ThunderPhone v3 is live.Self-serve, from 2¢/min.Read the announcement
Product updates from the ThunderPhone team

Changelog

Dates reflect when changes reached production.

v3.0.0 — November 9, 2026

  • ThunderPhone 3.0 is built around Zeus, an AI assistant that builds, tests and runs your voice agents with you. Describe what your calls need to do; Zeus drafts the agent, picks its voice, languages and settings, tests it, quotes its price and puts it live when you say so. The app opens on Zeus, with every page redesigned around it.
  • Try an agent before you sign up: describe your calls on thunderphone.com, talk to the agent Zeus drafts right in the browser, and keep building in Zeus after signup.
  • Test calls run in the chat, in the browser or to your phone, and Zeus can try a change on an agent's next calls, or split them between two versions, before you publish.
  • Every change Zeus makes shows the changed lines with Undo, and you can rewind a whole reply. Draft work needs no approval; anything that changes what callers get asks first.
  • Zeus keeps working with the tab closed and emails you when it's done or needs a yes. It runs scheduled routines, remembers account preferences, and takes @-mentions, / commands, dictation and pasted screenshots.
  • Zeus reads your calls: search calls by meaning, see where callers drop off, get problem turns marked and translated inline, save reports to refresh later, and listen in on a live call or whisper to the agent mid-call.
  • With your approval, Zeus works in your outside accounts: it connects carriers and numbers from the chat, calls Twilio, Telnyx, SignalWire, Vonage, n8n, Zapier and Make, and drives a browser you can watch for web apps with no API. It never stores the values you type.
  • The new agent editor shows an agent as steps or a single prompt, with a flow graph, Split into steps, and pages for voice and languages, knowledge, data to collect, call handling, tools and handoffs, tests and history.
  • Zeus quotes each agent's price per minute from its test and real calls, keeps it under a price target you set, and can end a call with a goodbye at a hard per-call cap. Billing and Calls break every call's cost down by part.
  • New pricing: each call costs the list prices of its parts (understanding, hearing the caller, the voice, the phone line, call review) plus infrastructure and a published markup, with no subscription or seat fees. New accounts start on it; existing accounts move on December 1 or switch early from Billing. Pricing.
  • Call settings have new names: Intelligence is Core or Pro, Hearing is Lite, Standard, Plus or Max, and faster replies is Speed boost. The API returns the new values and accepts the old ones; GET /v1/agents/{id}/reply-time estimates typical and slow reply times; the Python and TypeScript SDKs are at 0.3.0.
  • Full duplex, a speech-to-speech option for calls where naturalness matters most, is available in Zeus.
  • Terms of Service v3.1 cover Zeus and usage-based pricing and take effect December 1, 2026.

