Data handling and LLM providers
This page describes what data an agent can touch, where that data can go, and how prompt data is handled by the large-language-model providers your agents call.
The boundary
An agent's outbound traffic is governed by a declared, default-deny egress policy. Prompt and tool data transits only the hosts you put on the allow list. The credential used to reach a model provider is brokered by the gateway and never enters agent code. This is the same governed egress boundary as any other host the agent calls. See How enforcement works.
Which providers your prompts go to
The model provider, model, and credential binding are configured per agent. Prompts are sent to the provider endpoint you configure. Nothing is fanned out to providers you did not declare. To restrict an agent to a single provider, keep only that provider's host on the egress allow list.
Who controls whether your data trains a model
Training-use of your prompt data is a setting at your inference provider, not something AgentPaaS can enforce for you.
If you route through OpenRouter (or another inference layer or model provider), configure that provider's data controls to opt out of training on your data. AgentPaaS controls where your data is allowed to go; it cannot set the retention or training policy inside the provider's systems.
If you want AgentPaaS to help ensure your inference layer is configured not to train on your data, contact support and we will help you set it up.
What AgentPaaS does and does not claim
- Does: route prompt/tool data only to declared hosts, broker the credential, and record the access in the signed audit chain.
- Does not: set or guarantee the training/retention policy of an external model provider. That is a customer-controlled setting at the provider.
Retention
AgentPaaS is a platform service. We do not retain or use customer data for any training purpose, and we store only what is needed to operate the platform as a service. Stored data includes run metadata, the signed audit chain, and the brokered secrets you push. We do not use your prompts, files, or run content to train any model.
This is separate from the model provider's own data policy, covered above: what an external provider does with prompt data is governed by your agreement and settings with that provider, not by AgentPaaS.