agent-framework-lib 0.8.1.post6

A comprehensive Python framework for building and serving conversational AI agents with FastAPI

Agent Framework Library

A comprehensive Python framework for building and serving conversational AI agents with FastAPI. Create production-ready AI agents in minutes with automatic session management, streaming responses, file storage, and easy MCP integration.

Key Features:

• 🚀 Quick Setup - Create agents in 10-15 minutes
• 🔌 Easy MCP Integration - Connect to external tools effortlessly
• 🎯 Skills System - Markdown-based, on-demand capability loading for token optimization
• 🔄 Multi-Provider Support - OpenAI, Anthropic, Gemini
• 🎯 Smart Model Routing - Auto mode selects the best model per query complexity
• 💾 Session Management - Automatic conversation persistence
• 📁 File Storage - Local, S3, MinIO, GCP support
• 🤝 A2A Protocol - Agent-to-Agent communication via JSON-RPC
• 📊 Observability - Metrics, tracing, and logging via OpenTelemetry
• 🐘 PostgreSQL - Used by A2A Task Store and Memory Provider

Installation

Python: >=3.10,<3.14

# Install with LlamaIndex support (recommended)
uv add agent-framework-lib[llamaindex]

# Install with MCP support
uv add agent-framework-lib[llamaindex,mcp]

# Install with all features
uv add agent-framework-lib[all]

Available extras: llamaindex, mcp, mongodb, elasticsearch, postgresql, s3, minio, gcp, multimodal, memory, memori, graphiti, graphiti-falkordb, graphiti-neo4j, graphiti-all, observability, monitoring, websearch, dev, all

Optional: System Dependencies

The framework automatically detects and configures system libraries. Manual installation is only needed if you encounter issues:

For PDF Generation (WeasyPrint):

# macOS
brew install pango gdk-pixbuf libffi cairo

# Ubuntu/Debian
sudo apt-get install libpango-1.0-0 libpangoft2-1.0-0 libgdk-pixbuf2.0-0 libffi-dev libcairo2

# Fedora/RHEL
sudo dnf install pango gdk-pixbuf2 libffi-devel cairo

For Chart/Mermaid Image Generation (Playwright):

# Install Playwright and browser
uv add playwright
playwright install chromium

For MCP Python Server (Deno):

# macOS/Linux
curl -fsSL https://deno.land/install.sh | sh

# Windows (PowerShell)
irm https://deno.land/install.ps1 | iex

Post-Installation Script (Recommended)

The framework includes a CLI script that automatically installs all optional dependencies (Playwright browsers and Deno runtime):

# Run after installing the package
agent-framework-post-install

This script:

• ✅ Installs Playwright Chromium browser (for charts, mermaid diagrams, tables)
• ✅ Installs Deno runtime (for MCP servers like mcp-run-python)
• ✅ Works on Windows, macOS, and Linux
• ✅ Detects if dependencies are already installed (fast path)

Note: The framework also attempts lazy auto-installation when tools are first used, but running the post-install script ensures everything is ready upfront.

The framework handles library path configuration automatically on startup.

🤖 Framework Helper Agent

The framework includes a built-in AI assistant that helps you create agents! Access it at /helper when running any agent server.

Features:

• 🧠 Deep knowledge of framework documentation, examples, and source code
• 🔍 Search tools for docs and examples
• 💡 Code generation assistance
• 📚 Indexed knowledge base (30+ files)
• 🗄️ Persistent knowledge graph (FalkorDB) - survives server restarts

Access: http://localhost:8000/helper

The helper agent indexes:

• All documentation (docs/*.md)
• All examples (examples/*.py)
• Core framework source (tools, storage, memory, session management)

Re-indexing: If you update documentation or examples, trigger a re-index:

curl -X POST http://localhost:8000/helper/reindex

Model Configuration:

By default, the helper agent uses Claude (if ANTHROPIC_API_KEY is set) or GPT-5.2 (if OPENAI_API_KEY is set). You can override this with:

