Langroid is a lightweight, extensible Python framework (MIT license) designed explicitly around agents as first-class citizens, with multi-agent programming as its core paradigm. Created by researchers from Carnegie Mellon University and UW-Madison, it is inspired by the Actor model: developers set up agents, equip them with optional LLM, vector-store, and tool components, and assign them tasks that agents collaboratively solve by exchanging messages. The central abstraction is the Agent/Task pair. A Task wraps an Agent, provides it with instructions or goals, and manages iteration over the agent's responder methods. Sub-tasks are treated as additional responders, enabling hierarchical and recursive task delegation: a top-level task can silently hand off to sub-tasks without any special-case orchestration code. This makes complex multi-agent pipelines composable from simple building blocks. Langroid has strong built-in RAG (Retrieval-Augmented Generation) support via a DocChatAgent that incorporates advanced retrieval techniques. Supported vector stores include Qdrant, Chroma, LanceDB, Pinecone, PostgresDB (PGVector), and Weaviate. Tools and function calling share the same developer-facing Pydantic interface, covering both OpenAI function-calling and Langroid's own native tool system, making it straightforward to define structured tool use. The framework is model-agnostic: it supports OpenAI models as well as a wide range of non-OpenAI local and remote LLMs. Recent releases (active through mid-2026) have added Portkey integration, FastMCP client persistent connections, and expanded TaskConfig options. The project has over 4,000 GitHub stars and is available on PyPI (requires Python 3.11+). It is free and open-source with no commercial pricing tier. Key features: - Hierarchical multi-agent task delegation via composable Agent/Task pairs - Built-in DocChatAgent with advanced RAG techniques and source citation - Support for 6+ vector stores: Qdrant, Chroma, LanceDB, Pinecone, PGVector, Weaviate - Unified Pydantic-based interface for both OpenAI function-calling and native tools - Model-agnostic: supports OpenAI and non-OpenAI local/remote LLMs - Detailed message provenance and lineage logging for multi-agent interactions - SQLChatAgent and structured information extraction built in - Active development with FastMCP client support and Portkey integration (2025-2026)
Free and open-source (MIT license). No commercial pricing tier.