apflow

Task Orchestration and Execution Framework

Core Positioning

The core of apflow is task orchestration and execution specifications. It provides a unified task orchestration framework that supports execution of multiple task types. The core is pure orchestration with no LLM dependencies - CrewAI support is optional.

Core includes:

• Task orchestration specifications (TaskManager)
• Core interfaces (ExecutableTask, BaseTask, TaskStorage)
• Storage (DuckDB default, PostgreSQL optional)
• NO CrewAI dependency (available via [crewai] extra)

Optional features:

• CrewAI Support [crewai]: LLM-based agent crews via CrewaiExecutor (task executor implementation)
• HTTP/REST Executor [http]: Remote API calls via RestExecutor (task executor implementation)
• SSH Executor [ssh]: Remote command execution via SSH (task executor implementation)
• Docker Executor [docker]: Containerized command execution (task executor implementation)
• gRPC Executor [grpc]: gRPC service calls (task executor implementation)
• WebSocket Executor: Bidirectional WebSocket communication (task executor implementation)
• apflow API Executor: Inter-instance API calls for distributed execution (task executor implementation)
• MCP Executor: Model Context Protocol executor for accessing external tools and data sources (task executor implementation)
• MCP Server [a2a]: MCP (Model Context Protocol) server exposing task orchestration as MCP tools and resources
• LLM Executor [llm]: Direct LLM interaction via LiteLLM (supports OpenAI, Anthropic, Gemini, etc.)
• A2A Protocol Server [a2a]: A2A Protocol Server (A2A Protocol is the standard protocol for agent communication)
• CLI Tools [cli]: Command-line interface

Note: CrewaiExecutor and future executors are all implementations of the ExecutableTask interface. Each executor handles different types of task execution (LLM, HTTP, etc.).

Core Features

Task Orchestration Specifications (Core)

• TaskManager: Task tree orchestration, dependency management, priority scheduling
• Unified Execution Specification: All task types unified through the ExecutableTask interface

Task Execution Types

All task executors implement the ExecutableTask interface:

• Custom Tasks (core): Users implement ExecutableTask for their own task types
• CrewaiExecutor [crewai]: LLM-based task execution via CrewAI (built-in executor)
• RestExecutor [http]: HTTP/REST API calls with authentication and retry (built-in executor)
• SshExecutor [ssh]: Remote command execution via SSH (built-in executor)
• DockerExecutor [docker]: Containerized command execution (built-in executor)
• GrpcExecutor [grpc]: gRPC service calls (built-in executor)
• WebSocketExecutor: Bidirectional WebSocket communication (built-in executor)
• ApFlowApiExecutor: Inter-instance API calls for distributed execution (built-in executor)
• McpExecutor: Model Context Protocol executor for accessing external tools and data sources (built-in executor)
• GenerateExecutor: Generate task tree JSON arrays from natural language requirements using LLM (built-in executor)
• LLMExecutor [llm]: Direct LLM interaction via LiteLLM (supports 100+ providers)
• BatchCrewaiExecutor [crewai]: Batch orchestration container (batches multiple crews)

Supporting Features

• Storage: Task state persistence (DuckDB default, PostgreSQL optional)
• Unified External API: A2A Protocol Server (HTTP, SSE, WebSocket) [a2a]
• Real-time Progress Streaming: Streaming support via A2A Protocol
• CLI Tools: Command-line interface [cli]

Protocol Standard

• A2A Protocol: The framework adopts A2A (Agent-to-Agent) Protocol as the standard protocol for agent communication. A2A Protocol is a mature, production-ready specification designed specifically for AI Agent systems, providing:
  • Agent-to-agent standardized communication interface
  • Streaming task execution support
  • Agent capability description mechanism (AgentCard, AgentSkill)
  • Multiple transport methods (HTTP, SSE, WebSocket)
  • Task management and status tracking
  • JWT authentication support

Installation

Core Library (Minimum - Pure Orchestration Framework)

pip install apflow

Includes: Task orchestration specifications, core interfaces, storage (DuckDB) Excludes: CrewAI, batch execution, API server, CLI tools

