# Kiln

> A desktop app and open-source Python library for building, evaluating, and optimizing AI systems with evals, RAG, agents, fine-tuning, and synthetic data.

Kiln, made by Chesterfield Laboratories Inc, is a workbench for the full AI development loop. It combines evals, prompt optimization, RAG, fine-tuning, synthetic data, agents, and tools in one desktop app for macOS, Windows, and Linux, paired with an MIT-licensed Python library for running the same tasks in production. The latest listed release is Kiln Desktop v1.2.0.

## What It Is

Kiln is a local-first AI development workbench. A team defines a task and a dataset once, then runs evals, prompt optimization, fine-tuning, RAG, agents, and synthetic data generation against that same dataset. The app runs on your machine; you bring your own API keys or go fully offline with Ollama. Team sync works through Git, and datasets are stored as open JSON.

## Workflow and Capabilities

- **Evals:** an eval builder auto-generates a judge and a synthetic eval dataset, which can be aligned to your preferences. Golden datasets and human ratings are supported.
- **Auto-Optimize:** searches across prompt, model, tools, skills, subagents, and parameters to tune a task against its evals.
- **RAG:** documents (PDF, image, video, audio) can be dragged in to create a RAG, with evals generated from your own documents.
- **Agents and tools:** multi-agent hierarchies with subagents, plus Tools and MCP, Skills, structured output, and reasoning models.
- **Fine-tuning:** zero-code fine-tuning across 60+ models on Fireworks, Together, and Vertex, with serverless deployment.
- **AI Assistant:** proposes improvements, runs experiments, and creates evals through conversation.

## Team and Developer Model

Product managers, QA, and subject experts can rate outputs, leave feedback, and add data in the app without coding. Engineers and data scientists use the `kiln-ai` Python library to load Kiln projects, run tasks, build fine-tunes, and work in notebooks, with a REST API also available. Kiln says it has tested 190+ models across major providers.

## Licensing and Optional Service

The core Python library and REST server are MIT-licensed. The desktop app is source-available and free to use under a fair-code model. The AI Assistant, Auto-Optimize, and Eval Builder are provided through the opt-in Kiln Pro service; the core app remains functional without it.

## Features
- Eval builder with LLM-as-judge
- Auto-Optimize for prompts, models, tools and parameters
- AI Assistant for experiments and optimization
- RAG from PDF, image, video and audio documents
- Subagents and multi-agent hierarchies
- Tools and MCP support
- Skills
- Synthetic data generation
- Zero-code fine-tuning across 60+ models
- Git-based collaboration and auto-sync
- Open-source MIT Python library and REST API
- Local-first with Ollama offline support
- Tested model library of 190+ models

## Integrations
OpenAI, Anthropic, Gemini, Ollama, OpenRouter, Amazon Bedrock, Azure OpenAI, Groq, Fireworks AI, Together AI, Cerebras, Hugging Face, Vertex, Git, MCP

## Platforms
MACOS, WINDOWS, LINUX, DEVELOPER_SDK, API

## Pricing
Freemium — Free tier available with paid upgrades

## Version
v1.2.0

## Links
- Website: https://kiln.tech
- Documentation: https://docs.kiln.tech
- Repository: https://github.com/Kiln-AI/Kiln
- EveryDev.ai: https://www.everydev.ai/tools/kiln