v2.8.0 — November 7, 2026

  • Outbound deliverability: campaigns dial from a pool of up to 50 local numbers (from_numbers), call in each contact's local time, spread retries across days, and stop dialing from numbers carriers flag as spam. Each number shows its outbound health and expected STIR/SHAKEN attestation. Guide.
  • ThunderPhone numbers carry your organization's caller name, and numbers on your own Twilio or Telnyx account get free carrier registration.
  • A campaign can text each contact a heads-up before the first call, skip mobile or landline numbers, and default to each state's legal calling hours. Admins can stop outbound calls to chosen US states at once (blocked_call_states).
  • Schedule a call or text to one contact with POST /v1/scheduled. The opt-out list (/v1/opt-outs) applies to every outbound call and text, and outbound calls that hear "stop calling me" end politely and add the number to it.
  • Contact centers: inbound SIP headers and UUI reach the agent, webhooks and API; transfers go to sip: addresses with REFER headers or hand the call back to the platform with a short summary. SIP endpoints can get dial-in numbers and authenticate by IP allowlist. Setup guides cover Amazon Connect, Genesys Cloud, Five9, NICE CXone, Talkdesk, Twilio Flex, Dynamics 365 Contact Center and Zoom Contact Center.
  • The agent finishes its word and pauses when the caller makes a sound, then carries on if it wasn't an interruption; backchannels, repeats and crosstalk no longer cut it off.
  • The AI disclosure plays as a recorded intro in its own voice, in 47 languages, with a choice of voice, style and rings before the agent speaks (disclosure_voice, disclosure_style, disclosure_ring_count).
  • Pass free-text per-call context alongside variables in config webhooks, POST /v1/call and POST /v1/realtime/sessions. Realtime API sessions can run a saved agent with per-session instructions and variables.
  • An agent can have its own call setup webhook (call_setup_webhook_url), and numbers can be pinned to a version of their agent (pinned_versions).
  • Call summaries are written in your account's language, with a per-agent override (summary_language).
  • English calls get a second, more careful transcript after the call, shown on the call page.
  • Deleted agents can be restored for 30 days.
  • Agent history tracks every setting change, not just prompts.
  • Every voice in the voice library is selectable in the API and MCP; GET /v1/voices/library filters by several languages, ranks by description (like=) and splits included from paid voices (price=).
  • Cal.com agents can find and reschedule a caller's booking.
  • Transfers can be warm-only per destination, and Ask first confirms who answered before briefing them.
  • API responses, webhooks and call history use ThunderPhone labels and neutral transcript keys (fast_transcript, second_transcript). The old transcript keys leave the live stream after November 10.
  • /v1 JSON responses are gzip-compressed; tool names need to be unique only within an agent.
  • Admins get an email when their call setup webhook, carrier or a connected app keeps failing real calls.
  • Stay signed in to several accounts in one browser and switch between them.
  • Password changes, resets and sign-out-everywhere revoke OAuth grants; only a signed-in web session can approve an OAuth consent; MCP confirmation tokens confirm one call for one person.

v2.7.0 — October 23, 2026

  • One agent type replaces the three engines: every agent has the same intelligence, audio and reply-speed settings, and the builder shows the price per minute for what's on screen (GET /v1/agents/{id}/price, POST /v1/agents/price-quote). Old product values keep working in the API.
  • A deploy that would raise an agent's per-minute price asks for confirmation first, in the builder and in the API (409 price_change_needs_confirmation).
  • Call-start lookups: a custom tool can run with the caller's number before the greeting (trigger: "call_start"), so the agent can greet returning customers by name. Call details show each lookup.
  • Outbound voicemail: the agent stays quiet while a mailbox is suspected and leaves exactly one message after the beep. A person or call screener gets an answer at once.
  • Call transcripts translate into the viewer's language in one click, with English, Original and Both views.
  • Silence check-ins follow the conversation: a caller who asked for time gets it, and new settings choose how many check-ins come before the agent hangs up or transfers (silence_final_action).
  • Agents have a words list that sets how specific words are spoken and which words to listen for (words, with say_as, ipa and hear).
  • A caller returning a call from your number within 30 days reaches the agent with context from the earlier call; telephony.incoming carries returning_call.
  • Texts go out through your Twilio Messaging Service, with delivery status recorded; US senders need a verified 10DLC campaign.
  • Custom tool requests on phone calls carry the caller's number, the called number and the direction (X-ThunderPhone-From-Number, X-ThunderPhone-To-Number, X-ThunderPhone-Call-Direction).
  • MCP: 25 more tools, including organization settings, the caller blocklist, call analytics and text tests of an agent. A second URL, /v1/mcp/baa, serves call content once an admin confirms their AI assistant is covered by their own BAA. Guide.
  • Built-in Calculator tool, off by default, so agents total orders and prices exactly.
  • end_call, send_keypad_input and wait_on_hold can be turned on or off per agent; keypad and hold are offered on outbound calls by default.
  • Tool results up to 100,000 characters reach the agent in full, and long calls drop old tool results before the agent's context fills.
  • Auto-recharge falls back to a backup card when the bank declines the default.
  • Adding a viewer to a client portal emails them a sign-in link; client-portal custom domains verify reliably.
  • A reply that stalls partway through switches to a backup instead of leaving silence.
  • Fixed choppy agent audio right after an interruption or pause.
  • pipecat-thunderphone 0.1.1 supports Pipecat 1.13. Guide.

