JAF (Juspay Agent Framework) - Python Implementation

A purely functional agent framework with immutable state and composable tools, professionally converted from TypeScript to Python. JAF enables building production-ready AI agent systems with built-in security, observability, and error handling.

Production Ready: Complete feature parity with TypeScript version, comprehensive test suite, and production deployment support.

Read the Full Documentation

Get Started → | API Reference → | Examples →

Key Features

Complete TypeScript Conversion

• Full Feature Parity: All TypeScript functionality converted to Python
• Type Safety: Pydantic models with runtime validation
• Immutable State: Functional programming principles preserved
• Tool Integration: Complete tool calling and execution system

Production Ready Server

• FastAPI Server: High-performance async HTTP API
• Auto Documentation: Interactive API docs at /docs
• Health Monitoring: Built-in health checks and metrics
• CORS Support: Ready for browser integration

Model Context Protocol (MCP)

• MCP Client: Full MCP specification support
• Stdio & SSE: Multiple transport protocols
• Tool Integration: Seamless MCP tool integration
• Auto Discovery: Dynamic tool loading from MCP servers

Enterprise Security

• Input Guardrails: Content filtering and validation
• Output Guardrails: Response sanitization
• Permission System: Role-based access control
• Audit Logging: Complete interaction tracing
• Proxy Support: Corporate proxy integration with authentication

Observability & Monitoring

• Real-time Tracing: Event-driven observability
• OpenTelemetry Integration: Distributed tracing with OTLP
• Langfuse Tracing: LLM observability and analytics
• Structured Logging: JSON-formatted logs
• Error Handling: Comprehensive error types and recovery
• Performance Metrics: Built-in timing and counters

Agent-as-Tool Architecture

• Hierarchical Orchestration: Use agents as tools in other agents
• Conditional Tool Enabling: Enable/disable agent tools based on context
• Session Management: Configurable session inheritance for sub-agents
• Flexible Output Extraction: Custom extractors for agent tool outputs

Developer Experience

• CLI Tools: Project initialization and management
• Hot Reload: Development server with auto-reload
• Type Hints: Full mypy compatibility
• Rich Examples: RAG, multi-agent, agent-as-tool, and server demos
• Visual Architecture: Graphviz-powered agent and tool diagrams

Core Philosophy

• Immutability: All core data structures are deeply immutable
• Pure Functions: Core logic expressed as pure, predictable functions
• Effects at the Edge: Side effects isolated in Provider modules
• Composition over Configuration: Build complex behavior by composing simple functions
• Type-Safe by Design: Leverages Python's type system with Pydantic for runtime safety

Quick Start

Installation

# Install from GitHub (development version)
pip install git+https://github.com/xynehq/jaf-py.git

# Or install with all optional dependencies
pip install "jaf-py[all] @ git+https://github.com/xynehq/jaf-py.git"

# Install specific feature sets
pip install "jaf-py[server] @ git+https://github.com/xynehq/jaf-py.git"        # FastAPI server support
pip install "jaf-py[memory] @ git+https://github.com/xynehq/jaf-py.git"        # Redis/PostgreSQL memory providers
pip install "jaf-py[visualization] @ git+https://github.com/xynehq/jaf-py.git" # Graphviz visualization tools
pip install "jaf-py[tracing] @ git+https://github.com/xynehq/jaf-py.git"       # OpenTelemetry and Langfuse tracing
pip install "jaf-py[dev] @ git+https://github.com/xynehq/jaf-py.git"           # Development tools

CLI Usage

JAF includes a powerful CLI for project management:

# Initialize a new JAF project
jaf init my-agent-project

# Run the development server
jaf server --host 0.0.0.0 --port 8000

# Show version and help
jaf version
jaf --help

Development Setup

git clone https://github.com/xynehq/jaf-py
cd jaf-py
pip install -e ".[dev]"

# Run tests
pytest

# Type checking and linting
mypy jaf/
ruff check jaf/
black jaf/

# Documentation
pip install -r requirements-docs.txt
./docs.sh serve  # Start documentation server
./docs.sh deploy # Deploy to GitHub Pages

Documentation

Official Documentation Website

The complete, searchable documentation is available at xynehq.github.io/jaf-py with:

