// ZDZAgentsOS · ZDZCloud's flagship AI product

AI agents that do, not just talk.

An operating system for fleets of AI agents that execute real work on real machines — driving VS Code, git and the shell — with a person approving every action that matters, right from their phone.

.NET 10 backend · Semantic Kernel 26 LLM providers Human-in-the-loop via WhatsApp
// The problem

Chatbots talk. The work still has to get done.

The first wave of generative AI delivered assistants that answer, suggest and summarize. Useful — but for real work, someone still has to open the editor, run the command, make the commit. Conversation doesn't turn into delivery on its own.

// AI that only talks

Assistants that stop at text

  • Suggests code, but you copy, paste and tweak it by hand.
  • Lives in an isolated tab, far from your real tools.
  • Never touches git, the shell or your environment.
  • No approval trail: either you trust it blindly, or you don't use it.
// AI that does

Agents that execute and deliver

  • Read Work Items and follow your organization's standards.
  • Drive the tools already installed on the machine (VS Code, git, shell).
  • Write code + tests, commit/push and open Pull Requests.
  • Every action that matters goes through human approval — with an audit trail.
// How it works

From a natural-language goal to an open PR.

Three pieces keep the loop turning: a visual flow designer, a cowork agent that runs on the team's own machine, and an orchestrator that turns a goal into work routed across the fleet.

01

Design the flow, visually

An n8n-style flow designer: start, action, prompt, condition, parallel, join and loop nodes. The same flow contract holds, with zero drift, across web, desktop, mobile and tablet.

start · action · prompt · condition · parallel · join · loop
02

A cowork agent on the machine

An Electron desktop app becomes a local work agent: it drives the tools already installed and produces real deliverables. It connects outbound-only, with no open port, with a token encrypted in the system keystore — and runs headless on servers.

outbound-only · no open port · token in the keystore
03

The fleet organizes itself

The orchestrator breaks a natural-language goal into a task graph and routes each task to the best-suited agent and machine — with automatic cost estimation, budgeting, retries and reconciliation.

goal → task graph → right agent
// AI + People

Autonomy with a person in command.

ZDZCloud's method is copilot first, tested autonomy later. Every action that matters waits for a human — without ever blocking the flow. See the trust ladder →

📱

Action Inbox on your phone

Approve, reject or pick a branch of the flow right from your phone. With auditing and timeouts, so the work never sits stuck waiting.

💬

In and out via WhatsApp

A message triggers a flow; the result comes back as a push notification or audio on your device. Work starts and ends wherever you are.

🛡️

Spending under control

Capability and budget gates before any LLM call. The fleet only spends what was cleared — and every step is logged.

// Capabilities

Real AI at the core — not a slide.

Real multi-agent orchestration on a .NET 10 foundation, with total AI-vendor freedom: cloud or 100% local, on-premise, when the data demands it.

🧠

Semantic Kernel + Magentic

A .NET 10 backend on Microsoft Semantic Kernel, including the Magentic multi-agent pattern (Magentic-One) — real orchestration of an agent fleet.

26

LLM providers

Cloud or 100% local/on-premise, with no AI-vendor lock-in. You choose where the model runs — including inside your own network.

6

IDE/CLI integrations

Agents drive the development tools your team already uses, instead of demanding yet another isolated window.

🔌

Native MCP

Any Model Context Protocol server instantly becomes an agent tool. Connect what you already have.

🔒

Enterprise security

A desktop agent with an outbound-only connection, no open port, a token encrypted in the keystore, shell/FS sandboxing and per-organization scoping.

🏷️

Multi-tenant and white-label

A multi-tenant foundation, ready for you to operate under your own brand — or resell the platform as your own product.

// Proof · dogfooding at scale in production

Not theory. A fleet of agents already builds our own products.

Pet Planner, a ZDZCloud multi-platform SaaS (web PWA + native app), is built by an AI-agent software factory — in a spec-driven pipeline with human approval between phases. And the AI economics are measured, not estimated.

10
AI agents across 7 phases, with human approval between them
385
LLM calls in an 8-story sprint
3.28M
input tokens in the same sprint
71%
cache-hit rate — AI cost measured, not estimated
// Frequently asked questions

What people usually ask

Is this just another chatbot?
No. The agents drive the tools already installed on the machine — VS Code, git, shell — and produce real deliverables: code, commits and Pull Requests. Conversation becomes executed work, not more text in a tab.
Do I have to send my data to someone else's cloud?
Not necessarily. 26 LLM providers are supported — you can run in the cloud or 100% local/on-premise, inside your own network. No AI-vendor lock-in, and the option to keep data exactly where it needs to stay.
How does human approval work?
Every action that matters lands in an Action Inbox: you approve, reject or pick a branch of the flow from your phone — including via WhatsApp. Every decision is audited and has a timeout, so the flow never gets stuck waiting.
Is the agent on my machine safe?
It was designed on the enterprise-agent model: outbound-only connection, no open port, a token encrypted in the operating system's keystore, shell and file sandboxing, and per-organization scoping. It also runs in headless mode on servers.
Do you actually use this, or just sell it?
We use it. Pet Planner, a ZDZCloud multi-platform product, is built by a fleet of AI agents in a 10-agent, 7-phase pipeline with human approval — and we measure AI cost in tokens per story. It's dogfooding at industrial scale.
// Let's talk

Watch a goal become a Pull Request.

Book a demo and a diagnosis of your workflow. We'll show you, live, the agent fleet executing — with you approving every step from your phone.