v2.6.0 — October 16, 2026

  • Inbound calls can open with a prerecorded intro, a stock recording notice or your own upload, and a ring before the agent speaks.
  • Language switching is an agent setting: Ask first (default) or Follow the caller (language_switching). A full sentence in another language switches at once; a short reply gets "Would you like to continue in Spanish?"
  • 563 more premium voices across all 47 languages, with search and an accent filter in the voice picker. Voice library.
  • With native voices on, an agent's voice only needs to cover its primary language.
  • When a caller says "one sec" or "hold on", the agent says "Sure." and waits.
  • When a caller cuts the agent off, the agent knows which part of its line was heard, so a "yes" no longer confirms something the caller never heard.
  • Agents wait briefly when a caller pauses mid-thought, so replies no longer start while someone is still reading out a number or address.
  • Phone numbers and email addresses are read back at a natural pace; confirmation codes stay paced.
  • Webhook endpoints can regenerate their signing secret from the dashboard or POST /v1/developer/webhook-endpoints/{endpoint_id}/rotate-secret, optionally rotating copies on other agents.
  • The copilot keeps replying in the conversation's language and transcribes dictation in it.
  • LiveKit Agents plugin arguments are now model, caller_number and callee_number. Guide.
  • VoIP connection secrets are encrypted at rest, and the production database accepts only private network connections.

v2.5.0 — October 9, 2026

  • Switch from Vapi, Retell or ElevenLabs Agents gradually: connect the old platform, see baseline outcomes, replay its past calls through the imported agent, run side-by-side test calls, then move inbound and outbound traffic over on a ramp with automatic rollback. How it works.
  • Three warm-transfer styles per agent: introduce, ask first, or whisper (warm_transfer_style). Hold music plays while the destination rings, and the handoff waits for the person who answered to finish their greeting. Guide.
  • Tool-wait speech is a setting: tool_start_line (always, only_if_slow, never), a switch for filler phrases (filler_phrases_enabled), and still-working lines worded for each call on slow tools.
  • Six typing sounds to choose from while tools run (typing_sound).
  • Agents use the agent's time zone, then the call's number, then the organization's; each call can set its own.
  • The agent editor shows who else is editing and warns before you deploy someone else's changes.
  • Tools saved with no inputs get a warning in the builder and in agent API responses (tool_has_no_inputs).
  • When a reply fails, the agent asks the caller to repeat instead of going silent.
  • Inbound greetings start 0.6 s after a connect tone or recording notice instead of after a 2–3 s pause.
  • Phone numbers carry number_type (thunderphone or voip).
  • Owners get an email and a dashboard banner with a Re-activate button when a failing webhook endpoint is paused.
  • Self-serve account deletion in Settings; deleted calls are removed from storage along with every stored previous version.
  • Realtime API: streamed replies arrive complete in response.output_audio_transcript.done and response.done.

