Pezzo is an open-source, cloud-native LLMOps platform that centralizes prompt design, version management, instant delivery, collaboration, troubleshooting, and observability for teams building on large language models.

What it is

Pezzo is a developer-first LLMOps platform written in TypeScript and released under the Apache 2.0 licence. It runs as a set of services built on open-source infrastructure and exposes its functionality through a GraphQL API, with a management console for prompt authoring and an observability layer for inspecting LLM traffic. The project positions itself as entirely cloud-native: PostgreSQL provides relational storage, ClickHouse handles analytical and observability data, Redis provides caching, and Supertokens manages authentication. Official clients exist for Node.js and Python, and LangChain is supported as an integration surface, so prompt management, observability, and caching are reachable from the three supported client paths.

The concrete problem it replaces is the practice of hardcoding prompt strings directly into application source code and shipping a full application release every time a prompt needs to change. When prompts live in code, iterating on wording requires a deploy, comparing variants requires manual bookkeeping, and diagnosing a bad model response requires digging through logs that were never designed for LLM traces. Pezzo moves prompt definitions and their versions out of the codebase into a managed registry that applications fetch at runtime, so prompt edits ship without redeploying the application, and request-level telemetry is captured in a store suited to inspecting it.

Key capabilities

• Prompt management with versioning and instant delivery, exposed through the GraphQL API and supported in the Node.js, Python, and LangChain clients.
• Observability over AI operations, backed by ClickHouse for request and trace data.
• Caching of model responses, with Redis among the infrastructure dependencies.
• GraphQL schema with code generation, kept current in watch mode through npm run graphql:codegen:watch.
• Docker Compose deployment path described as "Running With Docker Compose" in the official documentation at docs.pezzo.ai.
• Management console served locally on http://localhost:4200 for prompt authoring and operations work.
• Server health endpoint at http://localhost:3000/api/healthz for verifying a running deployment.

Who uses it and how

• Application teams that keep prompts in source control and want prompt changes to ship independently of application deploys.
• Node.js and Python services that already call LLM providers and need trace-level visibility into requests without building a logging stack.
• Teams using LangChain that want prompt management, observability, and caching applied to existing chains rather than rewriting them.
• Self-hosting platform or infrastructure engineers who run PostgreSQL, ClickHouse, Redis, and Supertokens and prefer that stack over a managed vendor service.
• Contributors interested in the hacktoberfest topic, working across the NestJS server, the console, and the GraphQL schema.

Getting started

For a full local stack, the documentation's "Running With Docker Compose" guide is the entry point; the repository also supports running infrastructure dependencies via docker-compose -f docker-compose.infra.yaml up. Development mode requires Node.js 18+ and Docker, then npm install, Prisma migration deployment with npx prisma migrate deploy --schema apps/server/prisma/schema.prisma, npx nx serve server for the API, and npx nx serve console for the console, with environment files based on .env.example and .env.docker.

How it compares

The provided facts do not name any paid products that Pezzo replaces, and they name no directly comparable open-source LLMOps tools either, so the honest statement is that Pezzo stands alone in this registry and no side-by-side comparison can be made from the supplied material. LangChain appears only as a supported client integration, not as a competing product.

When to use it — and when not to

A self-hoster must operate PostgreSQL, ClickHouse, Redis, and Supertokens, and in development must run a Prisma migration step, a server process, a codegen watcher, and the console alongside whatever LLM provider they call, which is a non-trivial footprint for a small team. Anyone who wants a managed service with no database operations, or who needs a supported client outside Node.js, Python, and LangChain, should look elsewhere. The README excerpt is also thin on the shape of the console and on production hardening guidance, so teams evaluating it should read docs.pezzo.ai and the issue tracker before committing.

npm install

docker-compose -f docker-compose.infra.yaml up

npx dotenv-cli -e apps/server/.env -- npx prisma migrate deploy --schema apps/server/prisma/schema.prisma

npx nx serve server

npm run graphql:codegen:watch

npx nx serve console