Why do vertical agent products still need a real computer?

APIs cover part of a vertical workflow. The rest lives in desktop apps, portals and files, and that is where an agent needs its own computer.

OrgoUpdated November 5, 2026

Quick answer: a vertical agent product still needs a real computer because the software a vertical runs on often has no API for the step the agent has to take, and the work moves across a browser, desktop apps, files and a terminal. Orgo gives each agent, or each of your customers, its own cloud computer with a full desktop. You can bring your own agent or run one Orgo ships.

TL;DR

APIs first, computer for the rest. Use an API when it has the operation you need. Use a computer for the steps that only exist in an interface.
Vertical work crosses applications. A task often means a portal in the browser, a file on disk, a desktop app and a command, in one working environment.
Each customer needs its own state. Their logins, files, connected apps and secrets should live on a computer that is theirs alone.
People need to see it. A real screen lets your customer watch the agent, and lets your team debug what went wrong.

What is a "real computer" for an AI agent?

A real computer for an agent is an isolated virtual machine with its own disk, a full desktop, a browser and a terminal, that the agent controls through screenshots, mouse, keyboard and shell commands. It keeps its state between tasks.

A vertical agent product is an AI agent built for one industry or role, such as sales, legal research or an agency's client work, and sold to many customers. You own the workflow. The computer is the layer it works on.

On Orgo, a Linux or Windows computer is that kind of machine, self-serve and reachable over HTTP and VNC (Introduction). Macs and iPhones come from Orgo's stock while it's in stock, and an iPhone is a real phone with a screen, not a desktop with a terminal (pricing.md).

Why isn't an API enough for a vertical agent?

An API is the best interface when it exposes the operation you need. It is fast, structured and easy to test. The trouble is coverage. In many verticals the system a customer actually works in is a desktop application, a vendor portal, or a web app whose API leaves out the step the agent has to take.

When that happens, the agent has two choices: stop and hand the task back, or use the software the way a person does. Driving a computer is the second option.

Interface	Use it when	What it can't do on its own
Application API	The API has the exact operation you need	Reach steps the API leaves out
Browser automation	The whole task stays inside websites	Open native apps or work across applications
A real computer	The task needs desktop apps, files, a terminal or several apps together	Remove the need to test how reliably the agent completes it

Most good agents combine all three. What is computer use? explains the screenshot, decide, act and check loop behind the third row.

Why not just give the agent a headless browser?

A headless browser runs web pages without a screen. It is a good fit when every step happens in a website. A vertical workflow often doesn't stay there: the agent downloads a file, opens it in another program, checks a result, and records it somewhere else.

A computer keeps all of that in one place. The browser, the files and the installed apps share one filesystem, so the output of one step is already where the next step needs it. You can still run browser scripts inside the computer for the repetitive parts. The trade-offs are covered in Cloud desktops vs. headless browsers.

Why does each customer need their own computer?

In a vertical product, the agent works inside each customer's accounts and data. That state should not sit on a machine another customer's agent can reach.

Each computer is isolated. It is its own virtual machine with its own disk.
Connected apps belong to one computer. An app connected on one computer is that computer's alone, even inside the same workspace, and deleting the computer signs its apps out (Connect apps to an agent). The agent on that computer uses them through Orgo's tool server, so they are built for agents that run on the computer (Hermes, OpenClaw, Claude Code or Codex get a setup sheet). An agent loop you run elsewhere doesn't get them automatically.
Secret values are injected per computer. A template names the secrets it needs. At launch, Orgo copies the launching user's values (saved account-wide or for one workspace) onto that computer, never into the shared snapshot. A restricted API key without Secrets write launches with no secrets (Secrets).
The state stays. Stopping a computer archives its disk, so files, installed software and logins are there when it starts again (Stop computer).

How to set this up (a workspace and a scoped key per customer, Invoicing, offboarding) is in How do you give every customer of your vertical agent its own computer?.

