LLMProxyA self-hosted LLM gateway with an audit log you can verify

OpenAI-compatible. Records every request it handles in a hash chain, applies policy to what a response may do, and runs as one process on your infrastructure.

Quick start

What the audit log proves

GitHub

Verifiable audit log

Served, refused and failed requests are rows of a hash chain. The chain can be keyed with HMAC, its head recorded outside the database, and retention and erasure leave a record in the chain.

Tool policy

Which tools a response may call, and which may follow a tool result. It stops the call an injected instruction asks for without having to recognise the instruction.

Measured detection

The firewall and the shield are lexical. What they stop on public datasets is measured, published and regenerated from a results file, misses included.

Provider translation and fallback

Adapters for OpenAI, Anthropic, Google, Azure OpenAI and Ollama, plus OpenAI-compatible providers. Fallback chains, a circuit breaker per endpoint, a daily budget limit.

PII masking

Regular expressions by default, Presidio when installed. Masked before the request leaves, restored in the response.

One process, your infrastructure

SQLite by default, Postgres optional. No hosted component and no telemetry. Not built to run as several replicas.