🦆 DevDuck

Self-modifying AI agent that hot-reloads its own code—builds itself as it runs.

One Python file that adapts to your environment, fixes itself, and expands capabilities at runtime.

Learn more: https://duck.nyc

🎬 See It In Action

Feature	What You'll See	Demo
🔥 Hot-Reload	Agent detects code changes and restarts instantly	Watch
🌐 Web UI	Clean web interface with real-time streaming	Watch
🛠️ Dynamic Tools	Save .py file in ./tools/ → use instantly	Watch
🌊 TCP Streaming	Connect via netcat, apps, or other agents	Watch
🔗 Zenoh P2P	Auto-discover & coordinate multiple DevDucks	Multi-terminal magic ✨
🎬 Session Recording	Record, replay & resume agent sessions	Time-travel debugging 🕰️
🌙 Ambient Mode	Background thinking while you're idle	Auto-explores topics 🧠
🔌 IPC & Tray	macOS menu bar + Unix socket IPC
💬 Ambient Overlay	Floating AI input with glassmorphism UI	Watch

Quick Start

# Install & run
pipx install devduck && devduck

# With speech-to-speech capabilities (optional)
pipx install "devduck[speech]" && devduck

# One-shot query
devduck "create a REST API with FastAPI"

# Python API
python -c "import devduck; devduck('analyze this code')"

# Session recording (time-travel debugging)
devduck --record "analyze this codebase"
# → Exports to /tmp/devduck/recordings/session-*.zip

# Resume from recorded session
devduck --resume session-20250202-123456.zip "continue where we left off"

Requirements: Python 3.10-3.13, AWS credentials (or Ollama/Anthropic/GitHub/MLX)

Optional extras:

• devduck[speech] - Real-time speech-to-speech conversations (Nova Sonic, OpenAI Realtime, Gemini Live)

Core Capabilities

Developer Setup

git clone git@github.com:cagataycali/devduck.git
cd devduck
python3.13 -m venv .venv
source .venv/bin/activate

# Basic install
.venv/bin/pip3.13 install -e .

# With speech capabilities
.venv/bin/pip3.13 install -e ".[speech]"

devduck

Architecture

graph TB
    A[User Input] -->|CLI/TCP/WS/MCP/IPC| B[DevDuck Core]
    B -->|Auto RAG| C[Knowledge Base]
    C -.->|Context Retrieval| B
    B -->|Tool Calls| D[40+ Built-in Tools]
    D --> E[shell/editor/calculator]
    D --> F[GitHub/AgentCore]
    D --> G[TCP/WebSocket/MCP/IPC]
    D --> H[tray/ambient/cursor/clipboard]
    B -->|Hot-reload| I[./tools/*.py + __init__.py]
    I -.->|Load Instantly| D
    B -->|Runtime| K[manage_tools/install_tools]
    K -.->|Expand| D
    B <-->|Zenoh P2P| L[Other DevDucks]
    L -.->|Auto-discover| B
    B -->|Response| J[User Output]
    J -.->|Store Memory| C
    
    style B fill:#e1f5ff
    style C fill:#d4edda
    style I fill:#fff3cd
    style K fill:#ffe6cc
    style L fill:#f0e6ff

Self-adapting loop: Query → RAG → Tools → Response → Memory → Hot-reload/Runtime-load → Repeat

Model Setup

DevDuck auto-detects providers based on credentials:

Priority: Bedrock → Anthropic → OpenAI → GitHub → Gemini → Cohere → Writer → Mistral → LiteLLM → LlamaAPI → MLX → Ollama

Just set your API key - DevDuck handles the rest:

export ANTHROPIC_API_KEY=sk-ant-...
devduck  # Auto-uses Anthropic

export OPENAI_API_KEY=sk-...
devduck  # Auto-uses OpenAI

export GOOGLE_API_KEY=...
devduck  # Auto-uses Gemini

Manual override:

export MODEL_PROVIDER=bedrock
export STRANDS_MODEL_ID=us.anthropic.claude-sonnet-4-20250514-v1:0
devduck

Tool Management

Runtime Tool Management

Add, remove, or reload tools while agent is running:

# List all loaded tools
manage_tools(action="list")

# Add tools from a package at runtime
manage_tools(action="add", package="strands_fun_tools", tool_names="cursor,clipboard,bluetooth")

