What is an AI agent cloud?

It is easy to describe app.nz as a model gateway, a coding agent, a GPU host, or an application platform. Each description is true, and each one is too small.

The product is an AI agent cloud: a development and runtime system where agents can change software, the platform can build and deploy it, applications can route across models, and specialized inference can run behind durable endpoints.

That category matters because an agent does not stop at generating text. A useful software agent needs somewhere to read and change code, run tests, create artifacts, call tools, store data, deploy results, observe the release, and come back when a monitor fires.

Three familiar categories, one larger loop

Most comparisons start in one of three categories.

Those specialists are often the right choice. app.nz is for the case where the handoffs between them have become the system you are spending time maintaining.

What the integrated loop looks like

A task can begin as a prompt in the web app, CLI, desktop app, scheduler, webhook, or MCP client. A coding agent works in an isolated environment and leaves durable logs, diffs, screenshots, and artifacts. The same account can build a private container, deploy a static or server application, attach queues or data services, configure a domain, and observe the result.

The application can then call one OpenAI-compatible gateway for text, image, video, audio, embeddings, search, and 3D routes. Exact models remain available when determinism matters; intent routes such as app/auto-code can choose a fit when provider availability, price, and workload differ.

When a workload needs its own model, Cog and Comfy deployments can expose typed prediction endpoints and scale from zero to GPU capacity. The important distinction is not that every feature shares one dashboard. It is that an agent can move through the whole loop without the developer rebuilding identity, billing, artifacts, and deployment glue at every boundary.

The primitives an agent-friendly cloud needs

• Durable work: tasks, leases, retries, event streams, diffs, and artifacts survive beyond one HTTP request.
• A boring API surface: HTTP, CLI, desktop, and MCP expose the same capabilities instead of separate product islands.
• Model choice without provider lock-in: exact models, BYOK, fallbacks, and intent routes share one observable gateway.
• Runtime ownership: sites, containers, domains, queues, logs, secrets, datasets, notebooks, and GPUs are close to the code-changing agent.
• Cost boundaries: prepaid credits, usage records, scale-to-zero inference, idle reaping, and hardware routing make autonomous work governable.
• Verification: agents can run tests, visual checks, and production probes, then leave evidence rather than a claim that the task is done.

Where app.nz is not the answer

Use a hyperscaler when you need its full enterprise service catalogue, regional footprint, or organization-wide governance. Use a frontend specialist when the polished framework workflow is the whole problem. Use a dedicated inference provider when a specific model and latency profile matter more than the rest of the application lifecycle. Use a focused coding agent when you already have a platform team that owns every downstream handoff.

The honest position is not “replace every cloud.” It is: give AI-native product teams one coherent path from agent work to a running, model-backed application.

A practical way to evaluate the difference

Take one real change—not a toy chat prompt—and follow it to production:

1. Ask an agent to implement it in a repository.
2. Require tests and a reviewable artifact or diff.
3. Build and deploy the result.
4. Route its model calls with observable provider and cost data.
5. Add any background work, inference endpoint, secret, or domain it needs.
6. Verify the public result and make the workflow repeatable.

Count the separate identities, billing systems, deployment contracts, adapters, and manual handoffs required. That is the comparison app.nz is designed to change.

Browse the platform comparisons, start with a maintained prompt, or inspect the same surface through the CLI and MCP server.