# Force a specific model (useful if your Anthropic key has reached its limit)
HELPER_AGENT_MODEL=gpt-5.2

Example questions:

• "How do I create an agent with memory?"
• "Show me how to use PDF tools"
• "What's the difference between Memori and Graphiti?"
• "How do I configure S3 storage?"
• "Search the web for LlamaIndex best practices"

🐳 Docker Development Environment

For local development, use Docker Compose to run all external services (Elasticsearch, MongoDB, PostgreSQL, FalkorDB, MinIO):

# Start all services
docker-compose --profile all up -d

# Copy environment template
cp .env.docker .env
# Edit .env to add your LLM API keys

# Stop services
docker-compose down

Use profiles to start only what you need:

docker-compose --profile storage up -d  # Elasticsearch, MongoDB, MinIO
docker-compose --profile memory up -d   # PostgreSQL, FalkorDB

Full documentation: See Docker Setup Guide for service details, ports, credentials, and troubleshooting.

🚀 Getting Started

Create Your First Agent

Here's a complete, working agent with LlamaIndex:

from typing import List
from agent_framework import LlamaIndexAgent, create_basic_agent_server

class MyAgent(LlamaIndexAgent):
    def __init__(self):
        super().__init__(
            agent_id="my_calculator_agent",
            name="Calculator Agent",
            description="A helpful calculator assistant that can perform basic math operations."
        )
    
    def get_agent_prompt(self) -> str:
        """Define your agent's behavior and personality."""
        return "You are a helpful calculator assistant."
  
    def get_agent_tools(self) -> List[callable]:
        """Define the tools your agent can use.
        
        Tools are automatically converted to LlamaIndex FunctionTool instances.
        The function name becomes the tool name, and the docstring becomes the description.
        """
        def add(a: float, b: float) -> float:
            """Add two numbers together."""
            return a + b
        
        def multiply(a: float, b: float) -> float:
            """Multiply two numbers together."""
            return a * b
        
        # Just return the functions - automatic conversion to FunctionTool
        return [add, multiply]

# Start server - includes streaming, session management, web UI
create_basic_agent_server(MyAgent, port=8000)

Required Methods:

• __init__() - Call super().__init__(agent_id, name, description) with required identity info
• get_agent_prompt() - Return system prompt string
• get_agent_tools() - Return list of tools (can be empty)

Optional Methods (have default implementations):

• create_fresh_context() - Create new LlamaIndex Context (default provided)
• serialize_context(ctx) - Serialize context for persistence (default provided)
• deserialize_context(state) - Deserialize context from state (default provided)
• initialize_agent() - Customize agent creation (default: FunctionAgent)
• configure_session() - Add session setup logic

That's it! The framework provides default implementations for context management (state persistence), so you only need to implement the three core methods above.

Run it:

# Set your API key
export OPENAI_API_KEY=sk-your-key-here

# Run the agent
uv run python my_agent.py

# Open http://localhost:8000/ui

⚙️ Configure Your Agent

Environment Setup

Create a .env file:

# Required: At least one API key
OPENAI_API_KEY=sk-your-openai-key
ANTHROPIC_API_KEY=sk-ant-your-anthropic-key
GEMINI_API_KEY=your-gemini-key

# Model Configuration
DEFAULT_MODEL=gpt-5.4-mini

# Multi-Model Routing (Auto Mode)
DEFAULT_MODEL_MODE=auto                    # "auto" or specific model name
AUTO_CLASSIFIER_MODEL=gpt-5.4-nano           # Model for complexity classification
PREFERRED_LIGHT_MODELS=gpt-5.4-nano,claude-haiku-4-5-20251001
PREFERRED_STANDARD_MODELS=gpt-5.4-mini
PREFERRED_ADVANCED_MODELS=gpt-5.2,claude-opus-4-6

# Session Storage (optional)
SESSION_STORAGE_TYPE=memory  # or "mongodb" or "elasticsearch"
MONGODB_CONNECTION_STRING=mongodb://localhost:27017
MONGODB_DATABASE_NAME=agent_sessions

# File Storage (optional)
LOCAL_STORAGE_PATH=./file_storage
AWS_S3_BUCKET=my-bucket
S3_AS_DEFAULT=false

Remote Configuration (Elasticsearch-Managed Agents)

For production deployments, you can configure agents to be managed entirely via Elasticsearch, allowing ops teams to modify prompts and models at runtime without code deployments.

