What Wani does and refuses to do: no call without consent re-checked at dispatch, inside the recipient's calling window. Capabilities, the API, and what still needs a telephony credential.
Wani answers inbound calls and places outbound follow-ups, reminders, and lead-qualifying calls, built on a voice-agent flow your workspace designs from a template, a hand-built configuration, or a plain-language description. Every outbound dial passes through one shared gate that checks consent, the National DND list, your suppression list, and the recipient's own local calling window — the same gate, whichever route the call came from, so nothing gets to skip it.
No call places without consent re-read at dispatch
Not just when the call was queued. The same function that authorised the call at queue time runs again, fresh from the database, immediately before an attempt row is written — so consent withdrawn in between is honoured.
Every call stays inside the recipient's own local calling window
Enforced as database CHECK constraints — not application logic a bug could skip. A workspace can narrow the window; it cannot widen it, at the API, through a migration, or through direct SQL.
The blocked-dial log is evidence, not an error list
Every refused or deferred dial is written with the reason, the recipient's local time, and what was attempted — retrievable on its own route, the record a regulator would ask for.
The AI-disclosure and human-handoff steps cannot be removed
Re-inserted into a call flow on every write, whether or not a client tried to remove them. Their wording is editable; their existence is not — removing one from a request simply does not take effect.
A collections agent cannot skip the tone check
A collections-purpose voice agent cannot be saved without its tone being checked, and cannot be saved at all if the check finds debt-collection-threat language — enforced as a database CHECK constraint, not only as a review step someone could skip.
Nothing dials without a telephony provider credential
Every call sits at awaiting_provider — queue, compliance gate, retry backoff, and dispositions all work without one; the response says plainly that nothing was dialled.
A caller who asks for a human is transferred, full stop
There is no branch in the handoff route that refuses an explicit request, whatever else the call was about.
Getting started
Capture consent per phone number with a basis and evidence for it. Expiry is computed server-side from the basis you gave — never accepted as a value from the caller — so a workspace cannot quietly extend it by passing a later date. Opting a number out is one-way: it cancels every pending and deferred call already queued for that number.
Start from one of six shipped templates, hand-configure the steps, or describe what the agent should do and let a model draft a plan. All three paths write through the same normaliser, so the AI-disclosure and human-handoff steps land in every flow regardless of which path you used, and a step naming an unregistered type, action, or field is dropped and reported in your own words rather than silently ignored.
Queueing checks consent, DND, your suppression list and the calling window once, as an early courtesy refusal. Dispatch checks all four again, fresh, the moment before an attempt is made — and a call outside the calling window is deferred with the next permitted time returned in the response, not silently dropped.
Capabilities
Consent capture with basis-derived expiry, permanent opt-out, a workspace suppression list, and the blocked-dial log all work locally — no queue or dial route can be reached without them.
Needs: The National DND feed needs a telecom access-provider credential. Until it exists, an unknown DND answer blocks promotional dialling rather than permitting it.
The permitted window (09:00-21:00) is evaluated in the recipient's own local time zone, and a call outside it is deferred with a visible next-attempt time. A workspace can narrow the window; nothing can widen it.
Queue, the compliance gate, deferral, per-contact frequency cap, escalating retry backoff, dispositions and callbacks are all real.
Needs: Dispatch ends at awaiting_provider: placing the call itself needs a telephony credential.
Reminders queue off a Wavy appointment as independent, individually gated service-category calls.
Needs: Placing them needs a telephony credential.
Queues a qualifying call against a Jeet lead through the same gate as every other entry point.
Needs: The call itself needs a telephony credential; scoring the answers needs the call to have happened.
A campaign launches straight at a contact list: POST /campaigns queues every contact through the same gate a single manual call would use, reporting queued, deferred and blocked per contact rather than aborting the whole batch on one bad row. Pacing is recomputed on every read from the minutes left in the recipient's window, clamped to the carrier's documented ceiling, and a completion monitor — recomputed the same way, never stored — moves a running campaign to completed once nothing is left pending. A paused campaign is refused at the gate, not by filtering a queue that could still leak through elsewhere.