# Remove tools you don't need
manage_tools(action="remove", tool_names="cursor,clipboard")

# Reload specific tools after editing
manage_tools(action="reload", tool_names="shell,editor")

# Reload all tools (restarts agent)
manage_tools(action="reload")

# Load custom tool from file
manage_tools(action="add", tool_path="./my_custom_tool.py")

Dynamic Package Installation

Install Python packages and load their tools at runtime:

# Discover available tools before loading
install_tools(action="list_available", package="strands-fun-tools", module="strands_fun_tools")

# Install package and load all tools
install_tools(action="install_and_load", package="strands-agents-tools", module="strands_tools")

# Install and load specific tools only
install_tools(
    action="install_and_load",
    package="strands-fun-tools", 
    module="strands_fun_tools",
    tool_names=["clipboard", "cursor", "bluetooth"]
)

# Load tools from already installed package
install_tools(action="load", module="strands_tools", tool_names=["shell", "calculator"])

# List currently loaded tools
install_tools(action="list_loaded")

Static Tool Configuration

Format: package1:tool1,tool2;package2:tool3,tool4

# Minimal (shell + editor only)
export DEVDUCK_TOOLS="strands_tools:shell,editor"

# Dev essentials
export DEVDUCK_TOOLS="strands_tools:shell,editor,file_read,file_write,calculator"

# Full stack + GitHub
export DEVDUCK_TOOLS="devduck.tools:tcp,websocket,mcp_server,use_github;strands_tools:shell,editor,file_read"

devduck

Hot-Reload Tools from Directory

Create ./tools/weather.py:

from strands import tool
import requests

@tool
def weather(city: str) -> str:
    """Get weather for a city."""
    r = requests.get(f"https://wttr.in/{city}?format=%C+%t")
    return r.text

Enable directory auto-loading:

export DEVDUCK_LOAD_TOOLS_FROM_DIR=true
devduck
# Save weather.py → use instantly (no restart needed)

Default: Directory loading is OFF. Use manage_tools() or install_tools() for explicit control.

Speech-to-Speech (Optional)

Install speech capabilities:

pip install "devduck[speech]"

Real-time voice conversations with multiple providers:

# Start speech session with Nova Sonic (AWS Bedrock)
speech_to_speech(action="start", provider="novasonic")

# Start with OpenAI Realtime API
speech_to_speech(action="start", provider="openai")

# Start with Gemini Live
speech_to_speech(action="start", provider="gemini_live")

# Custom voice and settings
speech_to_speech(
    action="start",
    provider="novasonic",
    model_settings={
        "provider_config": {"audio": {"voice": "matthew"}},
        "client_config": {"region": "us-east-1"}
    }
)

# Stop session
speech_to_speech(action="stop", session_id="speech_20250126_140000")

# Check status
speech_to_speech(action="status")

# List conversation histories
speech_to_speech(action="list_history")

# List available audio devices
speech_to_speech(action="list_audio_devices")

Supported Providers:

• Nova Sonic (AWS Bedrock): 11 voices (English, French, Italian, German, Spanish)
• OpenAI Realtime API: GPT-4o Realtime models
• Gemini Live: Native audio streaming

Environment Variables:

• OPENAI_API_KEY - For OpenAI Realtime
• GOOGLE_API_KEY or GEMINI_API_KEY - For Gemini Live
• AWS credentials - For Nova Sonic (boto3 default credential chain)

Features:

• Background execution (parent agent stays responsive)
• Tool inheritance from parent agent
• Conversation history saved automatically
• Natural interruption with VAD
• Custom audio device selection

MCP Integration

As MCP Server (Expose DevDuck)

Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "devduck": {
      "command": "uvx",
      "args": ["devduck", "--mcp"]
    }
  }
}

Or start HTTP MCP server:

mcp_server(action="start", port=8000, stateless=True)
# Connect at: http://localhost:8000/mcp

Modes: --mcp (stdio for Claude Desktop) | http (background server) | stateless=True (multi-node)

As MCP Client (Load External Servers)

Expand capabilities by loading tools from external MCP servers:

export MCP_SERVERS='{
  "mcpServers": {
    "strands-docs": {"command": "uvx", "args": ["strands-agents-mcp-server"]},
    "remote": {"url": "https://api.example.com/mcp", "headers": {"Auth": "Bearer token"}},
    "custom": {"command": "python", "args": ["my_server.py"]}
  }
}'
devduck

Supported transports: stdio (command/args/env) | HTTP (url/headers) | SSE (url with /sse path)

Tool prefixing: Each server's tools get prefixed (e.g., strands-docs_search_docs)

Zenoh Peer-to-Peer Networking

Auto-discover and coordinate multiple DevDuck instances across terminals or networks.