Enable remote configuration:

from agent_framework import LlamaIndexAgent

class OpsMangedAgent(LlamaIndexAgent):
    def __init__(self):
        super().__init__(
            agent_id="ops_managed_agent",
            name="Ops Managed Agent",
            description="An agent configured via Elasticsearch."
        )
    
    @classmethod
    def get_use_remote_config(cls) -> bool:
        """Enable Elasticsearch-only configuration."""
        return True
    
    def get_agent_prompt(self) -> str:
        # Fallback prompt if ES config not available
        return "You are a helpful assistant."
    
    def get_agent_tools(self) -> list:
        return []

Behavior:

use_remote_config	Server Startup	Session Init
False (default)	Pushes hardcoded config to ES if different	Merges ES config with hardcoded
True	Skips pushing to ES	Reads ES config only (no merge)

When to use:

• use_remote_config=False (default): Code-managed agents where developers control the config
• use_remote_config=True: Ops-managed agents where configuration is modified via ES/Kibana

Fallback: If use_remote_config=True but no ES config exists, the system falls back to hardcoded config and pushes it to ES with a warning.

🎯 Multi-Model Selection

The framework includes intelligent model routing that automatically selects the best model based on query complexity.

Auto Mode (Default)

When DEFAULT_MODEL_MODE=auto, the system analyzes each query and routes it to the appropriate tier:

Tier	Icon	Use Case	Example Models
Light	💨	Simple queries, greetings, basic info	gpt-5.4-nano, claude-haiku-4-5
Standard	⚖️	Typical questions, explanations	gpt-5.4-mini, claude-sonnet-4-6
Advanced	🧠	Complex analysis, creative tasks	gpt-5.2, claude-opus-4-6

Benefits:

• 💰 Cost optimization - Use cheaper models for simple queries
• ⚡ Speed - Faster responses for trivial messages
• 🎯 Quality - Powerful models for complex tasks

Manual Model Selection

Users can also select a specific model from the UI dropdown:

• Models grouped by tier with availability indicators (✓/✗)
• Preference persisted in localStorage
• Real-time routing indicator shows selected model

Configuration

# Default mode when no user preference
DEFAULT_MODEL_MODE=auto

# Model used for complexity classification (should be fast and cheap)
AUTO_CLASSIFIER_MODEL=gpt-5.4-nano

# Preferred models per tier (comma-separated, in order of preference)
PREFERRED_LIGHT_MODELS=gpt-5.4-nano,claude-haiku-4-5-20251001,gemini-2.5-flash-lite
PREFERRED_STANDARD_MODELS=gpt-5.4-mini,gemini-2.5-flash
PREFERRED_ADVANCED_MODELS=gpt-5.2,claude-opus-4-6,gemini-2.5-pro

API Endpoint

# Get available models
curl http://localhost:8000/api/models

# Response
{
  "models_by_tier": {
    "light": [{"id": "gpt-5.4-nano", "provider": "openai", "available": true}, ...],
    "standard": [...],
    "advanced": [...]
  },
  "default_mode": "auto",
  "classifier_model": "gpt-5.4-nano"
}

Backward Compatibility

Agents with hardcoded models continue to work without changes:

class MyAgent(LlamaIndexAgent):
    def __init__(self):
        super().__init__(...)
        self._default_model = "gpt-5.2"  # This model will always be used