Needs: Dialling at that paced rate needs a telephony provider.
The voicemail drop message itself is enforced, not just written free-hand: sender identification and an opt-out instruction cannot be saved blank, the same CHECK-constraint discipline as the AI-disclosure node, and the rendered message composes both in by construction rather than trusting free text to remember them. Connect-analysis storage, and its honest no_detection default when nothing has actually detected a machine, are real too.
Needs: The telephony provider's own answering-machine-detection feature (Exotel/Ozonetel AMD) to actually tell a voicemail box from a person answering, plus TTS from that same provider to speak the message — neither is connected, so nothing but this API writes a row here yet.
Disclosure logging, the consent gate, and a mandatory retention clock are enforced. The recording's consent reference is NOT NULL at the schema level, so a recording that names no disclosure cannot be written.
Needs: Capturing audio needs a provider recording API and encrypted object storage.
An explicit request for a person is classified before scope or confidence, and the route has no branch that refuses one. A staff roster with department, skill and availability, plus a round-robin routing worker, now picks who the call goes to, and that department, skill and target staff member are stored on the handoff. The context packet records that something was withheld from the summary, never what.
Needs: The warm-transfer telephony leg and pushing an alert to the receiving agent's device both need providers this deployment does not have; the handoff response names which one is missing (no_staff_available or no_push_provider_configured) rather than claiming a device was paged.
Six templates, a hand-built configuration, or a plain-language description, all normalised through the same function so disclosure and handoff steps are always present. The model proposes and a registry disposes — an unregistered step type, action, or field is dropped and reported, never silently accepted. The knowledge base retrieves by keyword overlap, and an empty result is treated as the important answer, not hidden.
Needs: Generation from a plain-language description needs an LLM key; templates and the validator do not.
Voice selection over Sarvam's own catalogue. Listing returns the real catalogue or an empty list naming the missing credential — never a plausible but unverified voice id, because one nobody checked fails at synthesis on a live call.
Needs: Holding a full conversation in that language additionally needs speech-to-text, which is not built.
A register of the numbers a business calls from with its DLT header, and a record of which published agent answers on which number. A draft cannot be deployed, and an inbound deployment without a number is refused by a CHECK constraint.
Needs: Nothing is provisioned and no call is answered: buying, porting, or masking a number is a carrier relationship, and there is deliberately no inbound webhook until a provider exists to sign it.
Writes the disposition and summary onto the contact's Wavy conversation as a summary and a tag, never as a message in the customer's own thread. Captured fields stay flagged voice-sourced and unconfirmed.
Daily volume, outcome breakdown, and blocked-dial counts from this workspace's own rows only. Connect rate and average talk time are null, not zero, when nothing has an outcome yet — a zero would be a false statement about performance.
Summary generation runs end to end off the call's captured fields, notes and disposition through the shared LLM layer, tagging every point as either a confirmed captured field or the model's own inference rather than presenting both the same way, cached after the first read and refreshed on demand. A transcript can never be saved without an explicit language, to guard against a Hindi or code-mixed call being silently mangled by an English default.
Needs: An LLM key for the summary itself. The actual transcript text is a separate gap: this deployment records no call audio and has no speech-to-text credential, so GET /transcript always reports why rather than fabricating one — the storage and validation for a real transcript exist and are tested, waiting on that provider.
Quality flags are classified in the same LLM pass that builds a call's summary, so this needs no audio or transcript either. A flag naming a type outside the closed set is dropped rather than stored under a guess. Listing, review — which always requires a human's own note, never an auto-verdict — and a summary that groups repeated unhandled-request flags by topic into a suggested call-script-builder fix all run locally.
Needs: An LLM key (LLM_API_KEY or a rotation pool) — the only blocker.
