Sublayer

Sublayer is an open-source Ruby gem for building AI agents directly inside your Rails or Ruby codebase. Instead of shipping AI logic off to a Python service, you compose it from three plain Ruby classes: Generators take a string in and return LLM-generated string output, Actions do something useful with that output, and Agents pick which action to run. Because everything is ordinary object-oriented Ruby, it drops into the patterns Rails developers already use. The framework is model-agnostic. Provider-agnostic adapters let the same code target OpenAI or Anthropic, and the docs cover running models locally with Ollama (there's a Llama 3.1 walkthrough) or llamafile when privacy or latency matters. Setup is a gem install, usually in a Gemfile, with an OPENAI_API_KEY set as an environment variable; the docs' CodeFromDescriptionGenerator example turns a plain-language description plus a technology list into working code. The recent additions are the interesting part for anyone past the prototype stage. Streaming responses give you real-time output, custom agent triggers can launch workflows from external events or schedules, and Generator Output Adapters plus dynamic output structures let you shape exactly what an agent returns. There's even a guide for standing up a voice chat assistant on Rails. Sublayer bills itself as a self-assembling, model-agnostic agent framework for Ruby, and that's an honest description of its scope: lightweight, code-first, no visual builder in the box. It fits Ruby and Rails developers who want AI automation living in Ruby rather than a Python sidecar or a hosted agent platform. Compared with Python-first frameworks like LangChain or CrewAI, or drag-and-drop tools like Flowise, Sublayer trades breadth for staying inside the stack you already maintain.