LlamaIndex Agent Configuration

Control model behavior in your agent:

class MyAgent(LlamaIndexAgent):
    def __init__(self):
        super().__init__(
            agent_id="my_agent",
            name="My Agent",
            description="A helpful assistant."
        )
        # Default model config (can be overridden per session)
        self.default_temperature = 0.7
        self.default_model = "gpt-5.4-mini"

Runtime Configuration:

Users can override settings per session via the API or web UI:

• Model selection (gpt-5.2, claude-sonnet-4-6, gemini-2.5-pro)
• Temperature (0.0 - 1.0)
• Max tokens
• System prompt override

🔧 Create Custom Tools

Custom tools extend your agent's capabilities. The tool name and docstring are crucial - they tell the agent when and how to use the tool.

Basic Custom Tool

def get_weather(city: str) -> str:
    """Get the current weather for a specific city.
    
    Args:
        city: The name of the city to get weather for
        
    Returns:
        A description of the current weather
    """
    # Your implementation here
    return f"The weather in {city} is sunny, 22°C"

# Add to your agent
class MyAgent(LlamaIndexAgent):
    def get_agent_tools(self):
        # Just return the function - automatic conversion to FunctionTool
        # Function name = tool name, docstring = tool description
        return [get_weather]

Important:

• Function name should be explicit and descriptive (e.g., get_weather, not weather)
• Docstring is added as the tool description - the agent uses this to understand when to call the tool
• Type hints help the agent understand parameters
• Args/Returns documentation provides additional context

Custom Tool with Dependencies

For tools that need file storage or other dependencies, use closures to capture context:

from agent_framework import LlamaIndexAgent
from agent_framework.storage.file_system_management import FileStorageFactory

class MyAgent(LlamaIndexAgent):
    def __init__(self):
        super().__init__(
            agent_id="my_agent",
            name="My Agent",
            description="A helpful assistant with custom tools."
        )
        self.file_storage = None
    
    async def _ensure_file_storage(self):
        if self.file_storage is None:
            self.file_storage = await FileStorageFactory.create_storage_manager()
    
    async def configure_session(self, session_configuration):
        await self._ensure_file_storage()
        self._user_id = session_configuration.get('user_id', 'default_user')
        self._session_id = session_configuration.get('session_id')
        await super().configure_session(session_configuration)
    
    def get_agent_tools(self):
        storage = self.file_storage
        user_id = self._user_id
        session_id = self._session_id
        
        async def store_result(param1: str, param2: int) -> str:
            """Process data and store results.
            
            Args:
                param1: Description of first parameter
                param2: Description of second parameter
                
            Returns:
                Result description
            """
            result = f"Processed {param1} with {param2}"
            file_id = await storage.store_file(
                user_id=user_id,
                session_id=session_id,
                filename="result.txt",
                content=result.encode()
            )
            return f"Result stored with ID: {file_id}"
        
        return [store_result]

Tool Naming Best Practices

# ✅ GOOD - Explicit and clear
def calculate_mortgage_payment(principal: float, rate: float, years: int) -> float:
    """Calculate monthly mortgage payment."""
    pass

def send_email_notification(recipient: str, subject: str, body: str) -> bool:
    """Send an email notification to a recipient."""
    pass

# ❌ BAD - Too vague
def calculate(x: float, y: float) -> float:
    """Do calculation."""
    pass

def send(data: str) -> bool:
    """Send something."""
    pass

🔌 Adding MCP Servers

MCP (Model Context Protocol) allows your agent to connect to external tools and services.

