orxhestra

Multi-agent orchestration framework for Python — turn any agent setup into a CLI or server.

Compose multi-agent AI systems with async event streaming, agent hierarchies, and built-in support for MCP and A2A protocols.

Orx CLI

Turn any orx.yaml agent setup into an interactive terminal agent. Ships with a coding agent out of the box — or compose your own.

pip install orxhestra[cli,openai]
orx

+-- orx - terminal coding agent ------------------------------------+
|  model: gpt-5.4   workspace: ~/my-project   /help for commands    |
+-------------------------------------------------------------------+

orx> add error handling to the API routes

  > read_file(src/api/routes.py)
  > grep(pattern="raise", path=src/api/)
  > write_todos(3 tasks)

  Tasks
  * Add try/except to all route handlers  [in progress]
  - Add custom error response model
  - Write tests for error cases

  > edit_file(src/api/routes.py)
  > shell_exec(pytest tests/test_api.py)
  4 passed

  Done - added structured error handling to all 4 route handlers
  with a custom ErrorResponse model. All tests pass.

Features

• Any LLM — OpenAI, Anthropic, Google via --model gpt-5.4 / claude-sonnet-4-6 / gemini-2.0-flash
• Streaming — real-time token rendering with Markdown formatting
• Tool approval — prompts before destructive operations (write, edit, shell)
• Task planning — structured todo lists visible in the terminal
• Sub-agent delegation — spawn isolated agents for complex subtasks
• AGENTS.md memory — persistent project context across sessions
• Local context injection — auto-detects language, git state, package manager, project tree
• Context summarization — auto-compacts long conversations, /compact command
• Orx YAML — run any orx.yaml agent team: orx my-agents.yaml

Usage

orx                               # interactive REPL (default model)
orx --model claude-sonnet-4-6     # use a specific model
orx -c "fix the failing tests"    # single-shot command
orx my-agents.yaml                # run a custom orx file
orx --auto-approve                # skip approval prompts
orx orx.yaml --serve -p 9000      # start as A2A server

Commands

Quickstart (SDK)

pip install orxhestra
# or
uv add orxhestra

from orxhestra import LlmAgent, Runner, InMemorySessionService

agent = LlmAgent(
    name="assistant",
    model="gpt-5.4",
    instructions="You are a helpful assistant.",
)

runner = Runner(agent=agent, session_service=InMemorySessionService())
response = await runner.run(user_id="user1", session_id="s1", new_message="Hello!")

for event in response:
    print(event.content)

[!TIP]
For full documentation, guides, and API reference, visit docs.orxhestra.com.

Features

• Agent ensemble - LLM, ReAct, Sequential, Parallel, and Loop agents
• Event streaming - Async event-driven architecture with real-time streaming
• Composer - Conduct entire agent orchestras declaratively with YAML
• Tools - Function tools, filesystem tools, agent-as-tool, shell, and long-running tool support
• Planners - Choreograph task execution with PlanReAct and TaskPlanner strategies
• Skills - Reusable, composable agent repertoires
• MCP - Model Context Protocol integration for tool servers
• A2A - Agent-to-Agent protocol for cross-service harmonization
• Memory - Pluggable memory stores for persistent agent context
• Tracing - Built-in support for Langfuse, LangSmith, and custom callbacks

Agents at a glance

Composer

Define entire agent orchestras in a single YAML file — no Python wiring needed. Compose LLM agents, loops, pipelines, tools, and review cycles declaratively. The example below builds a coding agent that plans, implements with filesystem + shell access, and self-reviews in a loop:

defaults:
  model:
    provider: openai
    name: gpt-5.4

tools:
  exit:
    builtin: "exit_loop"
  filesystem:
    builtin: "filesystem"
  shell:
    builtin: "shell"

agents:
  planner:
    type: llm
    description: "Plans the implementation steps for the coder agent."
    instructions: |
      Output a numbered list of concrete steps the coder
      should execute. Each step must be an actionable file
      operation or shell command.

  coder:
    type: llm
    description: "Implements code changes with filesystem and shell access."
    instructions: |
      Follow the plan from the previous step exactly.
      Use filesystem tools to create files and shell to
      run commands. Never ask the user to do anything.
    tools:
      - filesystem
      - shell

  reviewer:
    type: llm
    description: "Reviews changes and approves or requests fixes."
    instructions: |
      Check files exist and look correct. If done, call
      exit_loop. Otherwise describe what needs fixing.
    tools:
      - exit

  dev_loop:
    type: loop
    agents: [coder, reviewer]
    max_iterations: 10

  coordinator:
    type: sequential
    agents: [planner, dev_loop]

main_agent: coordinator

runner:
  app_name: coding-agent
  session_service: memory

Run it as an interactive CLI or expose it as an A2A server:

orx orx.yaml                    # interactive terminal agent
orx orx.yaml --serve -p 9000    # A2A server on port 9000

# test the server
curl -X POST http://localhost:9000/ \
  -H "Content-Type: application/json" \
  -d '{
    "jsonrpc": "2.0", "id": "1",
    "method": "message/send",
    "params": {
      "message": {
        "role": "user",
        "parts": [{"text": "Hello!", "mediaType": "text/plain"}]
      }
    }
  }'

Docker

docker run -e OPENAI_API_KEY=$OPENAI_API_KEY \
  -v ./orx.yaml:/app/orx.yaml \
  nicolaimtlassen/orxhestra

Documentation

• Getting Started - Installation and first agent
• Agents - Agent types and configuration
• Tools - Built-in and custom tools
• Composer - YAML-based agent composition
• Integrations - MCP and A2A setup

Acknowledgments

This project is built on the shoulders of several outstanding open-source projects and research efforts:

• LangChain
• Google Agent Development Kit (ADK)
• LangGraph
• Model Context Protocol (MCP)
• Agent-to-Agent Protocol (A2A)

Special thanks to the open-source AI community for pushing the boundaries of what's possible with agent frameworks.