How It Works

1. Each DevDuck joins a Zenoh peer network
2. Multicast scouting (224.0.0.224:7446) auto-discovers peers on local network
3. Peers exchange heartbeats to maintain presence awareness
4. Commands can be broadcast to ALL peers or sent to specific peers
5. Responses stream back in real-time

Quick Start

# Terminal 1: Start DevDuck (Zenoh enabled by default)
devduck
# 🦆 ✓ Zenoh peer: hostname-abc123

# Terminal 2: Start another DevDuck
devduck
# 🦆 ✓ Zenoh peer: hostname-def456
# Auto-discovers Terminal 1!

# Terminal 1: See discovered peers
🦆 zenoh_peer(action="list_peers")

# Terminal 1: Broadcast to ALL DevDucks
🦆 zenoh_peer(action="broadcast", message="git status")
# Both terminals execute and stream responses!

# Send to specific peer
🦆 zenoh_peer(action="send", peer_id="hostname-def456", message="what files are here?")

Cross-Network Connections

Connect DevDuck instances across different networks:

# Machine A (office): Listen for remote connections
export ZENOH_LISTEN="tcp/0.0.0.0:7447"
devduck

# Machine B (home): Connect to office
export ZENOH_CONNECT="tcp/office.example.com:7447"
devduck

# Now they can communicate!
🦆 zenoh_peer(action="broadcast", message="sync all repos")

Use Cases

Actions

# Start Zenoh networking (auto-starts by default)
zenoh_peer(action="start")

# Stop Zenoh
zenoh_peer(action="stop")

# Check status and peer count
zenoh_peer(action="status")

# List all discovered peers
zenoh_peer(action="list_peers")

# Broadcast to ALL peers (waits for responses)
zenoh_peer(action="broadcast", message="your command", wait_time=60)

# Send to specific peer
zenoh_peer(action="send", peer_id="hostname-abc123", message="your command", wait_time=120)

# Start with remote connection
zenoh_peer(action="start", connect="tcp/remote.example.com:7447")

# Start listening for remote connections
zenoh_peer(action="start", listen="tcp/0.0.0.0:7447")

Environment Variables

Features

• Zero Config: Just start multiple DevDucks - they find each other
• Real-time Streaming: Responses stream as they're generated
• Peer Awareness: Dynamic context injection shows connected peers
• Cross-Network: Connect instances across different networks
• Fault Tolerant: Peers auto-detect disconnections (15s timeout)

Advanced Features

🎬 Session Recording (Time-Travel Debugging)

Record entire sessions for replay, debugging, and state restoration:

# CLI: Start with recording enabled
devduck --record
devduck --record "analyze this codebase"

# Resume from recorded session
devduck --resume ~/Desktop/session-20250202-123456.zip
devduck --resume session.zip "continue where we left off"

# Resume from specific snapshot
devduck --resume session.zip --snapshot 2 "what was I working on?"

Interactive recording:

🦆 record              # Toggle recording on/off
🦆 session_recorder(action="start")
🦆 session_recorder(action="snapshot", description="before refactor")
🦆 session_recorder(action="stop")  # Exports to /tmp/devduck/recordings/

Captures three layers:

• sys: OS-level events (file I/O, HTTP requests)
• tool: All tool calls and results
• agent: Messages, decisions, state changes

Python API for session analysis:

from devduck import load_session, resume_session, list_sessions

# List all recordings
sessions = list_sessions()
# [{'name': 'session-20250202-123456.zip', 'size_kb': 45.2, ...}]

# Load and analyze a session
session = load_session("~/Desktop/session-20250202-123456.zip")
print(session)  # LoadedSession(events=156, snapshots=3, duration=342.5s)

# Get events by layer
tool_calls = session.get_events_by_layer("tool")
file_ops = session.get_events_by_type("file.open")

