caudena/beam_weaver

BeamWeaver is an Elixir library for building AI agents and multi-step workflows that use language models. It brings together ideas from Python tools like LangChain and LangGraph, but is written from scratch in Elixir and designed to work naturally with Elixir's OTP runtime, which is known for building reliable, concurrent systems. It is not a wrapper around any Python library. The library gives developers the building blocks to wire language models into real applications. Agents can call tools, maintain conversation memory, produce structured output, and pause to wait for human input before continuing. Graph-based workflows let you define a series of steps with branching logic, checkpoints that save progress, and the ability to resume from a saved point if something fails or if a long task is interrupted. A deeper agent mode adds planning, virtual file systems, subagents that run concurrently, and summarization for long-running tasks. BeamWeaver connects to AI providers including OpenAI, Anthropic, Google Gemini, xAI, and Moonshot. It includes built-in profiles for specific model versions from each provider, which carry the token limits, cost rates, and capability flags for that model. Retrieval features let you load documents, split them into chunks, create embeddings, and store them in a vector database for use in search-augmented agent responses. Observability tools let you trace what the agent did, how many tokens were used, and what each tool call returned, either locally or through an external tracing service called WeaveScope. The library is added to an Elixir project through the standard package manager. Configuration is minimal: you provide API keys for whichever providers you use, and the library handles the rest. Agents can be defined using a module-based format that keeps model choice, system prompt, tools, and middleware in one place, or built dynamically from configuration maps at runtime. BeamWeaver is early in development, at version 0.1.x. It targets Elixir developers building production services that need to incorporate AI model calls alongside existing OTP supervision, database access, and telemetry.