Basic MCP Setup

from llama_index.tools.mcp import BasicMCPClient, McpToolSpec

class MyAgent(LlamaIndexAgent):
    def __init__(self):
        super().__init__(
            agent_id="my_agent",
            name="MCP Agent",
            description="An assistant with access to external tools via MCP servers."
        )
        self.mcp_tools = []
        self._mcp_initialized = False
    
    async def _initialize_mcp_tools(self):
        """Load tools from MCP servers."""
        if self._mcp_initialized:
            return
        
        # Configure your MCP server
        mcp_configs = [
            {
                "command": "uvx",
                "args": ["mcp-server-filesystem"],
                "env": {"FILESYSTEM_ROOT": "/path/to/workspace"}
            }
        ]
        
        for config in mcp_configs:
            client = BasicMCPClient(
                config["command"],
                args=config["args"],
                env=config.get("env", {})
            )
            
            # Load tools from the MCP server
            mcp_tool_spec = McpToolSpec(client=client)
            tools = await mcp_tool_spec.to_tool_list_async()
            self.mcp_tools.extend(tools)
        
        self._mcp_initialized = True
    
    async def initialize_agent(self, model_name, system_prompt, tools, **kwargs):
        # Load MCP tools before initializing agent
        await self._initialize_mcp_tools()
        
        # Combine with other tools
        all_tools = self.get_agent_tools()
        await super().initialize_agent(model_name, system_prompt, all_tools, **kwargs)
    
    def get_agent_tools(self):
        # Return built-in tools + MCP tools
        return self.mcp_tools

Multiple MCP Servers

def _get_mcp_configs(self):
    """Configure multiple MCP servers."""
    return [
        {
            "name": "filesystem",
            "command": "uvx",
            "args": ["mcp-server-filesystem"],
            "env": {"FILESYSTEM_ROOT": "/workspace"}
        },
        {
            "name": "github",
            "command": "uvx",
            "args": ["mcp-server-github"],
            "env": {
                "GITHUB_TOKEN": os.getenv("GITHUB_TOKEN")
            }
        },
        {
            "name": "python",
            "command": "uvx",
            "args": ["mcp-run-python", "stdio"]
        }
    ]

Popular MCP Servers

# Filesystem operations
uvx mcp-server-filesystem

# GitHub integration
uvx mcp-server-github

# Python code execution
uvx mcp-run-python

# Database access
uvx mcp-neo4j-cypher
uvx mcp-server-postgres

Installation:

# Install with MCP support
uv add agent-framework-lib[llamaindex,mcp]

# Or add MCP to existing installation
uv add agent-framework-lib[mcp]

# MCP servers are run via uvx (no separate install needed)

Using Deno-based MCP servers:

If you need to use Deno-based MCP servers (like TypeScript MCP servers), the framework provides a helper function to ensure Deno works correctly even if it's not in your PATH:

from agent_framework import get_deno_command

# Configure a Deno-based MCP server
mcp_config = {
    "command": get_deno_command(),  # Automatically uses correct Deno path
    "args": ["run", "-N", "jsr:@pydantic/mcp-run-python", "stdio"]
}

This helper function:

• ✅ Automatically finds Deno even if not in system PATH
• ✅ Works seamlessly after agent-framework-post-install
• ✅ Returns absolute path to Deno binary when needed

🧠 Memory Module

Add long-term semantic memory to your agents, enabling them to remember information across conversations and provide personalized responses.

Quick Start

from agent_framework import LlamaIndexAgent
from agent_framework.memory import MemoryConfig

class MyMemoryAgent(LlamaIndexAgent):
    def __init__(self):
        super().__init__(
            agent_id="memory_agent",
            name="Memory Agent",
            description="An agent with long-term memory."
        )
    
    def get_agent_prompt(self) -> str:
        return "You are a helpful assistant that remembers user preferences."
    
    def get_agent_tools(self) -> list:
        return []
    
    def get_memory_config(self):
        """Enable memory - just override this method!"""
        return MemoryConfig.memori_simple(
            database_url="sqlite:///memory.db"
        )

Memory Providers

Provider	Backend	Best For
Memori	SQLite, PostgreSQL, MySQL	Fast queries, simple setup
Graphiti	FalkorDB, Neo4j	Complex relationships, temporal queries
Hybrid	Both	Best of both worlds