# Resume from snapshot (restores conversation history!)
result = session.resume_from_snapshot(2, agent=devduck.agent)
print(f"Restored {result['messages_restored']} messages")

# Resume and continue with new query
result = session.resume_and_continue(2, "what files did we modify?", devduck.agent)
print(result['agent_result'])

# Replay with callback
def on_event(event, idx):
    print(f"[{idx}] {event.layer}/{event.event_type}: {event.data}")
session.replay_events(callback=on_event)

Session file structure (ZIP):

session-20250202-123456.zip
├── events.jsonl      # All events in JSON Lines format
├── snapshots.json    # State snapshots with conversation history
├── metadata.json     # Session info (duration, hostname, etc.)
└── session.pkl       # Serialized state for full restore (dill/pickle)

Recordings saved to: /tmp/devduck/recordings/

🌙 Ambient Mode (Background Thinking)

Continue working in the background while you're idle:

# Enable via environment
export DEVDUCK_AMBIENT_MODE=true
devduck

# Or toggle in REPL
🦆 ambient     # Toggle standard ambient mode
🦆 auto        # Toggle autonomous mode

Standard Mode: Runs up to 3 iterations when you go idle (30s)

# Configuration
export DEVDUCK_AMBIENT_IDLE_SECONDS=30      # Wait before starting
export DEVDUCK_AMBIENT_MAX_ITERATIONS=3     # Max background iterations
export DEVDUCK_AMBIENT_COOLDOWN=60          # Seconds between runs

Autonomous Mode: Runs continuously until done or stopped

export DEVDUCK_AUTONOMOUS_MAX_ITERATIONS=50  # Higher limit
export DEVDUCK_AUTONOMOUS_COOLDOWN=10        # Faster cycles

How it works:

1. You go idle (30s default)
2. DevDuck continues exploring the last topic
3. Background work streams with 🌙 prefix
4. When you return, findings are injected into your next query
5. Agent can signal completion with [AMBIENT_DONE]

Programmatic control:

# Enable standard ambient mode
devduck.ambient.start()

# Enable autonomous mode
devduck.ambient.start(autonomous=True)

# Stop ambient mode
devduck.ambient.stop()

# Check status
devduck.status()['ambient_mode']

State Management (Time-Travel)

Save and restore agent state for reproducibility:

# Export current state
state_manager(action="export", metadata={"note": "before refactor"})

# List saved states
state_manager(action="list")

# Load and display state
state_manager(action="load", state_file="~/.devduck/states/devduck_20250118_150000.pkl")

# Resume from state (ephemeral - doesn't mutate parent)
state_manager(
    action="resume", 
    state_file="~/.devduck/states/devduck_20250118_150000.pkl",
    query="continue the analysis from where we left off"
)

# Modify state metadata
state_manager(
    action="modify",
    state_file="path/to/state.pkl",
    metadata={"tags": ["important", "refactor"]}
)

# Delete state
state_manager(action="delete", state_file="path/to/state.pkl")

States saved to: ~/.devduck/states/

System Prompt Management

Self-improvement - agent updates its own system prompt:

# View current system prompt
system_prompt(action="view")

# Add new context (appends to prompt)
system_prompt(action="add_context", context="New learning: Always use FastAPI for APIs")

# Update entire prompt
system_prompt(action="update", prompt="You are a specialized DevOps agent...")

# Sync to GitHub (persist across deployments)
system_prompt(
    action="update",
    prompt="Updated system prompt with new learnings...",
    repository="cagataycali/devduck"
)

# Reset to default
system_prompt(action="reset")

Pattern: Learn → Add context → Sync to GitHub → Persist forever

Knowledge Base (Auto-RAG)

Automatic memory across sessions:

export DEVDUCK_KNOWLEDGE_BASE_ID=your_kb_id
devduck

How it works:

1. Before each query: Retrieves relevant context from KB
2. After each response: Stores conversation for future reference
3. No manual tool calls needed - fully automatic

Manual storage:

store_in_kb(
    content="Important information to remember...",
    title="Project Context",
    knowledge_base_id="optional-kb-id"
)

Sub-Agent Creation

Delegate tasks to specialized agents via GitHub Actions:

# Create sub-agent with specific model and tools
create_subagent(
    repository="owner/repo",
    workflow_id="agent.yml",
    task="Analyze this dataset and provide insights",
    model="us.anthropic.claude-sonnet-4-20250514-v1:0",
    provider="bedrock",
    max_tokens=60000,
    tools="file_read,python_repl,calculator,http_request"
)