• Interactive navigation with search and filtering
• Dark/light mode with automatic system preference detection
• Mobile-responsive design for documentation on any device
• Live code examples with syntax highlighting
• API reference with auto-generated documentation
• Always up-to-date with automatic deployments

Local Documentation

For offline access, documentation is also available in the docs/ directory:

• Documentation Hub - Your starting point for all documentation
• Getting Started - Installation and first agent tutorial
• Core Concepts - JAF's functional architecture principles
• API Reference - Complete Python API documentation
• Tools Guide - Creating and using tools
• Memory System - Persistence and memory providers
• Model Providers - LiteLLM integration
• Monitoring - Observability, metrics, and alerting
• Server API - FastAPI endpoints reference
• Deployment - Production deployment guide
• Examples - Detailed example walkthroughs
• Troubleshooting - Common issues and solutions

Project Structure

jaf-py/
├── docs/                    # Complete documentation
├── jaf/                     # Core framework package
│   ├── core/               # Core types and engine
│   ├── memory/             # Conversation persistence
│   ├── providers/          # External integrations (LiteLLM, MCP)
│   ├── policies/           # Validation and security
│   ├── server/             # FastAPI server
│   └── cli.py              # Command-line interface
├── examples/               # Example applications
└── tests/                  # Test suite

Architectural Visualization

JAF includes powerful visualization capabilities to help you understand and document your agent systems.

Prerequisites

First, install the system Graphviz dependency:

# macOS
brew install graphviz

# Ubuntu/Debian
sudo apt-get install graphviz

# Windows (via Chocolatey)
choco install graphviz

Then install JAF with visualization support:

pip install "jaf-py[visualization]"

Quick Start

import asyncio
from jaf import Agent, Tool, generate_agent_graph, GraphOptions

# Create your agents
agent = Agent(
    name='MyAgent',
    instructions=lambda state: "I am a helpful assistant.",
    tools=[my_tool],
    handoffs=['OtherAgent']
)

# Generate visualization
async def main():
    result = await generate_agent_graph(
        [agent],
        GraphOptions(
            title="My Agent System",
            output_path="./my-agents.png",
            color_scheme="modern",
            show_tool_details=True
        )
    )

    if result.success:
        print(f" Visualization saved to: {result.output_path}")
    else:
        print(f"❌ Error: {result.error}")

asyncio.run(main())

Features

• Multiple Color Schemes: Choose from default, modern, or minimal themes
• Agent Architecture: Visualize agents, tools, and handoff relationships
• Tool Ecosystems: Generate dedicated tool interaction diagrams
• Runner Architecture: Show complete system architecture with session layers
• Multiple Formats: Export as PNG, SVG, or PDF
• Customizable Layouts: Support for various Graphviz layouts (dot, circo, neato, etc.)

Example Output

The visualization system generates clear, professional diagrams showing:

• Agent Nodes: Rounded rectangles with agent names and model information
• Tool Nodes: Ellipses showing tool names and descriptions
• Handoff Edges: Dashed lines indicating agent handoff relationships
• Tool Connections: Colored edges connecting agents to their tools
• Cluster Organization: Grouped components in runner architecture views

Advanced Usage

from jaf.visualization import run_visualization_examples

# Run comprehensive examples
await run_visualization_examples()

# This generates multiple example files:
# - ./examples/agent-graph.png (agent system overview)
# - ./examples/tool-graph.png (tool ecosystem)
# - ./examples/runner-architecture.png (complete system)
# - ./examples/agent-modern.png (modern color scheme)

Key Components

Core Types

from dataclasses import dataclass
from pydantic import BaseModel, Field
from jaf import Agent, Tool, RunState, run

# Define your context type
@dataclass
class MyContext:
    user_id: str
    permissions: list[str]

# Define tool schema
class CalculateArgs(BaseModel):
    expression: str = Field(description="Math expression to evaluate")

# Create a tool
class CalculatorTool:
    @property
    def schema(self):
        return type('ToolSchema', (), {
            'name': 'calculate',
            'description': 'Perform mathematical calculations',
            'parameters': CalculateArgs
        })()

    async def execute(self, args: CalculateArgs, context: MyContext) -> str:
        result = eval(args.expression)  # Don't do this in production!
        return f"{args.expression} = {result}"

