Easily Build Smart Multi-Agent Systems

OpenMAS streamlines asynchronous Multi-Agent System (MAS) development in Python. By providing a lightweight framework, standardized structure, and helpful CLI tools, it handles the foundational setup, freeing you to concentrate on what matters most: designing and implementing sophisticated agent behaviors.

Inspired by modern development ecosystems and driven by real-world use cases like coding and gaming agents, OpenMAS aims to streamline the entire MAS lifecycle, with particular attention to integrating communication protocols like the Model Context Protocol (MCP) alongside standard web protocols.

Full Documentation: https://docs.openmas.ai

• Simplified Agent Development: Build agents inheriting from BaseAgent with a clear asynchronous lifecycle (setup, run, shutdown).
• Flexible Communication: Pluggable communicators for HTTP, Model Context Protocol (SSE & Stdio), gRPC, MQTT, with lazy loading to keep dependencies minimal. Easily extend with custom communicators. See Communication.
• Structured Projects: Standardized directory layout (agents/, shared/, extensions/, packages/) generated by openmas init promotes modularity and maintainability. See Project Structure.
• Layered Configuration: Robust system loading configuration from files (openmas_project.yml, config/*.yml), .env, and environment variables. See Configuration Guide.
• Prompt Management: Comprehensive prompt management system for organizing, versioning, and reusing prompts, with storage backends and template rendering capabilities. See Prompt Management.
• Sampling Interface: Consistent abstraction for interacting with language models, with support for sampling parameters and MCP integration. Use with any LLM provider. See Prompt Management.
• Enhanced MCP Agent: The PromptMcpAgent combines prompt management, sampling, and MCP capabilities for streamlined LLM-powered agents. See MCP Integration.
• Agent Reasoning Agnosticism: While BaseAgent inherently supports heuristic-based logic, OpenMAS facilitates integrating diverse reasoning mechanisms. Follow guides for LLM Integration (using official LLM client libraries like OpenAI, Anthropic, Google Gemini) or explore built-in support for BDI Patterns (including BdiAgent and SPADE-BDI integration examples).
• Workflow Implementation: Implement various agent interaction patterns (see Building Effective Agents). While specific helpers exist for the Orchestrator-Worker pattern, the core framework enables building custom workflows like prompt chaining, routing, and parallel execution, with more helpers planned for future releases. See Agent Patterns.
• Developer Workflow Tools: Use the openmas CLI tool for initializing projects (openmas init), validating configuration (openmas validate), running agents locally (openmas run), managing dependencies (openmas deps), and generating deployment artifacts (openmas generate-dockerfile, openmas generate-compose). See CLI Docs.
• Extensibility: Design encourages local project extensions (extensions/) and shareable external packages (packages/).
• Testing Utilities: Includes MockCommunicator and AgentTestHarness to facilitate unit and integration testing. See Testing Your Agents.

pip install openmas

OpenMAS has optional extras for different communication protocols ([mcp], [grpc], [mqtt], [all]).

See the full Installation Guide for details on prerequisites, virtual environments, and optional dependencies.

The best way to get started with OpenMAS is by following our detailed tutorial, which guides you through creating and running your first agent using the standard project structure:

• Getting Started Guide

This guide uses the openmas init command to set up a project with the proper directory structure and openmas run to execute the agent.

Contributions are welcome! Please see the Contributing Guide for details on how to get involved, set up your development environment, run tests (tox), and submit pull requests.

OpenMAS is licensed under the MIT License.