API surface
Every route below is mounted and reachable today. Requests and responses are real shapes, not illustrations.
/api/v1/agents/wani/consentCapture; expiry basis-derived, never caller-supplied expires_at is computed server-side as captured_at + 7 days for an explicit basis — the caller cannot pass a later date.
Request
{
"phone": "+919876543210", "basis": "explicit", "source": "website_form",
"evidence": "Checkbox ticked on pricing page enquiry form, 2026-08-15 14:02 IST, form_id=pricing-cta"
}Response
201
{ "phone_e164": "+919876543210", "basis": "explicit",
"expires_at": "2026-08-22T14:02:00+05:30", "callable": true }/api/v1/agents/wani/consent/{phone_number}State and callable flag
/api/v1/agents/wani/consent/{phone_number}/opt-outOne-way; cancels pending/deferred queue rows
/api/v1/agents/wani/suppressionsAdd to the workspace's own do-not-call list
/api/v1/agents/wani/compliance/check/{phone_number}Preview the real gate; nothing logged
/api/v1/agents/wani/compliance/blocked-logEvery refusal, with its reason An unknown DND answer blocks a promotional call rather than permitting it — the safe direction when the feed is not yet connected.
Response
200
{
"items": [{
"phone_e164": "+919812345678", "purpose": "promotional", "reason": "dnd_unknown",
"detail": "No fresh National DND answer for this number. Promotional calling stays blocked until a telecom access-provider DND feed is connected.",
"callee_timezone": "Asia/Kolkata", "callee_local_time": "2026-08-16 11:30",
"attempted_at": "2026-08-16T11:30:04+05:30"
}], "total": 1
}/api/v1/agents/wani/compliance/calling-windowPolicy plus open_now and next_open_at
/api/v1/agents/wani/compliance/calling-windowNarrow only; DB constraints are the real control
/api/v1/agents/wani/compliance/dndBulk-load a DND scrub result; numbers stored hashed
/api/v1/agents/wani/outbound/queueQueue a call; early compliance check
/api/v1/agents/wani/outbound/queue/{id}One queued call, with attempts and blocks
/api/v1/agents/wani/outbound/deferredCalls waiting for the window to open
/api/v1/agents/wani/outbound/queue/{id}/cancelCancel
/api/v1/agents/wani/outbound/queue/{id}/dispatchLive gate re-run; 202 or 409 with deferred_to The same gate ran once at queue time and once again, live, at dispatch — this is the second run.
Response
409
{
"message": "Refused: this call was not placed.",
"reason": "outside_calling_window",
"detail": "It is 22:14 in Asia/Kolkata; the permitted window is 09:00-21:00 local, and it is not a setting a business can widen.",
"deferred_to": "2026-08-17T09:00:00+05:30",
"checks": [
{ "check": "consent", "result": "active_explicit" },
{ "check": "dnd", "result": "clear" },
{ "check": "calling_window", "result": "closed_outside_calling_window" }
]
}/api/v1/agents/wani/remindersQueue T-24h/T-2h off a Wavy appointment
/api/v1/agents/wani/leads/{lead_id}/qualify-callQueue against a Jeet lead
/api/v1/agents/wani/calls/{id}One call
/api/v1/agents/wani/calls/{id}/outcomeRecord a disposition
/api/v1/agents/wani/calls/{id}/opt-outMid-call opt-out
/api/v1/agents/wani/calls/{id}/handoffClassify and transfer; explicit request never refused
/api/v1/agents/wani/handoffs/{id}/contextWithheld flag, never withheld content
/api/v1/agents/wani/calls/{id}/recording/consentLog the disclosure event
/api/v1/agents/wani/calls/{id}/recording/start409 without a logged disclosure
/api/v1/agents/wani/calls/{id}/connect-analysisHonest no_detection default without AMD
/api/v1/agents/wani/calls/{id}/transcriptReports why, rather than fabricating one, with no STT
/api/v1/agents/wani/calls/{id}/summaryLLM summary of captured fields, notes, disposition
/api/v1/agents/wani/calls/{id}/summary/regenerateRecompute; 503 naming LLM_API_KEY if unset
/api/v1/agents/wani/quality/flagged-callsClosed-set flag types; unknown types dropped
/api/v1/agents/wani/quality/flags/{id}/reviewAlways needs a human's note, never an auto-verdict
/api/v1/agents/wani/quality/summaryRepeated flags grouped into a suggested script fix
/api/v1/agents/wani/flowsCreate a flow; disclosure and handoff nodes always inserted The request supplied an empty nodes array; both system nodes were inserted anyway. Wording is editable, existence is not.
