zzldragon/metaagent

MetaAgent is a desktop application, built with PySide6, that gives you a visual canvas for designing AI agent workflows. Instead of writing code directly, you build a graph made of nodes and connecting lines, dragging in pieces like agents, language models, tools, knowledge bases for retrieval, memory, reusable skills, and checkpoints where a human needs to step in and approve something. You can wire these pieces together following known patterns, such as a simple chain, a router that picks a path, a supervisor that oversees other agents, an orchestrator, a fan-out and join structure, a map-reduce style split, or a voting setup. Once the graph is built, clicking Generate turns it into a standalone Python program that runs on its own, without depending on frameworks like LangChain. The tool does not lock you into one AI provider. It works with DeepSeek, OpenAI, Anthropic, and any other service that speaks the OpenAI-compatible API format, and it can automatically fall back to another provider if one fails. It supports typed shared state with custom nested data types, can search the web using multiple search engines, allows agents to run on a schedule in the background, and can be deployed as a multi-user web server. Two built-in AI assistants help while you work: one writes Python tools for you on request, and another, called the Designer Agent, can take a plain-language description of what you want and turn it into a working graph drawn directly on the canvas. When you run a workflow, you can watch it play back step by step on the canvas to debug it. Finished graphs are saved as portable files with the .mta extension, and any agent, or the designer tool itself, can be compiled into a single standalone Windows executable. The interface is available in both English and Simplified Chinese.