Trust & data handling
What happens to your data in the PAIRL Gateway, written for the people who have to sign off on it. Current as of October 2026.
How a request travels
- To the gateway. Your application sends the request to api.pairl.dev over an encrypted connection, with your PAIRL key and your provider key.
- Choosing excerpts. When a conversation is worth compressing, its text goes to an encoder model that picks what to keep. Today that is the Google Gemini API, on PAIRL’s own Google account, even if your target provider is Anthropic or OpenAI. Plain code then checks every excerpt against the source.
- Tool results. Older tool results are compacted by plain code, without a model.
- To your provider. The smaller request goes to your provider with your key, under your contract.
- Back to you. The response comes back as your provider sent it. Only the input token count differs: it shows the size before compression.
To keep a request away from the encoder, send X-PAIRL-Mode: off. The request then passes through without compression and without a fee.
What is stored
- Usage records. Request ID, time, key ID, provider, model, token counts, savings, fee and mode. No message content. They power your dashboard and billing.
- Compressed history. Compressed excerpts of your conversations stay in the gateway’s cache for up to 30 days, so that later turns don’t have to be encoded again. They contain original wording from your conversations and have no additional encryption of their own.
- Provider keys. Used to forward the request. Never written to a database or a log.
- Full requests. Off by default. If you turn on debug retention, requests are stored encrypted for the time you set (1 to 168 hours), then deleted.
- Conversations for improving compression. Off by default. If you opt in, encrypted copies of compressed conversations are kept to train our own encoder. Opting out permanently deletes them.
- Model responses. Not stored.
Who is involved
- Hetzner Online GmbH, Germany. Runs the gateway, its database and its cache (Nuremberg).
- Google LLC, USA. Gemini API for choosing excerpts.
- Your AI provider. Receives the compressed request under your own contract.
- enum. Stores encrypted database backups (Frankfurt, 35 days).
- Cloudflare, Inc., USA. Hosts this website and the console.
- Stripe. Processes payments.
- Brevo (Sendinblue GmbH), Germany. Sends account emails.
Dedicated deployments for regulated teams run on their own infrastructure. Their recipients, including the encoder, are agreed by contract.
Checking what an agent read
Every excerpt is an exact copy of your source, so it can be traced back to the original conversation. By default, the gateway does not keep the compressed context that went to the model. To inspect it case by case, turn on debug retention. PAIRL is not an audit log: keeping records stays with your own systems.
Availability
- PAIRL is in beta. Service levels are part of Enterprise agreements only, and there is no public status page yet.
- If compression fails, the request goes through unchanged and without a fee. You can switch this to fail closed through the Config API.
- If the gateway itself can’t be reached, requests fail. Keep your provider’s direct URL as a fallback in your client.
- If your credit balance falls below −$5, compression pauses and requests pass through unchanged.
Certifications and contracts
PAIRL holds no security certification such as SOC 2 or ISO 27001 today. It is operated by Bureau Wehrmann / WMNN in Germany. For a data processing agreement under Art. 28 GDPR or a security questionnaire, write to [email protected].