Privacy and data policy

This page describes what Pylo collects when you send a request, how long we keep it, and what we do with it. Pylo is an OpenAI-compatible inference provider. We route your request to an upstream model backend and stream the response back to you.

Read this alongside our Acceptable Use Policy and Terms. If anything here conflicts with a signed agreement, the signed agreement controls.

Pylo retains request metadata for security and legal compliance, and may keep a short-lived sample of request and response content to test reliability. We do not train on your data and we do not sell it. Pylo is not zero-data-retention.

What we collect

For every request, we collect request metadata and audit trace identifiers. This metadata records what happened, not what you wrote. It does not include the content of your prompts or the model's responses.

Trace identifiers include the X-Request-Id we assign to every inbound request. That value is the idempotency key for the request and lets us trace a single request end to end across a failover retry.

The audit log

Every request through Pylo writes one append-only line to an audit log. This log records metadata only. It does not contain the content of your prompts or the model's responses.

Each line stores exactly the following fields:

FieldWhat it holds
internal user identifierthe account the request was authenticated under
timestampwhen the request was received
endpointthe API endpoint the request hit, for example /v1/chat/completions
methodthe HTTP method, for example POST
modelthe model the request targeted
http_statusthe response status code
X-Request-Idthe trace identifier for the request
errorpopulated on failures, null otherwise
timingslatency markers such as time to first token and end to end duration
token countsprompt and output token counts for the request

The log is append-only. Lines are not edited after they are written.

Why we retain it

We keep audit metadata to detect and investigate abuse and to meet our legal and compliance obligations. We keep it while it serves those purposes. Pylo is not a zero-data-retention provider.

Reliability and quality

To keep the service fast and correct, we may keep a limited, short-lived sample of request and response content. We use it only to test reliability, reproduce bugs, and check output quality. We do not train models on it and we do not sell it. We delete these samples within 30 days.

What we do not do

Upstream backends

Pylo routes each request to an upstream inference backend to fulfill it. To serve your request, the request data is sent to that backend. Those backends are sub-processors and apply their own data policies to the request they receive. Pylo selects the backend and streams the response back to you over the same connection.

Making a data request

To make a data request or ask what data we hold about you, use the contact form at /contact and pick Technical support. Include the X-Request-Id values relevant to your request so we can locate the matching audit lines.

Governing terms

Operated from
United States
Governing law
State of Delaware, USA
Effective
2026

We may update this policy. Material changes will be reflected in the effective date above.