Configuration Options

# Memori with SQLite (simplest)
MemoryConfig.memori_simple(database_url="sqlite:///memory.db")

# Graphiti with FalkorDB
MemoryConfig.graphiti_simple(use_falkordb=True)

# Hybrid mode (both providers)
MemoryConfig.hybrid(
    memori_database_url="sqlite:///memory.db",
    graphiti_use_falkordb=True
)

Memory Modes

• Passive Injection: Relevant memories automatically injected into prompts
• Active Tools: Agent can explicitly recall_memory(), store_memory(), forget_memory()

Installation

# All memory support
uv add agent-framework-lib[memory]

# Or individual providers
uv add agent-framework-lib[memori]
uv add agent-framework-lib[graphiti]

More info: See Memory Installation Guide and Creating Agents Guide

🎯 Skills System

The Skills System provides modular, on-demand capability loading that reduces token consumption by ~80%. Instead of loading all instructions into every system prompt, skills deliver detailed instructions only when needed.

Skills are exclusively defined as SKILL.md markdown files with YAML frontmatter, loaded by MarkdownSkillLoader. Each skill uses ShellTool to execute standalone CLI scripts and WebFetchTool for web content retrieval.

How It Works

BEFORE: System Prompt = Base (~500) + Rich Content (~3000) = ~3500 tokens/message
AFTER:  System Prompt = Base (~500) + Skills Discovery (~200) = ~700 tokens/message
        + On-demand skill loading (~500 tokens, one-time per skill)

Quick Start

Skills are automatically available in all agents via BaseAgent. No need to explicitly inherit from SkillsMixin:

from agent_framework import LlamaIndexAgent

class MySkillsAgent(LlamaIndexAgent):
    def __init__(self):
        super().__init__(
            agent_id="skills_agent",
            name="Skills Agent",
            description="An agent with on-demand capabilities."
        )
        # Built-in markdown skills are automatically registered by BaseAgent.__init__
    
    def get_agent_prompt(self) -> str:
        # Skills discovery prompt is automatically appended by BaseAgent
        return "You are a helpful assistant."
    
    def get_agent_tools(self) -> list:
        # Skill tools are auto-loaded - no need to add them manually!
        return []  # Only return custom tools specific to your agent

Built-in Skills (21 total)

Category	Skills
Visualization	chart, mermaid, table
Document	file, unified_pdf, file_access, excel, drawio, powerpoint, word
Web	web_news_search
Multimodal	multimodal, image_gen
UI	form, optionsblock, image_display, email_template, skill_creator
Data	csv, data_format
Code	code_format

New in v0.9.0: 8 additional skills added (in bold):

• powerpoint — Generate .pptx presentations with slides, layouts, and themes
• word — Create .docx documents with formatted text, tables, and images
• image_gen — AI image generation via DALL-E 3 and DALL-E 2
• csv — Create, read, and transform CSV files
• data_format — Convert JSON ↔ YAML and validate schemas
• code_format — Format Python, JavaScript, JSON, YAML code
• email_template — Generate responsive HTML email templates
• skill_creator — Guide for creating custom skills via API or code

Each skill is a SKILL.md file in agent_framework/skills/builtin/skills/ with an associated CLI script executed via ShellTool.

Agent Workflow

1. Agent receives user request: "Create a bar chart"
2. Agent calls list_skills() → sees available skills
3. Agent calls load_skill("chart") → gets chart instructions from SKILL.md
4. Agent constructs the shell command as described in the instructions
5. Agent executes via shell_exec → script generates the chart
6. Optionally calls unload_skill("chart") when done

More info: See Custom Skills Guide, Creating Agents Guide and skills_demo_agent.py

📝 Rich Content Capabilities (Automatic)

All agents automatically support rich content generation including:

• 📊 Mermaid diagrams (version 10.x syntax)
• 📈 Chart.js charts (bar, line, pie, doughnut, polarArea, radar, scatter, bubble)
• 📋 Interactive forms (formDefinition JSON)
• 🔘 Clickable option buttons (optionsblock)
• 📑 Formatted tables (tabledata)

This is automatic! The framework injects rich content instructions into all agent system prompts by default. You don't need to add anything to your get_agent_prompt().

