Waldiez builds open-source frameworks for designing, orchestrating, and deploying AI agents — from visual workflow design to resilient actors running on edge devices. Our tools span the full spectrum of agentic AI: drag-and-drop orchestration for general-purpose multi-agent workflows, and an actor-model runtime for agents that live, adapt, and coordinate in the real world, 24/7.

Two complementary frameworks, one mission — make it easy to build AI agents that work together, whether they're reasoning in a notebook or detecting people on a Raspberry Pi.

An innovative platform that enables seamless collaboration among AG2 AI agents through an intuitive drag-and-drop interface. Design, orchestrate, and execute complex workflows by integrating various AI models and tools effortlessly. Learn more

Key features

• 🤖 Runs over AG2: Supports AG2 communication patterns for building agentic workflows.
• 🔬 JupyterLab Extension: Create, convert, and execute workflows directly within notebooks. Read more
• 🖥️ Visual Studio Code Extension: Design and manage Waldiez flows inside VS Code. Download
• 🎬 Waldiez Studio: A FastAPI-based web app for converting and executing Waldiez flows.
• 🚀 Rapid Prototyping: Export and import models, tools, agents, and workflows to accelerate iteration.
• 🧠 Multi-LLM Support: OpenAI, Anthropic, Google, NVIDIA NIM, Ollama, and several others.
• 🐳 Docker Support: Pre-configured images for the core package, JupyterLab extension, and Waldiez Studio. Check it out

An actor-model multi-agent framework built from scratch in Python for agents that run 24/7 on the edge. Describe what you want in natural language, and the LLM writes the code, wraps it in a <spawn> block, and a new live agent appears on the fly — no restarts, no handwritten infrastructure. Designed as part of the SYNAPSE project funded by dAIEDGE.

Key features

• 🎭 Actor Model at the Core: Agents are isolated actors spawned at runtime from natural-language descriptions. No prescribed paradigm — mix reinforcement learning, rule-based, Bayesian inference, active inference, LLM-driven, or plain deterministic logic in the same system.
• 📡 MQTT-Native Coordination: Agents communicate via MQTT — the real nervous system of IoT — alongside direct actor messaging.
• 🏠 Edge-First Design: Runs on modest hardware, fully offline. Spawn agents on a new node (e.g., Raspberry Pi) over SSH with a single command.
• 🔁 Crash-Resilient: Agents survive crashes and restore state automatically, with rolling conversation summarization so context is never lost.
• 🏗️ PlannerAgent: Decomposes complex tasks into dependency graphs and fans them out in parallel across actors.
• 🌐 Reactive Pipelines: "If a person is detected, turn on the lights" — pipelines are built and wired automatically, integrating with tools like Ultralytics YOLO and Home Assistant.
• 🧠 Multi-LLM Support with Cost Tracking: Works across Anthropic Claude, OpenAI GPT, Google Gemini, Ollama, and NVIDIA NIM, with per-agent LLM cost tracking built in.
• 💬 Multi-Platform Interfaces: Discord, Telegram, WhatsApp, REST, or CLI — whatever fits your workflow.

• 🏛️ waldiez — The central repository, including the React application for generating Waldiez flows, tools to convert flows into Python scripts or Jupyter notebooks, and the documentation.
• 🔬 jupyter — JupyterLab extension for working with workflows inside notebooks.
• 🖥️ vscode — VS Code extension for designing and managing Waldiez flows.
• 🎬 studio — FastAPI-based web interface for converting and running flows.
• 🏃 runner — Queues and runs Waldiez flows in isolated environments, streaming logs, input, and output via Redis.

• 🎭 wactorz — The actor runtime, MQTT coordination layer, PlannerAgent, spawn protocol, state persistence, and integrations with sensors, actuators, and home automation platforms.

Together, these projects cover both ends of the agentic spectrum: collaborative workflow design for general-purpose AI agents, and resilient, edge-deployable actors for the physical world.