# Custom system prompt for specialized behavior
create_subagent(
    repository="owner/repo",
    workflow_id="agent.yml",
    task="Review code and suggest improvements",
    tools="file_read,editor,shell",
    system_prompt="You are a senior code reviewer focused on best practices"
)

# Check sub-agent status
create_subagent(action="status", repository="owner/repo", workflow_id="agent.yml", run_id="12345")

# List recent runs
create_subagent(action="list", repository="owner/repo", workflow_id="agent.yml")

Hot-Reload Example

# ./tools/weather.py
from strands import tool
import requests

@tool
def weather(city: str) -> str:
    """Get weather for a city."""
    r = requests.get(f"https://wttr.in/{city}?format=%C+%t")
    return r.text

Save → use instantly:

🦆 weather(city="Tokyo")
# Clear sky +15°C

No restart. No configuration. Just works.

Access Methods

CLI Commands

# Interactive REPL
devduck

# One-shot query
devduck "your query here"

# MCP stdio mode (for Claude Desktop integration)
devduck --mcp

# Session recording
devduck --record                    # Start with recording enabled
devduck --record "do something"     # Record a one-shot query

# Resume from recorded session
devduck --resume session.zip        # Resume from latest snapshot
devduck --resume session.zip "continue"  # Resume and run new query
devduck --resume session.zip --snapshot 2 "continue"  # Resume from specific snapshot

REPL Commands

Custom ports:

export DEVDUCK_TCP_PORT=9000 DEVDUCK_WS_PORT=8001 DEVDUCK_MCP_PORT=8002
devduck

Disable servers:

export DEVDUCK_ENABLE_TCP=false DEVDUCK_ENABLE_MCP=false
devduck

Configuration

Troubleshooting

Ollama model not found:

# DevDuck auto-pulls models, but if it fails:
ollama pull qwen3:1.7b

Port already in use:

# Change ports
export DEVDUCK_TCP_PORT=9000
export DEVDUCK_WS_PORT=8001
devduck

Hot-reload not working:

# Ensure tools directory exists
mkdir -p ./tools

# Check file watcher logs
devduck
🦆 view_logs(action="search", pattern="watcher")

Memory/performance issues:

# Use lighter model
export STRANDS_MODEL_ID="qwen3:0.5b"

# Reduce context
export DEVDUCK_LOG_LINE_COUNT=20
export DEVDUCK_LAST_MESSAGE_COUNT=50

Speech dependencies not found:

# Install speech extras
pip install "devduck[speech]"

# Or with pipx
pipx install "devduck[speech]"

Ambient overlay not starting:

# Make sure tkinter is installed
python3 -c "import tkinter"

# Install tkinter if missing
brew install python-tk@3.13  # macOS
sudo apt-get install python3-tk  # Ubuntu/Debian
sudo dnf install python3-tkinter  # Fedora

Tray app not starting (macOS):

# Install rumps
pip install rumps

# Or reinstall devduck
pip install -e .

View logs: devduck → 🦆 view_logs()

GitHub Actions

Run DevDuck in CI/CD pipelines:

name: AI Code Assistant
on: 
  issues:
    types: [opened, edited]
  pull_request:
    types: [opened, edited, synchronize]

jobs:
  devduck:
    runs-on: ubuntu-latest
    permissions:
      contents: read
      issues: write
      pull-requests: write
    steps:
      - uses: cagataycali/devduck@main
        with:
          task: "Analyze and help with this issue or PR"
          provider: "github"
          model: "gpt-4o"
          tools: "shell,file_read,file_write,use_github,calculator"
        env:
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}

Sub-agent workflows:

devduck("Create a sub-agent to analyze test coverage")

Resources

• Strands SDK: github.com/strands-agents/sdk-python
• Documentation: strandsagents.com
• Web UI: cagataycali.github.io/devduck

Citation

@software{devduck2025,
  author = {Cagatay Cali},
  title = {DevDuck: Self-Modifying AI Agent with Hot-Reload and Multi-Protocol Servers},
  year = {2025},
  url = {https://github.com/cagataycali/devduck}
}

Apache 2.0 | Built with Strands Agents | @cagataycali