Disabling Rich Content

If you need to disable automatic rich content injection for a specific agent or session:

Via Session Configuration (UI or API):

# When initializing a session
session_config = {
    "user_id": "user123",
    "session_id": "session456",
    "enable_rich_content": False  # Disable rich content
}

Via Web UI: Uncheck the "Enable rich content capabilities" checkbox when creating a session.

Format Examples

Chart:

```chart
{
  "type": "chartjs",
  "chartConfig": {
    "type": "bar",
    "data": {
      "labels": ["Mon", "Tue", "Wed"],
      "datasets": [{
        "label": "Sales",
        "data": [120, 150, 100]
      }]
    }
  }
}
```

Options Block:

```optionsblock
{
  "question": "What would you like to do?",
  "options": [
    {"text": "Continue", "value": "continue"},
    {"text": "Cancel", "value": "cancel"}
  ]
}
```

Table:

```tabledata
{
  "caption": "Sales Data",
  "headers": ["Month", "Revenue"],
  "rows": [["Jan", "$1000"], ["Feb", "$1200"]]
}
```

🌐 Web Interface

The framework includes a built-in web UI for testing and interacting with your agent.

Access: http://localhost:8000/ui

Features:

• 💬 Real-time message streaming
• 🎨 Rich format rendering (charts, tables, mermaid diagrams)
• 📁 File upload and management
• ⚙️ Model and parameter configuration
• 💾 Session management
• 📊 Conversation history
• 🎯 Interactive option blocks and forms

Quick Test:

# Start your agent
uv run python my_agent.py

# Open in browser
open http://localhost:8000/ui

The UI automatically detects and renders:

• Chart.js visualizations from chart blocks
• Mermaid diagrams from mermaid blocks
• Tables from tabledata blocks
• Interactive forms from formDefinition JSON
• Clickable options from optionsblock

API Documentation: http://localhost:8000/docs (Swagger UI)

📚 Additional Resources

Documentation

• Installation Guide - Detailed setup instructions
• Configuration Guide - Environment and settings configuration
• Creating Agents Guide - Guide to building custom agents
• Tools and MCP Guide - Tools and MCP integration
• Memory Installation Guide - Memory module setup
• API Reference - Complete API documentation
• A2A Guide - Agent-to-Agent protocol documentation
• Observability Guide - Metrics, tracing, and logging
• Architecture Diagram - System architecture overview
• File Storage Guide - Local, S3, MinIO, GCP storage backends
• Multimodal Tools Guide - Multimodal processing tools
• Streaming Events Frontend - SSE event format for frontends
• Docker Setup Guide - Docker Compose development environment

Examples

• Simple Agent - Basic calculator agent
• File Storage Agent - File management
• MCP Integration - MCP integration
• Memory Agent - Agent with long-term memory
• Multi-Skills Agent - Complete multi-skills agent
• Custom Framework Agent - Custom framework implementation
• Skills Demo Agent - Skills system demonstration

API Endpoints

Core:

• POST /message - Send message to agent
• POST /init - Initialize session
• POST /end - End session
• GET /sessions - List sessions

Files:

• POST /files/upload - Upload file
• GET /files/{file_id}/download - Download file
• GET /files - List files

Full API docs: http://localhost:8000/docs

Authentication

# API Key Authentication
REQUIRE_AUTH=true
API_KEYS=sk-key-1,sk-key-2

curl -H "Authorization: Bearer sk-key-1" \
  http://localhost:8000/message \
  -H "Content-Type: application/json" \
  -d '{"query": "Hello!"}'

---

Quick Links:

• 🎨 Web UI
• 📖 API Docs
• ⚙️ Config Test