Your cloud and subscriptions

Your agents need computers to work on. Fly.io runs those computers in your own cloud account. When a workplace is set up, an admin connects that account once, on the Connect your Fly cloud screen, by pasting a single Org deploy token from Fly’s dashboard. Nothing else is asked for: the organization is read from the token, and the token is checked with Fly before it is saved. See Connect your cloud.

Until an admin has done that, a member sees the workplace’s Cloud account card read No cloud account connected yet — an admin connects one here.

The token is stored encrypted and is never handed to an agent. An agent’s machine is created and started with it, but the agent itself never sees it, and it never appears in a chat. The form says so under its field: Checked with Fly before saving. Stored encrypted and never shared in agent chats.

The Workspace recommends a separate Fly organization for your agents, and the screen explains why: a separate organization keeps our token on your agents’ ground only, and everything else you run on Fly stays untouched. The platform only ever acts on machines and sites its own records claim; an app you put in that organization yourself is never touched. It also keeps the Fly invoice readable, because everything on it is agents.

The payment method stays with Fly, on your own bill. We never see it and never charge you. To cut the platform off, delete the token in Fly’s dashboard; it stops working at once. To rotate it, paste a new one, as the next sections describe.

Three kinds of thing live in your Fly account, and all three are yours:

• One machine per agent, including each person’s Workspace Agent. The machine keeps its files between turns, sleeps after about three quiet minutes, and wakes in seconds when there is work. See The machine and Sleep and wake.
• The sites your agents deploy, each on its own small machine. See Production sites.
• A storage bucket for files: the photos and files sent in chats, put on notes, proposals and knowledge pages, and everything in Drive. It is created the first time the workplace stores a file. See Where files live.

Repositories are not in your Fly account. Code lives on the platform’s own git host, with its backups. See Where code lives. For the full split of what sits where, read the Privacy Policy.

The workplace’s screen has a Cloud account card that every member can read: the Fly organization’s name, a line such as Your own cloud · private tunnel up, and how many machines are active and how many are sleeping right now. Admins also see See cloud details, which opens the spend view.

Cloud is the admins’ view of what the workplace runs and roughly what it has cost this month. Its first line is the estimate so far this month, then how many machines are running and how many are asleep. It is the month to date, not a forecast.

Under a card for the account itself, Your team’s cloud · everything below runs here, the Machines list shows one row per machine: the agent’s face and name, how long the machine has been awake this month, when it has been up since, what it costs for each awake hour, and its cost so far. A machine that no longer belongs to any agent reads Leftover machine and offers Remove. The foot of the screen says when the figures were checked; they are read again on every visit.

Every figure is an estimate from Fly’s published prices. Fly’s own dashboard is the authority on what you will actually be billed, and Fly’s pricing page is the source for what a machine costs. The storage bucket is not counted here.

The biggest lever is one you do not have to pull. A machine costs processor and memory only while it is awake; asleep it costs almost nothing beyond its storage, and agents sleep on their own when idle. There is no approval step on spend. What stands in its place is this visibility, an account-wide ceiling on the number of sites, the rule that only an admin’s agent creates new infrastructure, and sleep.

Paste a new token from the workplace’s Cloud account card or with Replace this cloud on the Cloud screen. Both open the same form as the first time.

A token for the same Fly organization rotates quietly. Nothing moves, nothing restarts, and the screen confirms that the organization is still your cloud. This is the routine case: a token you revoked and reissued, or one you simply prefer to renew.

A token for a different organization is a move, and the Workspace asks first. It names both organizations and says what a move means: every agent, machine and site made from now on is created in the new one, on that account’s bill. You choose Keep the old organization or Move to the new one. Everything already in the old account stays there, keeps billing on that account, and is no longer reachable from the platform. A move is for a workplace that is starting over on a new account, not for a token refresh.

A seat is a subscription you have signed in: a Claude Pro or Max plan, which runs agents on Claude Code, or a ChatGPT Plus or Pro plan, which runs them on Codex. You sign in on the provider’s own page, and your password never touches us. See Sign in a subscription.

A seat is yours alone. It runs only your own agents, and no teammate, admin included, can see its credential or its usage. You can assign the same seat to several of your workplaces at the same time; each workplace’s admins still see nothing of it. Several agents share one seat: every Claude agent of yours in a workplace runs on the same chosen Claude seat.