# Define an agent
def create_math_agent():
    def instructions(state):
        return 'You are a helpful math tutor'

    return Agent(
        name='MathTutor',
        instructions=instructions,
        tools=[CalculatorTool()]
    )

Running the Framework

import asyncio
from jaf import run, make_litellm_provider, generate_run_id, generate_trace_id
from jaf.core.types import RunState, RunConfig, Message

async def main():
    model_provider = make_litellm_provider('http://localhost:4000')
    math_agent = create_math_agent()

    config = RunConfig(
        agent_registry={'MathTutor': math_agent},
        model_provider=model_provider,
        max_turns=10,
        on_event=lambda event: print(event),  # Real-time tracing
    )

    initial_state = RunState(
        run_id=generate_run_id(),
        trace_id=generate_trace_id(),
        messages=[Message(role='user', content='What is 2 + 2?')],
        current_agent_name='MathTutor',
        context=MyContext(user_id='user123', permissions=['user']),
        turn_count=0,
    )

    result = await run(initial_state, config)
    print(result.outcome.output if result.outcome.status == 'completed' else result.outcome.error)

asyncio.run(main())

Security & Validation

Composable Validation Policies

from jaf.policies.validation import create_path_validator, create_permission_validator, compose_validations

# Create individual validators
path_validator = create_path_validator(['/shared', '/public'])
permission_validator = create_permission_validator('admin', lambda ctx: ctx)

# Compose them
combined_validator = compose_validations(path_validator, permission_validator)

# Apply to tools
secure_file_tool = with_validation(base_file_tool, combined_validator)

Guardrails

from jaf.policies.validation import create_content_filter, create_length_guardrail

config = RunConfig(
    # ... other config
    initial_input_guardrails=[
        create_content_filter(['spam', 'inappropriate']),  # Requires blocked patterns
        create_length_guardrail(max_length=1000, min_length=1)
    ],
    final_output_guardrails=[
        create_content_filter(['harmful', 'unsafe'])
    ],
)

Agent-as-Tool Functionality

JAF 2.2+ introduces powerful agent-as-tool capabilities, allowing you to use agents as tools within other agents for hierarchical orchestration:

from jaf.core.agent_tool import create_agent_tool
from jaf.core.types import create_json_output_extractor

# Create specialized agents
spanish_agent = Agent(
    name="spanish_translator",
    instructions=lambda state: "Translate text to Spanish",
    output_codec=TranslationOutput
)

french_agent = Agent(
    name="french_translator",
    instructions=lambda state: "Translate text to French",
    output_codec=TranslationOutput
)

# Convert agents to tools with conditional enabling
spanish_tool = spanish_agent.as_tool(
    tool_name="translate_to_spanish",
    tool_description="Translate text to Spanish",
    max_turns=3,
    custom_output_extractor=create_json_output_extractor(),
    is_enabled=True  # Always enabled
)

french_tool = french_agent.as_tool(
    tool_name="translate_to_french",
    tool_description="Translate text to French",
    max_turns=3,
    custom_output_extractor=create_json_output_extractor(),
    is_enabled=lambda context, agent: context.language_preference == "french_spanish"
)

# Create orchestrator agent using agent tools
orchestrator = Agent(
    name="translation_orchestrator",
    instructions=lambda state: "Use translation tools to respond in multiple languages",
    tools=[spanish_tool, french_tool]
)

Key Features:

• Conditional Enabling: Enable/disable agent tools based on runtime context
• Session Management: Configure whether sub-agents inherit parent session state
• Custom Output Extraction: Define how to extract and format agent tool outputs
• Error Handling: Robust error handling for failed agent tool executions

🔗 Agent Handoffs

from jaf.policies.handoff import create_handoff_guardrail, HandoffPolicy
from pydantic import BaseModel
from enum import Enum

class AgentName(str, Enum):
    MATH_TUTOR = 'MathTutor'
    FILE_MANAGER = 'FileManager'

class HandoffOutput(BaseModel):
    agent_name: AgentName

def create_triage_agent():
    def instructions(state):
        return 'Route requests to specialized agents'

    return Agent(
        name='TriageAgent',
        instructions=instructions,
        tools=[],  # Regular tools would go here
        handoffs=['MathTutor', 'FileManager'],  # Allowed handoff targets
        output_schema=HandoffOutput,
    )