Request
{ "name": "Appointment reminder", "flow_type": "outbound", "business_name": "Aster Clinic", "nodes": [] }Response
201
{
"nodes": [
{ "id": "system-ai-disclosure", "type": "prompt", "system": true, "removable": false,
"text": "Hello — this is an automated AI assistant calling on behalf of Aster Clinic. I am not a human. You can ask to speak to a person at any time, or ask me not to call again.", "next": "system-human-handoff" },
{ "id": "system-human-handoff", "type": "transfer", "system": true, "removable": false,
"text": "Of course — putting you through to a person now.",
"trigger": "caller asks for a human, at any point" }
],
"status": "draft", "version": 1
}/api/v1/agents/wani/flows/{id}Update; re-normalised on every write
/api/v1/agents/wani/flows/{id}/publishBlocks (400) on an invalid flow rather than warning
/api/v1/agents/wani/voice-agents/templatesSix code-shipped templates
/api/v1/agents/wani/voice-agents/vocabularyThe full closed step/action/field registry
/api/v1/agents/wani/voice-agents/draftLLM-drafted plan; creates nothing
/api/v1/agents/wani/voice-agents/plan/validateSame validator, no model needed
/api/v1/agents/wani/voice-agents/{id}/publishBlocks (400) rather than warns on an invalid flow
/api/v1/agents/wani/voice-agents/{id}/queue-callUses the agent's own declared purpose, never the caller's
/api/v1/agents/wani/knowledge-bases/{id}/searchKeyword overlap; an empty result names itself
/api/v1/agents/wani/voicesReal catalogue, or empty with the missing credential named
/api/v1/agents/wani/voice-agents/{id}/deployCHECK-refused without a phone number for inbound
/api/v1/agents/wani/calls/{id}/syncWrites to the Wavy conversation, never the customer thread
/api/v1/agents/wani/analytics/dailyNull, not zero, when nothing has an outcome yet
Limits and gotchas
Queueing, the compliance gate, retry backoff, dispositions and callbacks all work today. Every call stops at awaiting_provider until a telephony credential (Exotel, Ozonetel, or Twilio Voice) is configured.
Wani enforces a 09:00-21:00 recipient-local calling window as database CHECK constraints (start_hour_local >= 9, end_hour_local <= 21). The seven-day explicit-consent-expiry rule is verified against TRAI's own text: the Telecom Commercial Communications Customer Preference (Second Amendment) Regulations, 2025, notified 12 February 2025, states that explicit consent to complete an ongoing transaction "shall be for seven days or as directed by the Authority" — and this product's consent-expiry constant matches that figure exactly. The 09:00-21:00 window is a different case: that amendment establishes calling-time band as a registrable per-subscriber preference but does not state a single national window in the text read for this product. Until the 2018 principal regulation's Schedule is read to confirm one, "TRAI mandates 09:00-21:00" could not be independently confirmed against a primary regulatory source, and this window is documented as this product's own, deliberately conservative commercial limitation — never call outside it — rather than a citation of a specific regulation.