With Optional Features

# CrewAI LLM task support (includes batch)
pip install apflow[crewai]
# Includes: CrewaiExecutor for LLM-based agent crews
#           BatchCrewaiExecutor for atomic batch execution of multiple crews

# A2A Protocol Server (Agent-to-Agent communication protocol)
pip install apflow[a2a]
# Run A2A server: python -m apflow.api.main
# Or: apflow-server (CLI command)

# CLI tools
pip install apflow[cli]
# Run CLI: apflow or apflow

# PostgreSQL storage
pip install apflow[postgres]

# SSH executor (remote command execution)
pip install apflow[ssh]

# Docker executor (containerized execution)
pip install apflow[docker]

# gRPC executor (gRPC service calls)
pip install apflow[grpc]

# LLM support (LiteLLM, supports 100+ providers)
pip install apflow[llm]

# Everything (includes all extras)
pip install apflow[all]

🚀 Quick Start

Get started with apflow in minutes!

Installation

# Minimal installation (core only)
pip install apflow

# With all features
pip install apflow[all]

As a Library (Pure Core)

Using Task Orchestration Specifications:

from apflow import TaskManager, TaskTreeNode, create_session

# Create database session and task manager (core)
db = create_session()  # or: db = get_default_session()
task_manager = TaskManager(db)

# Create task tree (task orchestration)
# Use task_repository to create tasks
root_task = await task_manager.task_repository.create_task(
    name="root_task",
    user_id="user_123",
    priority=2
)

child_task = await task_manager.task_repository.create_task(
    name="custom_task",  # Task name corresponds to specific executor
    user_id="user_123",
    parent_id=root_task.id,
    dependencies=[],  # Dependency relationships
    inputs={"url": "https://example.com"}
)

# Build task tree and execute (task orchestration core)
task_tree = TaskTreeNode(root_task)
task_tree.add_child(TaskTreeNode(child_task))
result = await task_manager.distribute_task_tree(task_tree)

Creating Custom Tasks (Traditional External Service Calls):

from apflow import ExecutableTask
from typing import Dict, Any
import aiohttp

class APICallTask(ExecutableTask):
    """Traditional external API call task"""
    
    id = "api_call_task"
    name = "API Call Task"
    description = "Call external API service"
    
    async def execute(self, inputs: Dict[str, Any]) -> Dict[str, Any]:
        async with aiohttp.ClientSession() as session:
            async with session.post(inputs["url"], json=inputs.get("data")) as response:
                result = await response.json()
                return {"status": "completed", "result": result}
    
    def get_input_schema(self) -> Dict[str, Any]:
        return {
            "type": "object",
            "properties": {
                "url": {"type": "string", "description": "API endpoint"},
                "data": {"type": "object", "description": "Request data"}
            }
        }

With CrewAI Support [crewai]

Executing CrewAI (LLM) Tasks:

# Requires: pip install apflow[crewai]
from apflow.extensions.crewai import CrewaiExecutor

# CrewAI task execution
crew = CrewaiExecutor(
    name="Analysis Crew",
    agents=[{"role": "Analyst", "goal": "Analyze data"}],
    tasks=[{"description": "Analyze input", "agent": "Analyst"}]
)
result = await crew.execute(inputs={...})

With Batch Support [crewai]

Using BatchCrewaiExecutor to batch multiple crews (atomic operation):

# Requires: pip install apflow[crewai]
from apflow.extensions.crewai import BatchCrewaiExecutor, CrewaiExecutor

# BatchCrewaiExecutor is a batch container - executes multiple crews as atomic operation
batch = BatchCrewaiExecutor(
    id="my_batch",
    name="Batch Analysis",
    works={
        "data_collection": {
            "agents": [{"role": "Collector", "goal": "Collect data"}],
            "tasks": [{"description": "Collect data", "agent": "Collector"}]
        },
        "data_analysis": {
            "agents": [{"role": "Analyst", "goal": "Analyze data"}],
            "tasks": [{"description": "Analyze data", "agent": "Analyst"}]
        }
    }
)

# All crews execute sequentially, results are merged
# If any crew fails, entire batch fails (atomic)
result = await batch.execute(inputs={...})