Observability

Real-time Tracing

from jaf.core.tracing import ConsoleTraceCollector, FileTraceCollector, create_composite_trace_collector

# Console logging
console_tracer = ConsoleTraceCollector()

# File logging
file_tracer = FileTraceCollector('./traces.log')

# Composite tracing
tracer = create_composite_trace_collector(console_tracer, file_tracer)

config = RunConfig(
    # ... other config
    on_event=tracer.collect,
)

OpenTelemetry Integration

JAF 2.2+ includes built-in OpenTelemetry support for distributed tracing:

import os
from jaf.core.tracing import create_composite_trace_collector, ConsoleTraceCollector

# Configure OpenTelemetry endpoint
os.environ["TRACE_COLLECTOR_URL"] = "http://localhost:4318/v1/traces"

# Tracing will be automatically configured when creating a composite collector
trace_collector = create_composite_trace_collector(ConsoleTraceCollector())

config = RunConfig(
    # ... other config
    on_event=trace_collector.collect,
)

Langfuse Integration

For LLM-specific observability and analytics:

import os
from jaf.core.tracing import create_composite_trace_collector, ConsoleTraceCollector

# Configure Langfuse credentials
os.environ["LANGFUSE_PUBLIC_KEY"] = "pk-lf-your-public-key"
os.environ["LANGFUSE_SECRET_KEY"] = "sk-lf-your-secret-key"
os.environ["LANGFUSE_HOST"] = "https://cloud.langfuse.com"  # or your self-hosted instance

# Langfuse tracing will be automatically configured
trace_collector = create_composite_trace_collector(ConsoleTraceCollector())

config = RunConfig(
    # ... other config
    on_event=trace_collector.collect,
)

Error Handling

from jaf.core.errors import JAFErrorHandler

if result.outcome.status == 'error':
    formatted_error = JAFErrorHandler.format_error(result.outcome.error)
    is_retryable = JAFErrorHandler.is_retryable(result.outcome.error)
    severity = JAFErrorHandler.get_severity(result.outcome.error)

    print(f"[{severity}] {formatted_error} (retryable: {is_retryable})")

Provider Integrations

A2A (Agent-to-Agent) Communication

JAF provides a robust A2A communication layer that allows agents to interact with each other. This is useful for building multi-agent systems where different agents have specialized skills.

from jaf.a2a import create_a2a_agent, create_a2a_client, create_a2a_server

# Create agents
echo_agent = create_a2a_agent("EchoBot", "Echoes messages", "You are an echo bot.", [])
ping_agent = create_a2a_agent("PingBot", "Responds to pings", "You are a ping bot.", [])

# Create a server
server_config = {
    "agents": {"EchoBot": echo_agent, "PingBot": ping_agent},
    "agentCard": {"name": "Test Server"},
    "port": 8080,
}
server = create_a2a_server(server_config)

# Create a client
client = create_a2a_client("http://localhost:8080")

LiteLLM Provider

from jaf.providers.model import make_litellm_provider

# Connect to LiteLLM proxy for 100+ model support
model_provider = make_litellm_provider(
    'http://localhost:4000',  # LiteLLM proxy URL
    'your-api-key'           # Optional API key
)

MCP (Model Context Protocol) Integration

JAF includes full Model Context Protocol support for seamless tool integration:

from jaf.providers.mcp import create_mcp_stdio_client, create_mcp_tools_from_client

# Connect to MCP server via stdio
mcp_client = create_mcp_stdio_client(['python', '-m', 'my_mcp_server'])

# Initialize and get all available tools
await mcp_client.initialize()
mcp_tools = await create_mcp_tools_from_client(mcp_client)

# Use MCP tools in your agent
def create_mcp_agent():
    def instructions(state):
        return "You have access to powerful MCP tools for various tasks."