Why does a real screen matter to your customers?

When an agent works on someone's behalf, they want to see what it did. A real computer has a real screen you can show.

In your app. Every Orgo computer streams over one WebSocket URL, so you can put the live screen in your product with any websockify-compatible VNC client (Embed computers). See How to embed a live AI agent computer in your app. Setting rfb.viewOnly in the browser is not a security control: a visitor can turn it off.
In Orgo. A workspace member with Can chat access can watch screens, and Orgo's proxy enforces view-only on the server for them. Despite the name, they can't chat with the workspace's agents. Full access adds that, along with creating, changing and deleting computers, setting up agents and connecting apps (Workspaces, members and clients).
For your team. When a run goes wrong, you can look at the same interface the agent used instead of guessing from logs.

Which computer should a vertical agent run on?

Need	What to use on Orgo
Linux desktop with a browser and terminal	Self-serve
Windows desktop apps	`os: "windows"`. Included on Scale. On Starter and Growth each one is $47/month in all, a $15 license plus a Medium's memory, vCPU and disk. Without one, the create returns `403 WINDOWS_REQUIRES_SCALE` (pricing.md)
Mac or iPhone	From Orgo's stock in the app while it's in stock, billed on its own (pricing.md)
The same setup for every customer	A template, built once into a golden snapshot. See Templates for vertical AI agents
Many customers at once	An account on a paid plan can have 50 computers, and the plan's vCPU and RAM decide how many of them it runs. Past that room, a blank Linux computer or one from a curated starter template runs by the hour from credit, or you add capacity. Past 50, only those two run, by the hour, up to 999. Computers from your own template stay within the 50 and your plan's room (Account limits). Computers in a workspace a client pays for under Invoicing don't count toward the 50

For the agent itself, you have two options. Bring your own (Claude computer use, OpenAI's CUA, Gemini or your own loop), or run Orgo's: the OpenAI-compatible endpoint runs orgo-agent when you send no model, and curated templates launch Hermes Agent, OpenClaw, or Claude Code and Codex (Templates). Computer use models compared can help you pick a model.

When does a vertical agent not need a computer?

You don't need one when every step of the workflow is covered by APIs your agent can call, or when the whole task stays inside one website and a browser tool is enough. Start there. Add a computer when the work reaches desktop software, files across apps, or steps no API covers.

Building that computer layer yourself on raw VMs is a longer list than it looks. The build list and Orgo's prices are in Orgo vs rolling your own VMs.

FAQ

Can my agent use both APIs and the computer?

Yes. The agent can call APIs directly, run commands in the computer's terminal, and use the screen only for the steps that need it.

Does Orgo include an agent?

Yes, if you want one. The chat endpoint runs Orgo's computer use agent (orgo-agent) when you send no model, and the curated templates ship Hermes Agent, OpenClaw, and Claude Code with Codex. You can also bring your own agent or model (Introduction).

Can one customer's agent see another customer's apps?

Connected apps belong to one computer, and a key scoped to a workspace reaches only that workspace. Give each customer their own workspace and computers.

Can my customer watch the agent without controlling it?

Invite them to the workspace with Can chat access, and have them watch in Orgo. There the proxy filters their input on the server. An embed that hands out the computer password has no such filter. Despite the name, Can chat doesn't let them chat with the workspace's agents. That takes Full access, which also lets them operate the computers. Someone with Full access can have an agent use its connected apps, so connect only accounts that everyone in that workspace may use. Each workspace holds 5 people besides you on Free, Starter and Growth, and 25 on Scale.

How long does a computer stay around?

On a paid plan, it runs until you stop it and exists until you delete it. A stopped computer keeps its disk. You can also create one with an end time (1 to 720 hours), after which it is deleted with everything on it (Set when a computer ends). A computer running by the hour stops by itself when your credit can't pay for it. With no plan and no credit, it is deleted 7 days later (pricing.md).