kiln

Kiln is a free desktop application and open-source Python library for building, evaluating and optimizing AI systems. It is positioned as a workbench covering the full AI development loop rather than a single-purpose framework. Core capabilities described in the README: - Evaluations: an Eval Builder that can auto-generate evals (a judge plus a synthetic eval dataset) and align them to user preferences. - Auto-Optimize: searches across prompt mutations, model selection, tools, skills, subagents and parameters to improve results across eval dimensions. - RAG: drag-and-drop documents (PDF, image, video, audio) to build retrieval-augmented generation, with auto-generated RAG evals from your own documents. - Agents: compose multi-agent hierarchies, where each subagent runs in its own focused context window. - Synthetic data generation: produce data for evals or fine-tuning. - Fine-tuning: zero-code fine-tuning across many models (Qwen, Llama, GPT, Gemini and others) on Fireworks, Together and Vertex, with serverless deployment. - Dataset management and structured outputs, reasoning models, tools and MCP, and a model library. Collaboration and deployment: - The desktop app is aimed at the whole team: PMs, subject-matter experts and QA can rate outputs and add data without writing code. - Projects sync to Git automatically, so collaboration uses Git infrastructure the team already owns. - The same tasks can be shipped to production through the MIT-licensed kiln-ai Python library, which is the same library used inside the app. It supports loading Kiln projects, running tasks, building fine-tunes, working in notebooks, and integrating Pandas/Polars dataframes. A REST API is also documented. Local-first design: - Runs on the user's machine; users bring their own API keys or run fully offline with Ollama. - Datasets are stored as open JSON, and the README states users own and control their datasets. - The README lists support for OpenAI, Anthropic, Gemini, Bedrock, Ollama, OpenRouter, Fireworks, Groq, any OpenAI-compatible endpoint, and others, with a model library described as tested across providers. Licensing and availability: - 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. - Kiln Pro is an optional paid service adding the AI Assistant, Auto-Optimize and Eval Builder; the core app remains functional without it. - The Kiln name and logos are trademarks of Chesterfield Laboratories Inc. Getting started: one-click installers for macOS, Windows and Linux, a documented quickstart, a Python library quickstart, and community channels including Discord, a newsletter and GitHub issues.