// the find
Kiln-AI/Kiln
Build, Evaluate, and Optimize AI Systems. Includes evals, RAG, agents, fine-tuning, synthetic data generation, dataset management, MCP, and more.
Kiln is a desktop workbench for the full AI development loop — evals, prompt optimization, RAG, fine-tuning, synthetic data, and agents — where the same dataset flows through every stage without rewriting anything. The MIT-licensed Python library ships what you build in the GUI straight to production. Aimed at teams where engineers aren't the only people who need to interact with AI task data.
The 'one dataset, every technique' design is the real differentiator: define a task once and run evals, fine-tuning, RAG, and synthetic data generation against the same artifact. Most frameworks force you to maintain separate pipelines per technique. Git-native collaboration is a smart call — non-engineers use the GUI, the data lives in a Git repo you already control, no SaaS data custody required. Auto-Optimize searches across prompt mutations AND model selection simultaneously, not just prompt variants — which is more useful than eval tools that only score what you already have. Desktop app and Python library share the same engine and project files, so there's no rewrite penalty when moving from experimentation to production.
The most compelling features — AI Assistant, Auto-Optimize, and Eval Builder — are paywalled behind Kiln Pro. Without them, the free tier is basically a local UI for running prompts and manually labeling outputs, which is not a strong case for switching from whatever you already have. Fine-tuning is fully outsourced to Fireworks/Together/Vertex; if you have your own GPU cluster or need local fine-tuning, there's nothing here for you. The three-tier licensing (MIT core library + source-available fair-code desktop app + Pro subscription) is genuinely confusing — it takes a careful reading to understand which workflows are free and which ones hit a paywall mid-experiment. Headless/CI use is technically possible via the REST API but the whole system is clearly designed around the desktop GUI first, so pipeline integration is awkward.