This page lists all sub-processors Stagee uses to deliver the service. Sub-processors are third-party service providers that process personal data on our behalf - for example, the database provider or the speech recognition engine.
We publish this list because it's the right way to operate, and because our Data Processing Agreement (DPA) requires us to notify customers in advance of changes.
Current list
| Sub-processor | Function | Data category | Location | Safeguards |
|---|---|---|---|---|
| Supabase | PostgreSQL database - account config, call event records, phone hashes, admin events. | Account data, call events (no audio), phone hashes. | Ireland (EU) - eu-west-1 | DPA signed (PandaDoc, 12 Mar 2026) · AES-256 at rest · Daily backups. |
| Vercel | Hosting for website and client portal. | Portal navigation data · No call data. | Ireland (EU) - Dublin (eu-west-1) | DPA via ToS. |
| Upstash (Redis) | Rate limiting for the portal and APIs. | Rate-limit counters keyed by identifier (IP address). No call data. | EU (to confirm) | DPA via ToS. |
| Soniox | Real-time speech recognition (primary, stt-rt-v5) - call audio to text, via LiveKit Inference. | Call audio (streaming, not stored). | EU (via LiveKit Inference) | Direct DPA signed (separate from the LiveKit addendum) · EU processing · Streaming only. |
| AssemblyAI | Real-time speech recognition (fallback, u3-rt-pro) - call audio to text, via LiveKit Inference. | Call audio (streaming, not stored). | EU (via LiveKit Inference) | Under LiveKit addendum · EU processing · Streaming only. |
| Google (via LiveKit Inference) | Concierge language model (Gemma, gemma-4-31b-it) - semantic understanding and action decisions, via LiveKit Inference. | Transcribed call text (no audio). | EU (via LiveKit Inference) | Under LiveKit addendum · EU processing · No training on our data. |
| OpenAI (via LiveKit Inference) | Triage and manager models (gpt-4o-mini) - intent classification and transfer, via LiveKit Inference. | Transcribed call text (no audio). | EU (via LiveKit Inference) | Under LiveKit addendum · EU processing · No training on our data. |
| Cartesia | Voice synthesis (Sonic 3.6, pt-PT) - agent text to speech, via LiveKit Inference. | Agent response text (no caller audio). | EU (via LiveKit Inference) | Under LiveKit addendum · EU processing. |
| LiveKit | Real-time voice orchestration and Inference gateway - coordinates STT, LLM, TTS at low latency. | Audio and text in transit (not persisted in LiveKit). | Frankfurt (EU) | DPA signed (addendum v. 18 Nov 2025) · SCCs. |
| LiveKit Observability | Capture of audio, transcripts and traces for agent debugging. | Call audio and transcript. | Frankfurt (EU) | Public DPA via Trust Center · EU-hosted · Purpose limited to agent debugging. |
| Langfuse | Observability - per-turn traces (prompt and completion text) exported via LiveKit OTel. | Conversation content (per-turn prompt and completion). | EU (Langfuse Cloud EU) | EU-hosted · Purpose limited to observability · DPA to confirm. |
| Telnyx | SIP telephony and SMS - receives calls on agent number, routes to LiveKit, sends confirmation SMS and, when the restaurant enables it, the WhatsApp/SMS unresolved-call alerts to the team. | Caller phone number, audio in transit, SMS and WhatsApp messages (alert content: caller number, name if given, reason and time). | Germany (EU) - EU data location | DPA incorporated in Terms of Service · SCCs · ANACOM-licensed operator. |
| Resend | Transactional email - account email verification and password reset, staff invitations, and event-lead notifications. | Names and emails of restaurant owners and staff; names and emails of event leads. | USA | DPA incorporated in ToS · SCCs · No training on our data. |
| Meta Platforms Ireland Limited (WhatsApp) | WhatsApp network delivering the alerts to the restaurant team when a call ends unresolved (optional, enabled per restaurant; sent through Telnyx as Business Solution Provider). | Caller phone number (never shown when withheld), guest name if given, reason and time of the call; the team's destination number or group. No transcript or recording. | EU (Ireland); possible transfers to the USA | Meta WhatsApp Business Terms and Data Processing Terms · SCCs · Purpose limited to calling the guest back about that same call. |
| Stripe | Subscription payment processing. | Restaurant payment data (cards on Stripe infra, not stored by Stagee). | Ireland (EU) | DPA via ToS · PCI-DSS Level 1. |
| Cloudflare R2 | Call audio storage - private bucket with EU jurisdiction, signed-URL access only. | Call audio files (OGG). 90-day retention with automatic deletion. | EU jurisdiction | DPA via Cloudflare ToS · AES-256 at rest · No public access · Short-lived signed URLs. |
The speech and language engines (Soniox, AssemblyAI, Google, OpenAI, Cartesia) are reached through LiveKit Inference, with region restriction enabled (EU processing). On that path LiveKit is the direct sub-processor and the models operate under LiveKit's addendum, except Soniox, which has a direct DPA signed separately.
Additional sub-processors
Used for internal Stagee operations (code management, dev tools) and do not touch call personal data:
- GitHub (USA) - code repository. No personal data.
How we communicate changes
When we add, remove, or significantly change a sub-processor: we update this page, notify active customers by email 30 days in advance, and allow objection. Reasonable objections that cannot be resolved entitle the customer to terminate without extra cost.
Change log
- 12 October 2026: aligned the model stack with production - primary speech-to-text Soniox (stt-rt-v5) with AssemblyAI fallback; concierge LLM Google Gemma (gemma-4-31b-it) and triage/manager OpenAI (gpt-4o-mini); TTS Cartesia (sonic-3.6); all via LiveKit Inference in the EU. Added Langfuse (observability, per-turn traces) and Upstash (rate limiting). Soniox has a direct DPA; the other models run under the LiveKit addendum.
- 19 October 2026: added Meta Platforms Ireland Limited (WhatsApp, via Telnyx) for the optional unresolved-call alerts to the restaurant team (caller number, name if given, reason and time; no transcript or recording). Records deleted after 30 days.
- 24 September 2026: corrected the Vercel location to Ireland (Dublin, eu-west-1), with the function region pinned in vercel.json; removed Anthropic Claude (the offline QA judge was disabled via OBSERVABILITY_DISABLED) and OpenAI (embeddings) (the knowledge base is supplied to the model in full, no embeddings); LiveKit Observability is now in the EU (Frankfurt) after the LiveKit project migration.
- 22 September 2026: corrected Supabase location (Ireland eu-west-1, not Frankfurt); added Cloudflare R2 (call audio storage); updated Cartesia and Telnyx DPA status to "incorporated in Terms of Service"; replaced OpenAI (LLM) with Microsoft Azure (OpenAI models, via LiveKit Inference, EU); OpenAI retained for offline embeddings only; updated AssemblyAI to Universal 3, 3.5 and 3.6.
- 21 May 2026: location updates (Supabase EU Frankfurt, LiveKit EU Frankfurt, Telnyx Germany EU), inclusion of LiveKit Observability, reclassification of Anthropic Claude as a production sub-processor, DPAs signed with Supabase and OpenAI.
- 19 May 2026: initial version.
Contact
Questions about sub-processors: hello@stagee.ai.