Voice selection works against Sarvam's real catalogue, but holding an actual spoken conversation needs speech-to-text, which is not built — and neither is the batch STT a real transcript would need, so GET /transcript always reports why rather than fabricating one, even though the table and the ingestion code behind it are written and tested. A call summary does not wait on either: it is generated by an LLM from the call's captured fields, notes and disposition, not from audio, so it works today with only an LLM key. Actually detecting voicemail needs the telephony provider's own answering-machine-detection feature, which is also not connected.
Until a telecom access-provider DND credential exists, an unknown DND answer blocks promotional dialling rather than permitting it — the safe default, but it means no promotional call can go out at all in the meantime.
Retrieval works with no credential; embeddings-based semantic search needs one and is not wired up.
What this needs from you
Everything else on this page works with none of these connected.
| Credential | Unlocks |
|---|---|
| A telephony provider (WANI_TELEPHONY_PROVIDER — Exotel, Ozonetel, or Twilio Voice) | Placing or receiving any call at all, and capturing recording audio. Everything up to that moment works without one. |
| A telecom access-provider DND API | The National DND scrub. Until connected, an unknown answer blocks promotional dialling rather than permitting it. |
| A Sarvam API key | The real voice catalogue and TTS preview. Without it, GET /voices returns an empty list naming the missing key rather than a guessed one. |
| Speech-to-text | Holding a live spoken conversation, and a real, verbatim call transcript. The transcript table and its ingestion path are already built and tested — they are just unreachable with no recorded audio and no STT provider to transcribe it. Call summary and quality monitoring do not need this: both already run off captured fields with only an LLM key. |
| An embeddings provider | Semantic knowledge-base search. Keyword-overlap search works without one. |
| An LLM key | Drafting a voice-agent plan from a plain-language description, generating a call's summary, and classifying quality flags — the same shared analysis layer covers the last two, since there is no separate background worker in this deployment. Templates, the hand-built builder path, and the validator all work without one. |
Questions
No. Consent is checked once when a call is queued and again, live, immediately before dispatch — the same authorisation function runs both times, reading fresh from the database, so a withdrawal that happened in between is honoured at dispatch even if the call was queued before it.
09:00 to 21:00 in the recipient's own local time zone, enforced as database CHECK constraints a workspace can narrow but not widen. Wani enforces a 09:00-21:00 recipient-local calling window as database CHECK constraints (start_hour_local >= 9, end_hour_local <= 21). The seven-day explicit-consent-expiry rule is verified against TRAI's own text: the Telecom Commercial Communications Customer Preference (Second Amendment) Regulations, 2025, notified 12 February 2025, states that explicit consent to complete an ongoing transaction "shall be for seven days or as directed by the Authority" — and this product's consent-expiry constant matches that figure exactly. The 09:00-21:00 window is a different case: that amendment establishes calling-time band as a registrable per-subscriber preference but does not state a single national window in the text read for this product. Until the 2018 principal regulation's Schedule is read to confirm one, "TRAI mandates 09:00-21:00" could not be independently confirmed against a primary regulatory source, and this window is documented as this product's own, deliberately conservative commercial limitation — never call outside it — rather than a citation of a specific regulation.
No. Every flow write runs through the same normaliser, which strips whatever was submitted for the disclosure and human-handoff steps and rebuilds them from the system template, keeping only the wording. An attempt to omit them from a request simply does not take effect — the flow comes back with both steps present regardless.
They are transferred. The handoff route classifies an explicit request before scope or confidence, and there is no branch anywhere in it that refuses one.
GET /api/v1/agents/wani/compliance/blocked-log — every refusal, filterable by reason, with the recipient's local time at the moment it was attempted. This log exists specifically to be evidence, not just a debugging trail.
The compliance, consent, and flow-building machinery is built and tested, but no call can actually be placed without a telephony credential your workspace has not configured yet — every call stops at awaiting_provider. That is a configuration gap, not a review gate, but the boundary copy on this page should still be checked by someone before launch, per docs/REQUIRED_FROM_USER.md.
Coming soon