One policy.
Every model provider.
The same rules and the same audit trail whether an answer came from OpenAI, Bedrock, Vertex or a model you host — and a side-by-side view of how each provider behaves on your own traffic. Cloud guardrails stop at their own cloud; Xelurel doesn't.
Providers and how to connect them
| Provider | Drop-in proxy | SDK client wrapper | Vercel AI SDK | LiteLLM gateway | Their guardrail verdicts |
|---|---|---|---|---|---|
| OpenAI | ✓ | ✓ | ✓ | ✓ | Moderation, via API |
| Azure OpenAI | ✓ | ✓ | ✓ | ✓ | Content filters — automatic |
| Anthropic | ✓ | ✓ | ✓ | ✓ | — |
| Amazon Bedrock | Converse | ✓ | ✓ | ✓ | Guardrails — automatic |
| Google Gemini & Vertex AI | ✓ | — | ✓ | ✓ | Safety ratings — automatic |
| Mistral, Groq, Together, Fireworks, DeepSeek, xAI, OpenRouter | One-click | ✓ | ✓ | ✓ | — |
| Self-hosted (Ollama, vLLM) | Allowlisted URL | ✓ | ✓ | ✓ | — |
Anything else — including models not listed here — works through the direct API. Every decision records the provider, so policies can route by it (bedrock/* → your healthcare policy) and Analytics can compare providers.
Pick the integration that fits your stack
Drop-in proxy
Change the base URL of the SDK you already use. Every request is assessed on the way back; streams are gated while they flow.
Docs →
SDK client wrappers
Keep calling providers directly. wrapOpenAI, wrapAnthropic and wrapBedrock (JavaScript and Python) govern the client in place.
Docs →
Vercel AI SDK
One middleware for every model the AI SDK can call. The provider is read from the model, so decisions say where each answer came from.
Docs →
LiteLLM gateway
Add Xelurel as a guardrail in config.yaml and every model behind the gateway — more than 100 providers — shares one policy.
Docs →
LangChain & LlamaIndex
Callback handlers for chains and agents, in Python and JavaScript.
Docs →
Direct API
POST the prompt and answer to /v1/assess from anywhere, for any model. Send provider to record where it ran.
Docs →
Sits above your existing guardrails
Already using a cloud guardrail? Keep it. Send its verdict with the assessment — or let the proxy pick it up from the provider's response — and it's stored on the decision next to Xelurel's own. Policies can act on it (“block when Bedrock Guardrails blocked a prompt attack”), and anything another guardrail flags gets a closer look from the LLM tier.
Start with the provider you already use
Free for 10,000 assessments a month. No card required.
Start free →