    return Agent(
        name='MCPAgent',
        instructions=instructions,
        tools=mcp_tools  # Automatically converted JAF tools
    )

# SSE client is also supported
from jaf.providers.mcp import create_mcp_sse_client
sse_client = create_mcp_sse_client('http://localhost:8080/sse')

Development Server

JAF includes a built-in development server for testing agents locally via HTTP endpoints:

from jaf.server import run_server
from jaf.providers.model import make_litellm_provider

def create_my_agent():
    def instructions(state):
        return 'You are a helpful assistant'

    return Agent(
        name='MyAgent',
        instructions=instructions,
        tools=[calculator_tool, greeting_tool]
    )

model_provider = make_litellm_provider('http://localhost:4000')

# Start server on port 3000
await run_server(
    [create_my_agent()],
    {'model_provider': model_provider},
    {'port': 3000}
)

Server provides RESTful endpoints:

• GET /health - Health check
• GET /agents - List available agents
• POST /chat - General chat endpoint
• POST /agents/{name}/chat - Agent-specific endpoint

Function Composition

JAF supports functional composition patterns for building complex behaviors from simple, reusable functions:

from jaf import create_function_tool, ToolSource

# Enhanced tool with caching and retry
def with_cache(tool_func):
    cache = {}
    async def cached_execute(args, context):
        cache_key = str(args)
        if cache_key in cache:
            return cache[cache_key]
        result = await tool_func(args, context)
        if result.status == "success":
            cache[cache_key] = result
        return result
    return cached_execute

# Create enhanced tools by composition
enhanced_tool = create_function_tool({
    'name': 'enhanced_search',
    'description': 'Search with caching',
    'execute': with_cache(base_search_function),
    'parameters': SearchArgs,
    'source': ToolSource.NATIVE
})

Key Benefits:

• Reusability: Write once, compose everywhere
• Testability: Each function can be tested in isolation
• Type Safety: Full type checking support
• Performance: Optimize individual pieces independently

Example Applications

Explore the example applications to see the framework in action:

1. Multi-Agent Server Demo

cd examples
python server_example.py

Features demonstrated:

• Multiple specialized agents (math, weather, general)
• Tool integration (calculator, weather API)
• Agent handoffs and routing
• RESTful API with auto-documentation
• Real-time tracing and error handling
• Production-ready server configuration

Available endpoints:

• GET /health - Server health check
• GET /agents - List all available agents
• POST /chat - Chat with any agent
• GET /docs - Interactive API documentation

2. Agent-as-Tool Demo

cd examples
python agent_as_tool_example.py

# Or start as server
python agent_as_tool_example.py --server

Features demonstrated:

• Hierarchical agent orchestration
• Conditional tool enabling based on context
• Custom output extraction from agent tools
• Session management for sub-agents
• Translation agents working together

3. Tracing Integration Demos

# OpenTelemetry tracing example
cd examples
python otel_tracing_demo.py

# Langfuse tracing example
python langfuse_tracing_demo.py

Features demonstrated:

• OpenTelemetry distributed tracing setup
• Langfuse LLM observability integration
• Composite trace collectors
• Real-time monitoring and analytics

4. MCP Integration Demo

cd examples/mcp_demo
python main.py

Features demonstrated:

• Model Context Protocol integration
• Dynamic tool loading from MCP servers
• Secure filesystem operations
• MCP client configuration and management

Testing

pytest          # Run tests
ruff check .    # Lint code
mypy .          # Type checking
black .         # Format code

Architecture Principles

Immutable State Machine

• All state transformations create new state objects
• No mutation of existing data structures
• Predictable, testable state transitions

Type Safety

• Runtime validation with Pydantic schemas
• Compile-time safety with Python type hints
• NewType for type-safe identifiers

Pure Functions

• Core logic is side-effect free
• Easy to test and reason about
• Deterministic behavior

Effect Isolation

• Side effects only in Provider modules
• Clear boundaries between pure and impure code
• Easier mocking and testing

📜 License

MIT

🤝 Contributing

1. Fork the repository
2. Create a feature branch
3. Add tests for new functionality
4. Run the test suite
5. Submit a pull request

JAF (Juspay Agentic Framework) v2.2 - Building the future of functional AI agent systems