CLI Usage

# Run tasks (standard mode - recommended)
apflow run flow --tasks '[{"id": "task1", "name": "Task 1", "schemas": {"method": "executor_id"}, "inputs": {"key": "value"}}]'

# Or use the shorthand
apflow run flow --tasks '[{"id": "task1", "name": "Task 1", "schemas": {"method": "executor_id"}, "inputs": {"key": "value"}}]'

# Or legacy mode (executor ID + inputs)
apflow run flow executor_id --inputs '{"key": "value"}'

# Start API server
apflow serve --port 8000

# Start daemon mode
apflow daemon start

# Stop daemon mode
apflow daemon stop

A2A Protocol Server

The [a2a] extra provides an A2A (Agent-to-Agent) Protocol server built on Starlette/FastAPI.

from apflow.api import create_app

# Create A2A protocol server app
app = create_app()

# Run with: uvicorn app:app --port 8000
# Or use the entry point: apflow-server

Note: The current [a2a] extra focuses on A2A protocol support. Future versions may include additional FastAPI REST API endpoints for direct HTTP access without the A2A protocol.

Architecture Design

┌─────────────────────────────────────────────────────────────┐
│              Unified External API Interface Layer            │
│  - A2A Protocol Server (HTTP/SSE/WebSocket) [a2a]          │
│  - CLI Tools [cli]                                          │
└─────────────────────────────────────────────────────────────┘
                          ↓
┌─────────────────────────────────────────────────────────────┐
│        Task Orchestration Specification Layer (CORE)         │
│        - TaskManager: Task tree orchestration, dependency  │
│          management, priority scheduling                     │
│        - ExecutableTask: Unified task interface             │
└─────────────────────────────────────────────────────────────┘
                          ↓
┌─────────────────────────────────────────────────────────────┐
│                    Task Execution Layer                      │
│  - Custom Tasks [core]: ExecutableTask implementations      │
│    • Traditional external service calls (API, DB, etc.)     │
│    • Automated task services (scheduled tasks, workflows)  │
│  - CrewaiExecutor [crewai]: CrewAI (LLM) task execution        │
│  - BatchCrewaiExecutor [crewai]: Batch task orchestration           │
└─────────────────────────────────────────────────────────────┘
                          ↓
┌─────────────────────────────────────────────────────────────┐
│                    Supporting Features Layer                │
│  - Storage: Task state persistence (DuckDB/PostgreSQL)     │
│  - Streaming: Real-time progress updates                    │
└─────────────────────────────────────────────────────────────┘

Project Structure

See docs/architecture/DIRECTORY_STRUCTURE.md for detailed directory structure and module descriptions.

Installation Strategy:

• pip install apflow: Core library only (execution, base, storage, utils) - NO CrewAI
• pip install apflow[crewai]: Core + CrewAI support (includes BatchCrewaiExecutor)
• pip install apflow[a2a]: Core + A2A Protocol Server
• pip install apflow[cli]: Core + CLI tools
• pip install apflow[all]: Full installation (all features)

Note: For examples and learning templates, see the test cases in tests/integration/ and tests/extensions/.

📚 Documentation

Quick Links:

• 📖 Full Documentation - Complete documentation index
• 🚀 Quick Start - Get running in 10 minutes
• 📘 Tutorials - Step-by-step tutorials
• 📝 Guides - Comprehensive guides
• 💡 Examples - Practical examples
• 🔧 API Reference - Complete API documentation

For New Users:

• Start with Getting Started
• Follow the Quick Start Guide
• Try the First Steps Tutorial

For Developers:

• Task Orchestration Guide
• Custom Tasks Guide
• Best Practices
• API Reference

For Contributors:

• Development Setup
• Contributing Guide
• Architecture Overview

Full documentation is also available at flow-docs.aipartnerup.com.

🤝 Contributing

Contributions are welcome! Please see our Contributing Guide for setup instructions and contribution guidelines.

📄 License

Apache-2.0

🔗 Links

• Documentation: docs/index.md - Complete documentation
• Website: aipartnerup.com
• GitHub: aipartnerup/apflow
• PyPI: apflow
• Issues: GitHub Issues
• Discussions: GitHub Discussions