orxhestra
A Python multi-agent orchestration framework that turns a YAML-defined agent team into a runnable command-line agent or an A2A server, with sub-agent delegation, persistent per-project memory, and many supported model providers.
orxhestra composes multi-agent AI systems from a YAML specification and can run them either as an interactive terminal agent or as a server, with async event streaming and built-in support for the MCP and A2A protocols. Its bundled command-line tool ships with a coding agent out of the box, supporting streaming output, tool-use approval prompts before destructive operations, structured visible task planning, and delegation to isolated sub-agents for complex subtasks.
It connects to 29 language model providers, including OpenAI, Azure OpenAI, Anthropic, Google, Mistral, Cohere, Groq, DeepSeek, and Ollama, and keeps four types of persistent, per-project memory across sessions: user, feedback, project, and reference. Long conversations are automatically summarized to free context, and any custom YAML agent team file can be run directly from the command line, including as a served A2A endpoint.