v2.4.0 — October 2, 2026

  • Agent steps: split an agent into steps, each with its own instructions, tools, voice and transitions, and hand calls to other agents (steps, handoffs). Imports bring flows, pathways and workflows over as steps, and squads as handoffs. Guide.
  • SIP addresses: reach an agent over SIP with no phone number, through per-call URIs (POST /v1/calls/register-sip) or static endpoints (POST /v1/sip-endpoints). Guide.
  • Connect any VoIP provider: sip.thunderphone.com is the one inbound address, a setup assistant tests and fixes your provider's settings, incoming-call tests place a real call, and manual SIP connections set number and caller ID formats. Guide.
  • Call reports: Home and Call History show KPI tiles, outcome labels you define per agent, a needs-attention flag, and business-hours and test-number filters (GET /v1/calls/report-summary). Guide.
  • Transfers go only to named destinations in the agent's Transfers section (transfer_destinations); a refused transfer keeps the call going. Guide.
  • Cold transfers try SIP REFER first and bridge the call when the carrier refuses. Transfers can be turned on for a VoIP connection from the dashboard (POST /v1/voip-connections/{id}/enable-transfers).
  • Inbound caller blocklist per organization (/v1/blocked-callers), and permanent call deletion (DELETE /v1/calls/{call_id}, POST /v1/calls/bulk-delete).
  • Signed webhook and tool requests carry a timestamped X-ThunderPhone-Signature-V2 header; the Python and TypeScript SDKs 0.2.0 add verify_signature. Guide.
  • Organizations have a concurrent-call limit (20 by default, raised on request), new agents default to a 1-hour call limit, and campaign calling hours default to the from-number's time zone.
  • Outbound calls to emergency, sanctioned, invalid and premium-rate numbers are refused.
  • Organization API keys act only within their organization and can't manage members, billing or SSO; custom-tool and MCP header values are encrypted and write-only.
  • Outbound calls nobody answers end as outbound_no_answer, with busy and declined reported separately, so campaigns retry them correctly.
  • Spelled codes such as VINs and phone numbers are read back slower, in groups of three or four.
  • Agents know the current date and time, end with the closing line from your prompt, and never read out tool calls or language tags.
  • Realtime API: phone-menu handling is its own setting (ivr_mode).

v2.3.0 — September 25, 2026

  • The ThunderPhone app on Zapier starts Zaps when calls end or are graded, and places outbound calls and runs campaigns from any Zap. Guide.
  • The verified ThunderPhone node for n8n starts workflows on the same call events and places calls, runs campaigns, and reads call records. Guide.

v2.2.0 — September 21, 2026

  • Structured post-call data extraction produces typed records with transcript evidence for every value, delivered on call.complete or call.data_extracted; finished calls can be re-extracted.
  • Agents send text messages through the send_sms tool and the SMS API from SMS-capable numbers on their own carrier account.
  • OAuth 2.1 authorization with PKCE and the device flow lets MCP clients and the CLI sign in through browser consent instead of API keys. Guide.
  • MCP server at parity with the REST API, with 62 annotated tools, prompts and resources; tools that delete or dial require a confirmation-token exchange. Guide.
  • Developer tooling includes the thunderphone CLI, Python and TypeScript SDKs, and the @thunderphone/mcp stdio package; Agent Skills and a docs MCP server support coding agents. Developers.
  • Webhook delivery history in the dashboard, the API and the copilot.
  • Agent-scoped webhook endpoints deliver call events with the agent's identity.
  • Realtime API: saved-agent call events with call_events=1, response.metadata echoed on response.created, function_call_output acknowledged with conversation.item.done, and 32-character call IDs.
  • Added GET /v1/whoami connection checks, the campaign.completed event, outbound eligibility and a reason code on phone-number responses, and sample webhook payloads.
  • Agent-level max call and ring durations, plus per-call variables for saved agents.
  • Twilio: connect a parent account once and pick its subaccounts in the same flow; each subaccount becomes its own connection.
  • Calendar bookings send guest invitations by default, with per-connection control and attendee support in calendar_update_event.
  • Client-portal custom domains get their HTTPS certificate bound automatically after DNS verification.
  • ThunderPhone-provisioned numbers start inbound-only, with outbound access requiring use-case review.
  • Custom voices: gender is optional, with a warning for languages whose first-person grammar is gendered.
  • Copilot: proactive attention nudges, welcome updates, and import links for teams moving from another platform.
  • Auto-reload backs off after a card decline and tells the customer once instead of retrying.
  • Webhook test sends get one retry and never mark an endpoint failing.
  • Email tool arguments are normalized and validated before customer tools run.

v2.1.0 — September 11, 2026

  • Import your agents from Vapi, Retell, ElevenLabs Agents, or Bland: paste an API key or a JSON export, review the mapped prompts, voices, tools, and knowledge bases, and import in one go. How it works.
  • Migration guides for each platform under /guides.

