// the find
pezzolabs/pezzo
🕹️ Open-source, developer-first LLMOps platform designed to streamline prompt design, version management, instant delivery, collaboration, troubleshooting, observability and more.
Pezzo is a self-hosted LLMOps platform that treats prompts as versioned, deployable artifacts — commit a prompt version, publish it to an environment, ship it without a code deploy. It's aimed at teams running multiple LLM-powered features who want prompt iteration separated from their release cycle. The NX monorepo houses a NestJS GraphQL API, a React console, and an OpenAI proxy.
The prompt-as-artifact model (version → commit → publish to environment) solves a real problem: hardcoded prompt strings that require a full deploy to change. ClickHouse for request logging is the right call — append-only, aggregation-friendly, and won't hammer your main Postgres when you're logging thousands of LLM calls per day. NX monorepo with GraphQL codegen in watch mode means the frontend types stay in sync with the API schema automatically, which is the right way to run a TypeScript full-stack project. Provider API keys are stored encrypted at the org level, so they don't scatter across project env files.
The infrastructure footprint is punishing before you write a line of code: PostgreSQL + ClickHouse + Redis + Supertokens is four services just to get the platform running — teams wanting basic observability will balk. Provider support is thinner than advertised: the console only has settings schemas implemented for OpenAI and Azure OpenAI; Anthropic, Mistral, and Meta logos appear in the assets directory but the corresponding ProviderSettings renderers don't exist. Schema migrations stop in May 2024 despite the repo still receiving pushes, which strongly suggests core development has stalled and effort shifted to the paid cloud offering. The proxy architecture routes all LLM traffic through an extra hop with no documented failover behavior — if the proxy goes down, your LLM calls go with it.