Kiln-AI/Kiln

Kiln – A Local‑First Workbench for Building AI Products

What it is – Kiln is both a desktop application (Mac/Windows/Linux) and an MIT‑licensed Python library (kiln‑ai). It aims to cover the whole AI development loop: creating evaluation suites, automatically optimizing prompts and model choices, building Retrieval‑Augmented Generation (RAG) pipelines, composing multi‑agent systems, generating synthetic data, and fine‑tuning models. All of these steps can be done through a graphical UI that syncs to a Git repo, or programmatically via the library or a REST API.

Key capabilities

Feature	What you can do (as described in the README)
Eval Builder	Auto‑generate evaluation datasets (including synthetic data) and define custom judges; creates a full eval suite in ~10 min.
Auto‑Optimize	Searches across prompt variations, model selections, tool/skill configurations, and sub‑agent setups to find the best performing combination for your defined eval metrics.
AI Assistant	An “AI data‑science partner” that suggests improvements, creates evals, and runs experiments (available in the Pro tier).
RAG	Drag‑and‑drop documents of many types (PDF, image, video, audio) to build a searchable index; auto‑generates RAG‑specific evals.
Sub‑agents	Build hierarchical agent structures where each sub‑agent has its own context window.
Synthetic Data Generation	Produce data for evaluation or fine‑tuning in minutes.
Fine‑Tuning	Zero‑code fine‑tuning for 60+ models (e.g., Qwen, Llama, GPT, Gemini) across providers like Fireworks, Together, Vertex; includes server‑less deployment.
Python library (kiln‑ai)	Same engine that powers the desktop app; can be used in notebooks, scripts, or production services.
Model catalog	Tested >190 models from providers such as OpenAI, Anthropic, Gemini, Bedrock, Ollama, OpenRouter, Fireworks, Groq, etc.
Collaboration	Git‑native syncing so non‑technical teammates can rate outputs, add data, or flag regressions without touching code.
Local‑first / offline	Runs on the user’s machine; you bring your own API keys or use fully offline models via Ollama.

Who it’s for – Teams that want a unified, open‑source environment for iterating on LLM‑based products without being locked into a SaaS platform. It targets product managers, subject‑matter experts, QA, and engineers alike.

How to get started – Download the desktop app (one‑click installer) or install the library with pip install kiln-ai. The docs provide a 5‑minute quick‑start that walks you through creating a task, running an eval, and optimizing the prompt.

Licensing & availability – The core Python library and REST server are MIT‑licensed. The desktop app’s source is “source‑available” under a fair‑code model, free for individuals; a paid “Kiln Pro” tier adds the AI Assistant, Auto‑Optimize, and Eval Builder but the base app remains fully functional.

Community – Discord, newsletter, GitHub issues, and a contributing guide are provided for collaboration.

---

All information above is taken directly from the repository’s README.