v2.0.1 — September 2, 2026

  • Launched TranslateMyCall: a free public phone line — (228) 888-8666 — that live-interprets calls across 70+ languages. Merge it into any call or use speakerphone; no app, no account. Read the announcement.
  • Flux Interpreting is now ThunderPhone Interpreting — the same white-label over-the-phone interpreting platform for language service providers, under the ThunderPhone name.
  • Added a usage-spike monitor on the TranslateMyCall public line.

v2.0.0 — August 24, 2026

  • ThunderPhone is open to everyone: self-serve signup is live — build a voice agent, connect telephony, and go to production without talking to us first. Read the announcement.
  • Launched a public status page: status.thunderphone.com.
  • Added optional keyboard sounds while user-defined tools run.
  • Tool-wait announcements now match what the tool is actually doing.
  • Improved naturalness of acknowledgement behavior.
  • Improved Storm spelling readbacks.
  • Opening-turn completions now recover from content-filter rejections instead of leaving the call silent.
  • Realtime API: exactly one completed event per caller item, and late transcript corrections remain in call history rather than arriving as deletions.

v1.9.2 — August 17, 2026

  • Realtime API: added session.live_transcripts — caller transcript segments stream over the session socket as they finalize, instead of arriving only in the post-call record.
  • Moved the premium voice lineup to the latest synthesis model generation; Urdu voices re-cast.
  • Multilingual agents now switch voice language in the same turn the caller switches, not one turn late.
  • Exposed the silent-line hangup timeout as an agent builder setting.
  • Outbound greeting now yields when a live reply wins the pickup race — no more double-talk when a human answers mid-greeting.
  • Fixed a consent-gate redirect loop for invited members: the governing org now carries through the acceptance flow.

v1.9.1 — August 12, 2026

  • Added send_email agent tool — mid-call email delivery to a configured recipient.
  • Higher answer quality on turns that follow substantive tool results.
  • Spelling readback got faster and more accurate on names, addresses, and confirmation codes.

v1.9.0 — August 6, 2026

  • POST /v1/call fast-returns with a durable dial record and idempotent retries; dial progression continues async.
  • Made max_duration authoritative: the agent wraps the call before the cap fires rather than the media layer hard-dropping.
  • Exactly-once tool-call/response pairing — an orphaned tool call can no longer wedge delivery of later messages.
  • Added additional_audio_context (tri-state) controlling how much prior-turn audio the model hears; surfaced in builder advanced settings.
  • Outbound pre-conversation audio is now anchored on SIP answer supervision instead of wall-clock timing.
  • Hardened how integration credentials are stored.

v1.8.2 — August 5, 2026

  • Shipped docs in all 47 supported languages — 1,100+ localized pages with hreflang pairs and localized sitemaps.
  • Recording disclosure defaults on, per-agent override with an effective-state readout; the runtime fails safe to "disclose" on any config ambiguity.
  • Cut streaming first-audio latency at call start.

v1.8.1 — August 3, 2026

  • Per-agent recording opt-out; the consent announcement is muxed into retained recordings so the artifact is self-documenting.
  • Added a one-time per-org TCPA confirmation gating outbound calling features.
  • Fixed Cantonese being voiced with a Mandarin accent.
  • Tuned CJK responses down to conversational length — long-form answers read badly at TTS pace.
  • Synthesis stalls no longer leave dead air — the turn recovers automatically.

v1.8.0 — July 31, 2026

  • Custom voice cloning.
  • Regional consent regimes: announcement opt-in for EEA/UK/CH, default-on with opt-out elsewhere, resolved per call leg.
  • Retention: deletion cascade, erasure tombstones, and a scheduled dsar_sweep job for data-subject requests.
  • ToS v2.0 with versioned, evidenced acceptance records.
  • Support presence: heartbeat-based session presence with proactive support chat.

v1.7.4 — July 29, 2026

  • Knowledge bases: URL import behind an SSRF-guarded server-side fetcher.
  • Added the forgot-password flow.
  • Dashboard restyled to the new design system.
  • Widget 1.2.1: mic no longer stays hot during ringback.

