I Built an AI Agent Workspace That Runs 400+ Conversations at Once — No JS Frameworks, No Node Modules, Just Python

Cross-post note: This project is open source and free. If you like ambitious local-first tools, the repo links are at the bottom.

The problem

Most AI chat apps share one limitation: they're built around a single conversation. If you want to run 10 tasks in parallel, you open 10 tabs, babysit each one, and copy-paste context between them.

I wanted something different: a workspace where the LLM is not a chat partner but a team of workers on your computer. So I built Zhufeng Agent Unlimited (Infinity) — a local-first AI agent workspace for Windows.

What it is

One infinite canvas + one AI that actually gets things done. It doesn't just chat — it plans, acts, verifies, and repairs itself when it breaks things.

Highlights:

• 400+ simultaneous AI conversations. Each conversation is an independent agent with its own model, tools, and context. Agents can delegate, monitor each other, and collect results.
• 83 built-in tools across 4 categories (daily tasks, coding, writing, canvas) — read/write files, run code, fetch the web, Git commits, scheduled tasks, text-to-image, video generation.
• Infinite canvas multi-agent graphs: create nodes, wire dependencies, run them, fix broken nodes, lock finished ones. Think of it as a visual pipeline where each node is a full agent.
• Self-healing: when an agent hits an error, it re-reads the code, diagnoses the root cause, patches, and verifies. I watched it fix a real bug in its own frontend during testing.
• Long plans: tasks longer than one chat session get persisted as shareable plans, so a new conversation can pick up where the last one stopped.
• Zero-dependency startup: pure Python standard library server. Double-click to run, no pip install, no Node, no Docker.

Why pure Python stdlib?

Because setup friction kills adoption. If a user has to install Node, run three commands, and configure an API key before seeing anything, most of them are gone. With Infinity you download a folder, double-click, and the browser opens.

The trade-off: no WebSocket frameworks, no async megastack. Instead there's a hand-rolled HTTP layer with keep-alive, a small routing system, and a frontend with zero build step — plain JS files, versioned with cache-busting query params. It's ~6MB of JavaScript with no bundler. Ugly? Maybe. It works, and anyone can read every line.

The multi-agent canvas

The most interesting part is the node graph. You describe a goal, and the agent decomposes it into nodes on an infinite canvas — forward decomposition or backward reasoning from the goal. Each node:

1. Holds its own conversation and context
2. Can read results from upstream nodes
3. Writes conclusions into a global context pool other nodes can read
4. Can be locked ("fixed") when done, so later runs don't redo finished work

It's a different mental model from linear chat: you stop prompting and start orchestrating.

Who is it for?

People who use AI for real work: batch file processing, codebase-wide refactors, content pipelines, research with many parallel threads. If your workflow involves more than 2 chat tabs, this is for you.

Try it

MIT licensed, free forever.

Tags: ai, agents, python, opensource, llm, productivity, showdev, devtools