Grepture

About Grepture

Grepture is an open-source AI gateway that sits on the hot path of every LLM request rather than observing from the sidelines, which is what lets it do protection and observability from one integration. The distinction it draws is deliberate: most observability tools watch after the fact and most gateways only route, whereas Grepture redacts PII, blocks prompt injections, and enforces rules before data leaves your network, while simultaneously capturing traces, scoring responses, and tracking spend. Integration is a drop-in SDK with zero dependencies — you wrap your existing OpenAI or Anthropic client options and every call is traced. It works with OpenAI, Anthropic, Google AI, Azure OpenAI, Cohere, Mistral, AWS Bedrock, HuggingFace, Groq, and Replicate. On observability you get detailed request logs, multi-step agent traces with waterfall timelines, full-text search across all prompts and responses, request replay, and before/after diff views. Evals run LLM-as-judge scoring on live traffic with six built-in templates or custom judge prompts, with configurable sampling — no separate eval pipeline. Protection covers 80+ detection rules with reversible custom patterns, local AI models that find names, organizations, and addresses in freeform text without sending anything to third-party services, and mask-and-restore reversible redaction. The zero-data mode is the notable option: rules still fire in flight, but headers, bodies, and URLs never touch the database — only method, status, latency, and rule hits are stored. It is EU-hosted, and ships a CLI for local dev, pre-commit scanning, and CI.

Key Features

Tags