v1.7.3 — July 28, 2026

  • Redesigned website.

v1.7.2 — July 27, 2026

  • Premium voice tier with an accent-grouped catalog.
  • Premium synthesis failures now fail over to the standard voice mid-turn instead of dropping the turn.
  • Fixed PCM frame misalignment in streamed synthesis output.

v1.7.1 — July 24, 2026

  • Tool-wait reassurance: the agent narrates progress on long tool calls instead of going silent; placeholder cadence tuned to kill filler loops.
  • Drag-and-drop uploads for knowledge bases.
  • macOS app auto-updates.
  • Fixed a double-greeting race at call start.

v1.7.0 — July 22, 2026

  • Dashboard and site localized into 47 locales: namespace-split catalogs, ICU plural rules (including Arabic's six categories), locale cookie, per-account preference.
  • Transactional emails and system prompts localized to match.

v1.6.5 — July 22, 2026

  • Stored provider credentials re-encrypted under rotating keys, rotated online with no downtime.
  • Session tokens now revoke on password change/reset; added an admin break-glass path.
  • Per-device sessions with individual revocation.
  • Billing pipeline hardening: spend caps enforced under concurrency, idempotent auto-reload, correct handling of out-of-order refund reversals.

v1.6.4 — July 20, 2026

  • Security hardening sweep across authorization scoping, export validation, and outbound fetch guards.
  • DTMF digits are redacted from logs and stored artifacts.
  • Webhook egress is IP-pinned and SSRF-guarded.
  • Call exports now walk the full filtered set instead of the first page.

v1.6.3 — July 17, 2026

  • Org spending limits enforced at call admission with reservation semantics — calls are declined before budget is exceeded, not billed past it.
  • Realtime WebSocket API: bidirectional audio and event frames for building custom call frontends.
  • Webhook platform: HMAC signatures over canonical payload bytes, exponential-backoff retry ladder, coverage across all call events.
  • Idempotent top-ups and stricter payment-webhook verification.
  • Added a live prompt-size meter to the agent builder.

v1.6.2 — July 16, 2026

  • Client portals: scoped read-only access for your clients via emailed login codes.
  • In-app assistant: voice mode, guided walkthroughs that drive the UI step by step, human handoff.
  • Knowledge bases: document upload with chunked ingestion and grounded retrieval.
  • Per-agent connect tones and progress sounds.

v1.6.1 — July 13, 2026

  • In-dashboard number lifecycle: search → order → provision → assign.
  • New agent options: silence check-in kill switch; connect tone at call start.
  • Provisioning failures now return structured JSON errors instead of raw 500s.
  • Raw-audio lane recovers mid-call when its stream reader dies.

v1.6.0 — July 9, 2026

  • Warm (attended) transfers: hold leg, conversational consent on both sides, merge-on-consent.
  • iPhone home-screen install (PWA).
  • Acknowledgements now dispatch immediately when the caller stops speaking.

v1.5.3 — June 27, 2026

  • Suppressed TTS narration of keypad tones.
  • Storm tool-calling gains a model fallback path for upstream slowdowns.

v1.5.2 — June 18, 2026

  • Fewer mid-thought cutoffs on Storm: end-of-turn detection improved beyond simple pause thresholds. Benchmarked before rollout.
  • Configurable response cues on Storm.

v1.5.1 — June 15, 2026

  • Fixes to hold handling and IVR menu navigation on outbound calls.
  • Website refresh.

v1.5.0 — June 9, 2026

  • Improved spelling readback accuracy on ambiguous spans.
  • Language detection stays active for the full call; added a short-utterance switch detector and lowered the configured-language confidence floor.
  • Fixed autosave for additional agent languages.

v1.4.3 — May 28, 2026

  • TTS holds while incoming audio is still unconfirmed as speech — no more talking over a hesitating caller.
  • Assorted call-quality fixes.