Your seats appear as a strip on Home, one tile per subscription with its usage, and Add sub opens Your seats. That page lists the seats assigned to this workplace, with Sign in a subscription to add one, and a section Your other sign-ins for seats you use elsewhere, each with Use here. Each seat wears a status: Ready, Resting, Needs sign-in, Off or Subscription needs attention.

One seat’s screen shows its Status, Plan, This workplace (whether the workplace uses it first or keeps it as a fallback), and when it was Added. With more than one seat of a provider, Make primary here puts one first, and Earlier fallback and Later fallback order the rest; the order is per workplace. When the primary seat is paused, your agents fall back to the next ready seat of the same provider.

Two ways to take a seat away, and they differ. Unassign from this workplace leaves it signed in and working in every other workplace you assigned it to. Remove subscription everywhere signs it out of all of them, and you would sign in again to get it back.

When Claude pauses a subscription, usually a billing matter on their side, the seat’s screen says Claude paused this subscription and offers Open Claude billing and Check again. A sign-in that has lapsed offers Retry sign in. On Home, an agent waiting on its seat reads Paused — Claude subscription needs attention or the ChatGPT equivalent, and an agent that has used up a window reads Session limit until it can resume.

A subscription has windows: an amount you can use in a short period, and an amount for the week. A seat’s screen shows each window under Usage, or Last known usage when the newest reading is older. A Claude seat shows 5 hours, 7 days, and a weekly window per model, and it may show Extra usage, which is a spend cap rather than a limit. Before the first reading it says No reading yet.

Each window reads how much is used as a percentage, with a meter and Resets followed by when. On the meter, a tick marks where an even pace would be by now, and a line under it says how you are doing against it: ahead of an even pace, with a forecast of when the window would run out if it keeps up; under an even pace; or Right on an even pace. Within three points counts as even, and early in a window it simply says Period just started.

The pace line is the one to watch. A high percentage late in the window is fine; a high percentage early means the agents will hit the limit before it resets, and the forecast says when. When that happens, the agent pauses at Session limit, the seat’s screen shows Back with the time its limit resets, and the agent picks up again on its own once it does.

Every agent’s machine has a size: how many processor cores it has, whether they are shared or dedicated, and how much memory. The standard for a new agent is two dedicated cores and 4 GB of memory, and that is a good size for almost all work. The ladder runs from one shared core with 2 GB up to eight dedicated cores with 16 GB:

• Shared cores: one, two or four cores with 2, 4 or 8 GB. Cheapest, but a shared core is throttled hard under sustained load, and the picker says so beside them: the same job can take many times longer there.
• Dedicated cores: one, two, four or eight cores with 2, 4, 8 or 16 GB. The standard is in this list.

Memory is a wall, not a slope. There is no swap, so work that outgrows the memory is stopped rather than slowed, and the fix is one step up the ladder. The picker shows an approximate cost per awake hour beside each size; Fly’s dashboard is the authority on the real figure.

Two people decide sizes. An admin sets what new agents are built on, from Cloud details, in the New agents row, which opens Machine for new agents. That changes nothing already running. An agent’s owner, whatever their role, resizes that one agent from the agent’s own screen, in its Machine row. See The machine.

A resize is refused while the agent is mid-turn. An awake machine restarts, and anything it is running stops for about a minute; an asleep machine simply wakes up on the new size and nobody is interrupted. Files are kept either way. The platform never resizes a machine on its own, and an agent cannot resize its own.

Agentic Cloud Computer is free to use. There is no paid tier and no checkout, and no billing system behind the Workspace at all: it does not take payment details and cannot charge you.

What you pay, you pay to your providers directly, for what you choose to run with them:

• AI: your Claude or ChatGPT subscription, billed by Anthropic or OpenAI as it always was.
• Cloud: your Fly account, for the machines, sites and storage bucket that run there, billed by Fly on the card you gave them.
• Git and media: whatever you connect from outside, on that provider’s own terms.

None of those is a charge from the platform, and the platform resells none of them. The Cloud screen exists so that Fly’s bill never surprises an admin.