Open Swarm

What is Open Swarm?

Open Swarm is a modular, multi-agent framework for building intelligent agent systems. It is a fork and actively maintained extension of the OpenAI Swarm framework, adding stateless RESTful operations, a plugin system, and optional MCP integration. It runs via CLI, REST API, or as a Python SDK.

How to use Open Swarm?

Install via pip install open-swarm, configure an LLM provider in ~/.swarm/swarm_config.json (defaults to OpenAI), optionally add MCP servers, then add and run a blueprint using the swarm-cli commands: swarm-cli add /path/to/blueprint.py --name example and swarm-cli run example. Blueprints are Python modules encapsulating agent logic, tools, and configuration.

Key features of Open Swarm

• Multi-agent orchestration with handoffs and context sharing
• Blueprint-driven architecture for reusable modular agents
• Optional MCP integration for external tools (e.g., databases, web search)
• CLI & Django-powered REST API with interactive web pages
• OpenAI API–compatible endpoint (/v1/chat/completions)
• Configurable LLM providers per agent (OpenAI, Groq, Ollama, etc.)

Use cases of Open Swarm

• Build personal assistants combining weather lookups and documentation search
• Create database-querying agents with real-time web search via MCP
• Automate system administration tasks using filesystem, SQLite, and search MCP tools
• Develop multi‑agent university- or business-oriented task workflows

FAQ from Open Swarm

What is the status of Open Swarm?

The README states that OpenAI has released openai-agents as the successor to swarm, and that this project is no longer relevant. However, experimental integration with OpenAI Agents is available on the experimental branch.

How does Open Swarm differ from the original OpenAI Swarm?

Open Swarm is a fork with extensions including stateless REST operations, a plugin system, blueprint-driven architecture, and optional MCP server integration. It also supports multiple configurable LLM providers per agent.

Does Open Swarm support MCP?

Yes, MCP integration is optional. npx MCP servers work well, but uvx‑based MCP servers currently have issues, and some blueprints using uvx are marked as broken.

Is Open Swarm free and open source?

Yes, it is distributed under the MIT License.

What are known limitations of Open Swarm?

The framework adds a mandatory sender field in agent responses for proper handoffs, but most OpenAI‑compatible API clients ignore this field by default, requiring custom UI or logic to display agent names.