v1.4.2 — May 21, 2026

  • Rebuilt VAD around multiple acoustic signals — improves the precision/recall tradeoff of speech detection in noisy environments.
  • VAD spans now emit complete — no more truncated leading edges on utterances.
  • Fixed greeting-echo self-barge-in: the agent no longer interrupts itself on its own opening line.

v1.4.1 — May 13, 2026

  • Voice system overhaul: rebuilt catalog with consistent naming.

v1.4.0 — May 5, 2026

  • Live transcript streaming to the dashboard during in-progress calls.
  • Automatic language detection with mid-call switching.
  • A/B experiment framework: run agent variants side by side with outcome comparison.

v1.3.1 — April 28, 2026

  • Automatic issue investigation: flagged calls get a structured, reviewable investigation record.
  • Draft deploys: stage agent edits against an init snapshot, publish atomically.
  • Settings redesign; guided onboarding coach.
  • Call-success scoring redefined around task completion instead of call completion.
  • Configurable end-of-call silence tail.

v1.3.0 — April 23, 2026

  • Scenario generation: simulated callers with generated personas and goals, runnable against an agent pre-launch.
  • Test-call and issues-page overhaul.

v1.2.1 — April 21, 2026

  • Call history export.
  • API documentation rewrite.
  • On-request mid-call language switching.
  • Acknowledgements localized to the caller's language.
  • Composable, saveable filters across calls and numbers.

v1.2.0 — April 17, 2026

  • Agent builder redesign.
  • Multilingual agents switch languages mid-conversation without a re-prompt.
  • Voice catalog renamed for consistency.
  • Storm output restructured for complex requests.

v1.1.1 — April 14, 2026

  • Grading signal-quality pass: missing transcripts no longer grade as failures; grader inputs deduplicated.
  • Pipeline reports across call batches.

v1.1.0 — April 8, 2026

  • Phone number labels.
  • Transfers and goodbyes are interruptible without corrupting call state.
  • Per-agent endpointing configuration for IVR navigation.

v1.0.4 — March 30, 2026

  • Call history detail view upgraded.
  • Configurable completion prompt on outbound calls.
  • Storm degrades gracefully under slow upstream APIs instead of stalling the turn.

v1.0.3 — March 23, 2026

  • DTMF send: agents press keys to drive IVRs and enter codes, with per-digit acknowledgement.
  • Hold state with an explicit wait-on-hold trigger — the agent waits out hold music and resumes on a human voice.
  • Configurable speak/act ordering.
  • Fixed silent follow-ups after DTMF entry.

v1.0.2 — March 18, 2026

  • New HD voices.
  • Widget call-cost display fixed.
  • Stale in_progress calls now detected and cleaned up.

v1.0.1 — March 16, 2026

  • Call connect latency reduced.
  • Widget: ringback no longer bleeds into the session audio.
  • Phone-number management API.

v1.0.0 — March 10, 2026

  • Full call recordings, archived and playable from the dashboard.
  • Browser test calls.
  • Teammate invites through signup.
  • Per-call cost breakdowns.

v0.9.2 — March 8, 2026

  • Bolt streams TTS as tokens are generated instead of waiting on the full completion — cuts time-to-first-word.
  • Bolt transcription got faster and more robust to transcriber slowdowns.
  • Passkey sign-in with org-level enforcement.
  • Call history shows real billed cost per call.

v0.9.1 — March 5, 2026

  • Widget fast-connect plus built-in noise cancellation for browser calls.
  • Reduced Spark turn latency and improved recovery from failed turns.
  • Fixed sentence restarts after long silence tails.

v0.9.0 — March 2, 2026

  • Multi-language conversations.
  • Production-readiness hardening pass.

v0.8.0 — February 26, 2026

  • Added redundancy across model backends — single-backend failures no longer surface mid-call.

v0.7.1 — February 25, 2026

  • Configurable reasoning depth on Storm.
  • Tool-call placeholders: the agent acknowledges before executing longer actions.
  • Hangup timing recorded to the precise disconnect moment.
  • Webhook delivery dedup.

