Our Subprocessors
Last updated: November 9, 2026
To provide our services, we use the third-party companies below as subprocessors. For each one, the table shows what we use it for, where it processes data, and whether it can receive data that Toffu reads from your connected platforms (for example ad accounts, analytics, or an online store).
Each subprocessor is bound by a written data processing agreement that is at least as protective as our DPA. International transfers are covered in Section 6 of the DPA.
| Subprocessor | Purpose | Processing location | Receives connected-platform data |
|---|---|---|---|
| Anthropic | Primary AI model provider. Processes prompts and tool results to generate answers and actions. | United States | Yes, the data needed for each request (after shopper personal data is removed from store connectors). |
| AI models (Gemini on Vertex AI) for specific tasks and as a fallback; Google Analytics for website analytics. | Vertex AI: global endpoint, Google selects the region. Google Analytics: United States. | Vertex AI: yes, the data needed for each request. Google Analytics: no. | |
| OpenAI | AI models for specific features (for example images and transcription) and as a fallback. | United States | Yes, only for the requests routed to it. |
| Amazon Web Services | Application hosting, file storage, server logs, cache, and Anthropic models on AWS Bedrock as a fallback. | United States (us-west-2, Oregon) | Yes |
| MongoDB Atlas | Primary database: accounts, conversations, reports, and settings. Runs on AWS. | United States (AWS us-west-2, Oregon) | Yes |
| Render | Edge proxy for report links and standby compute. | United States (Oregon) | Yes, in transit. |
| E2B | Isolated sandboxes that run analysis scripts. Each sandbox is deleted when its run ends. | United States | Yes, during a script run only. |
| Cloudflare | Network security, DNS, and TLS for toffu.ai traffic. | Global edge network | Yes, in transit only (encrypted traffic passes through its network). |
| Sentry | Error tracking. Sensitive keys are scrubbed before an event is sent. | United States | Limited: an error report can include a fragment of the request that failed. |
| PostHog | Product analytics: which features are used. | United States | No |
| Frontegg | User login and identity (names, email addresses, roles). | United States | No |
| Vercel | Hosts the toffu.ai website and web app code. | United States | No. Your browser fetches workspace data from our API directly. |
| Stripe | Payment processing. We do not store card details. | United States | No |
| Slack | Internal team messages, including summaries of activity in unclaimed workspaces (the name or email of the login used to connect a platform). | United States | No |
AI providers: training and retention
- No AI provider may use your data to train or improve a model. We use each provider's commercial API, whose terms prohibit it (DPA Section 4).
- Anthropic (our primary model provider): zero data retention. Anthropic does not store API inputs and outputs after the response is returned, except where required by law or to combat misuse.
- Anthropic models on AWS Bedrock (fallback when Anthropic's API is unavailable): AWS does not store or log prompts and outputs.
- Google (Vertex AI) and OpenAI: retention follows each provider's published API terms, which allow short-term retention of inputs and outputs for abuse monitoring (up to 30 days).
Their own subprocessors
Some of our subprocessors use subprocessors of their own, mainly cloud infrastructure. Each publishes its list:
This list is subject to change, and we will update it as our use of subprocessors evolves.