems
Models do not call tenant systems directly. Transactional workflows expose explicit tool contracts, and connector capabilities resolve only through bindings owned by the authenticated tenant. Missing, mismatched, or unavailable bindings fail closed.
Replay-safe external actions
Retries reuse a stable operation identity. Leased tool-operation records prevent repeated requests from duplicating orders, bookings, refunds, or other connector mutations while still allowing interrupted workflows to resume safely.
Channel-neutral outcomes with end-to-end evidence
The runtime returns directives that chat and voice adapters can project into their own experiences. Correlation IDs and a persisted event timeline connect routing, tools, connectors, latency, outcomes, handoffs, and evaluation evidence for diagnosis.
Evaluation path
01
Pick one high-volume support reason.
02
Install Clara Assist SDK on a demo or staging site.
03
Load tenant knowledge, policies, and workflow actions.
04
Validate voice, chat, handoff, analytics, and guardrails.
05
Expand across channels and customer journeys.