Sub-processors
Every third party that processes data on our behalf, grouped below by core infrastructure, then LLM, speech-to-text, and text-to-speech. Each publishes a standard Data Processing Addendum (we haven't separately executed one with each yet — see our compliance page). Where a provider offers a no-training or zero-retention opt-out, we enable it by default.
There's always the option to bring your own API key for any of these providers — then you control these settings yourself.
Bring your own keys (BYOK)?Those opt-outs are yours to set — we don't change the no-training or retention settings on your own provider accounts.
For how we protect your data — encryption, hardened headers, and our A-rated security posture — see the security page.
Last reviewed: 2026-06-21 · Live status read from each provider's status page (refreshed every few minutes).
Core processors
Telephony, billing, auth, hosting, database, cache, and email — the infrastructure providers. Each publishes a standard Data Processing Addendum; we haven't separately executed one with each yet.
| Provider | Purpose & data | Region | Status |
|---|---|---|---|
| Twilio Does not train on customer data; IE1 region available for EU residency. | Phone numbers, call/SMS transport, recordings. Phone numbers, call metadata, audio, SMS content. | United StatesIreland (IE1, for EU) | Minor issues |
| Stripe PCI-DSS certified. | Billing and card processing. Billing contact + payment details (card data handled by Stripe, not stored by us). | United States | Operational |
| Clerk | User authentication & sessions. Account email, name, auth credentials. | United States | Minor issues |
| Vercel | Hosts the web app & APIs. Request data in transit; no primary data store. | United States | Operational· 100.00% · 30d |
| Hetzner | Hosts the primary real-time voice runtime and automation tooling. Live call audio in transit during a call; no persistent storage on the box. | European Union (Germany)United States (planned) | Status |
| Render Sessions are in-memory, not persisted. | Backup/fallback for the real-time voice runtime. Live call audio in transit during a call. | United States | Operational· 100.00% · 30d |
| Supabase Encrypted at rest. | Primary application database. Account, agent config, transcripts/recordings metadata (per retention). | United States | Operational |
| Upstash | Rate limiting & ephemeral cache. IP/identifier counters; no conversation content. | EU (Frankfurt)United States (per tenant) | Operational |
| Resend | Transactional & marketing email. Recipient email address & message content. | United States | Operational |
LLM
| Provider | Purpose & data | Region | Data handling | Status |
|---|---|---|---|---|
| OpenAI Logging disabled (store=false); ZDR requires BYOK. | Generates agent responses (selectable model). Conversation text sent during a turn. | United States | No training ZDR | Operational |
| Anthropic (Claude) 7-day default retention; ZDR requires BYOK. | Generates agent responses (Claude). Conversation text sent during a turn. | United States | No training ZDR | Operational |
| Mistral AI EU-hosted, not subject to US CLOUD Act. | Generates agent responses — strongest EU fit. Conversation text sent during a turn. | European Union (Paris) | No training ZDR | Status |
| Google (Gemini) Under our current API settings, may be used to improve Google's products outside the EEA/UK/CH. | Generates agent responses (Gemini, selectable). Conversation text sent during a turn. | Global (EEA/UK/CH: no-train regardless of tier) | No training ZDR | Status |
| Groq Zero Data Retention enabled on our account. | Low-latency LLM inference (selectable). Conversation text sent during a turn. | United States | No training ZDR | Operational |
| OpenRouter Routes only to non-training / ZDR endpoints. | Model router for additional LLMs. Conversation text sent during a turn. | United States | No training ZDR | Status |
| xAI (Grok) 30-day default retention; ZDR requires Enterprise + BYOK. | Generates agent responses (Grok, selectable). Conversation text sent during a turn. | United States | No training ZDR | Status |
Speech-to-text
| Provider | Purpose & data | Region | Data handling | Status |
|---|---|---|---|---|
| Deepgram Model Improvement opt-out applied; SOC 2 Type II. | Transcribes caller speech in real time. Call audio + resulting transcript. | United States | No training ZDR | Operational |
| AssemblyAI EU residency available; zero retention requires BYOK; SOC 2 Type II. | Transcription (fallback / selectable). Call audio + resulting transcript. | United StatesEU (Dublin) | No training ZDR | Operational |
Text-to-speech
| Provider | Purpose & data | Region | Data handling | Status |
|---|---|---|---|---|
| ElevenLabs Third-party training contractually prohibited; EU residency available. | Synthesizes the agent's voice. Response text + generated audio. | United StatesEU (available) | No training ZDR | Operational |
| Cartesia In-region inference; GDPR, SOC 2 Type II / HIPAA. | Synthesizes the agent's voice (selectable). Response text + generated audio. | United Statesin-region (available) | No training ZDR | Operational |
✓ = enabled on our account. No training: your data is never used to train the provider's models. ZDR (zero data retention): the provider stores no inputs/outputs (enabled where available; some offer it only on enterprise plans).
Questions about data processing or a copy of our DPA? Contact us. See also our Privacy Policy.