Subprocessors

Last updated 8 October 2026

Cailos routes your API requests to third-party LLM providers ("subprocessors") based on your routing configuration. The content of your requests is processed by these providers according to their own privacy policies and terms of service.

Before any request reaches a subprocessor, Cailos applies PII shielding (via LLMShield) to detect and cloak personally identifiable information in transit. However, the semantic content of your prompts is visible to the routed provider.

LLM inference providers

Your requests may be routed to any of the following providers depending on your model selection, optimisation strategy, and endpoint availability:

Provider Purpose Location Policy
OpenAILLM inferenceUSPrivacy policy
AnthropicLLM inferenceUSPrivacy policy
Google (Gemini)LLM inferenceUSAPI terms
CohereLLM inferenceCAPrivacy policy
GroqLLM inferenceUSPrivacy policy
Together AILLM inferenceUSPrivacy policy
DeepInfraLLM inferenceUSPrivacy policy
ScalewayLLM inferenceFR (EU)Privacy policy
CerebrasLLM inferenceUSPrivacy policy
NCompassLLM inferenceUSPrivacy policy
NScaleLLM inferenceEUPrivacy policy
InceptionLLM inferenceAEPrivacy policy
OpenRouterLLM inference routingUSPrivacy policy
Novita AILLM inferenceUSPrivacy policy
SambaNovaLLM inferenceUSPrivacy policy
TinfoilLLM inference (secure enclave)USPrivacy policy
xAILLM inferenceUSPrivacy policy
BytePlusLLM inferenceSG (AP-Southeast)Privacy policy
CloudflareLLM inferenceGlobalPrivacy policy
DarkbloomLLM inference (private hardware-attested Apple Silicon)GlobalHomepage

Request classification

When a team gives a custom model routing instructions, each request to that model is classified to apply them: the latest user message, after PII shielding, is sent to the provider below, which returns which kind of request it is and how demanding it is. Requests to other models are not sent.

Provider Purpose Location Policy
TypeSafe AIClassification for custom model routing instructions (Jev)USPrivacy policy

Infrastructure providers

Provider Purpose Location
RailwayApplication hosting, PostgreSQL database, Redis cachingEU
CloudflareDNS, CDN, DDoS protectionGlobal
AxiomStructured log ingestion and observabilityEU

PII shielding

Cailos uses LLMShield to detect and cloak PII before requests reach any LLM inference provider. Email addresses, phone numbers, credit card numbers and IP addresses are replaced with tokens before transit and restored in the response, so the upstream provider never sees them. Other personal data, such as names and postal addresses, is not cloaked and is sent to the provider as written.

Trust levels

Each endpoint in Cailos is assigned a trust level (0–3) based on the provider's data handling practices:

0LowMay use your data for training
1StandardRetention or training terms are ambiguous
2HighNo training on your data, deleted within 30 days
3MaximumNothing stored once the response is returned (ZDR)

You can control which trust levels your requests are routed to via the trust_level routing hint.

Changes

We will update this page when subprocessors are added or removed. Material changes will be communicated to account holders via email.