Kiln

Rapid iteration across the whole post-training loop is Kiln's pitch: synthetic data generation, evaluations, fine-tuning across more than 60 models, RAG, multi-agent workflows, and prompt optimization all operate on one shared dataset, so work done in one stage feeds the next. It ships as a free desktop app for macOS, Windows, and Linux that non-developers can run, backed by an MIT licensed Python library and REST server, published as kiln-ai on PyPI, for programmatic use; the desktop app itself is source-available under a fair-code license and free for individuals. Projects are stored as files and synced through Git, giving teams review and versioning without a hosted backend, and model access spans hosted provider APIs as well as local models through Ollama. An optional Kiln Pro subscription adds an AI assistant, auto-optimization, and an eval builder, while the core stays free. The repository has around 5,000 GitHub stars and documentation at docs.kiln.tech, and it suits teams that want structured evals and dataset workflows without stitching together five separate tools.