v0.7.0 — February 20, 2026

  • Consent controls for callers and users.
  • Token-streamed TTS path — synthesis starts before the sentence is complete.
  • Background ambiance decoupled from agent voice gain.

v0.6.2 — February 18, 2026

  • Prompt regeneration from agent config.
  • Fixed provisioning button states.

v0.6.1 — February 12, 2026

  • Developer docs launched.

v0.6.0 — February 11, 2026

  • Embeddable widget: a voice agent on your own site via script tag, authenticated with publishable keys.
  • Introduced the three-engine lineup — Spark, Bolt, Storm — differing in model size, latency, and per-minute rate.
  • Integration-based tool routing: tool calls resolve through the connected integration's endpoint.

v0.5.0 — February 10, 2026

  • The ThunderPhone dashboard: build, test, and manage agents from the browser.
  • Organizations and teams with email invites.
  • BYO-carrier SIP trunking and unified number management.
  • Integrations UI.
  • Ambient background sound options.
  • Google OAuth sign-in.

v0.4.2 — January 27, 2026

  • Call-transport hardening after a series of edge-case disconnects.
  • Acoustic-event classification moved off the response hot path — noise detection no longer adds latency.

v0.4.1 — January 14, 2026

  • Voice lineup expanded.
  • Fixed an audio-pump stall that could freeze playback mid-turn.
  • Reduced latency on the live-response path.

v0.4.0 — January 6, 2026

  • Usage logging and per-call cost reporting.
  • Collapsed per-turn model round trips into a single request, removing one network round trip from every reply.
  • Fixed dropped transcript spans.

v0.3.3 — December 18, 2025

  • Completion no longer blocks on a slow transcriber.
  • Transcript finalization timeout fixes.
  • Fixed artifacts in merged call audio.

v0.3.1 — December 12, 2025

  • Added an acoustic noise classifier to cut false speech triggers from background noise; tuned its confidence threshold against recorded calls.
  • End-of-call detection: added farewell-intent and post-farewell-silence signals to improve the precision/recall tradeoff — fewer premature hangups, fewer lingering dead lines.
  • Added a lenient VAD mode for quiet callers — lower onset threshold at the cost of some false accepts.
  • Faster acknowledgement dispatch.

v0.3.0 — December 4, 2025

  • Realtime transcription.
  • Recordings consolidated to a single mixed audio file per call.
  • Latency fixes across the pipeline.

v0.2.4 — November 24, 2025

  • Acknowledgements are checked against the pending completion so they can't contradict the final answer.
  • Raised the acknowledgement timeout to stop premature re-prompts.
  • Fixed audio glitches in longer replies.
  • Fixed handling of long caller silences.

v0.2.3 — November 20, 2025

  • Call recording.
  • Hold music during waits and transfers.
  • First version of Spark: moved simple call flows onto a smaller, lower-latency model.
  • Phone number availability expanded.

v0.2.2 — November 14, 2025

  • Call-completion webhooks now wait on finalized call state before firing.
  • Cut model reasoning effort on simple turns to reduce response latency.

v0.2.1 — November 11, 2025

  • Barge-in: the agent halts speech on caller speech onset.
  • Bounded runaway audio buffers.
  • Response latency reduced.
  • New voices.

v0.2.0 — November 5, 2025

  • Acknowledgements now condition on the caller's request instead of using generic filler.
  • Call-end webhook improvements.
  • Faster recovery when a turn stalls past five seconds.

v0.1.4 — October 31, 2025

  • Interruption handling.
  • The agent finishes its sign-off before hanging up — no more mid-sentence cutoffs.
  • Phone numbers formatted properly for speech.

v0.1.3 — October 29, 2025

  • Webhooks for call events.
  • Mid-call recovery on connection blips.

v0.1.2 — October 28, 2025

  • Spelling readback for names and confirmation numbers.
  • Reduced false speech triggers from background noise.
  • More voices and accents.

v0.1.0 — October 8, 2025

  • First release: AI voice agents on real phone calls, inbound and outbound.