Agenta

MIT-licensed prompt engineering, evaluation and deployment: self-host the whole loop.

Agenta is the open-source LLM engineering platform focused on prompt lifecycle: a visual playground for iterating prompts and model choices side by side, versioning with the ability to diff and roll back, evaluation on datasets, human feedback collection, and deployment that turns a prompt into an API endpoint.

Why founders use it: prompts change more often than code, and managing that churn in a repository full of string constants hurts. Agenta gives prompt engineering the workflow of software, experiment in the playground, evaluate against cases, promote a version, monitor in production, with non-technical teammates able to iterate safely inside guardrails.

A realistic setup: prompt variants for the product's LLM features, datasets of real inputs with expected behaviors, evaluations comparing variants before promotion, the deployed endpoint behind your API, and feedback signals feeding the next iteration.

Trade-offs: it centers prompt-and-model iteration, not deep tracing or complex agent graphs, and the project is younger than the category leaders. Compare with Langfuse for observability breadth, Dify for the full app canvas, and LangSmith for ecosystem depth.