A pool of resources.
Split it your way.

Your plan gives you a total budget of vCPU, running memory, and disk, with separate per-computer limits. For example, Solo can hold four computers with 1 vCPU, 2 GB RAM, and 20 GB disk each. Its 80 GB disk pool cannot all go into one computer: the per-computer disk limit is 40 GB.

Stop a computer before resizing its CPU, RAM, or disk. CPU and RAM can be reduced; disks can only grow.

Stopping or suspending a computer frees the memory it was using, and frees its vCPU and its slot for another computer to start: the memory pool, the vCPU pool, and the computer limit are all charged against what is running. Creating one is different — the computer limit, the vCPU pool, and the storage pool each count every computer you own, running or not — and a resize counts vCPU and storage the same way. Deleting a computer is what gives back its slot, vCPU, and disk.

Yes, from your billing page. Upgrades and downgrades apply immediately with prorated charges or credits for unused time. If payment fails, your plan stays unchanged. After a downgrade, the smaller plan’s limits apply to subsequent requests; you may need to reduce your resource use before creating, starting, or resizing computers.

When funded mode is enabled for your account, Solo includes $5, Studio $10, and Swarm $50 of Mandala AI credits per monthly billing period. These credits pay for Mandala's built-in computer agent and expire at the end of that period. Funded runs require a plan, a supported billing country in Stripe, and available credits; a subscription alone does not enable funded mode.

Funded usage costs 1.3× Anthropic's API list price for the model used. Sonnet 5.5 is the default, with Opus 5.5 selectable. Anthropic's fallback routing can use other models, charged at 1.3× their own API list prices. Each model call needs at least $0.25 available to start. A run's cost limit controls whether another call starts; the final call can exceed that limit, and all completed billable usage is charged. If usage exceeds your remaining credits, the shortfall carries forward even after those credits expire. Future grants and top-ups pay it first, leaving less to spend.

When enabled, account owners can buy $10, $25, or $50 top-ups from Billing. Top-ups expire 12 months from purchase. If your plan ends, existing credits freeze and their original expiry continues. A purchase paid after plan removal is retained and issued on resubscription only if it has not expired. Credits have no cash value, are nontransferable, and are non-refundable except where required by law. See the terms.

You can still bring your own Anthropic API key: send X-Model-Key with your Mandala API key. Anthropic bills that usage directly and no Mandala AI credits are spent. The OpenAI-shaped endpoint and agents inside a computer, including Claude Code, Hermes and OpenClaw, still require your own model provider. Funded prompts, screenshots and tool results go to Anthropic under Mandala's account; see subprocessors.

Every plan can serve a port inside a computer on its own HTTPS address under mandala.dev, open to anyone with the link or private to people signed in to your account: up to 2 per computer on Solo, 5 on Studio and 10 on Swarm.

The built-in computer agent keeps each conversation as a thread, so a later prompt carries on where the last one stopped, and it can work in the background: start a run, walk away, and read back its steps, cost and ending later. A run stops after 60 minutes.

Files of up to 2 GiB go into a computer as a resumable upload, sent in chunks and picked up again after a dropped connection.

Secrets are stored once and bound to computers. The API never hands a stored value back, but a bound secret is delivered into the computer, where code running there can read it. The credential gateway keeps a secret out of the computer instead: code inside calls an upstream API through the gateway, which adds the real credential on the way out, only for the methods and paths you allowed. The provider's responses come back unchanged, so use one that does not echo credentials, with a narrowly scoped key.

Linux desktops, from templates we maintain: a Debian desktop with Xfce, Omarchy (Arch Linux with Hyprland), and the Debian desktop with Claude Code, Hermes Agent, OpenClaw, a set of ten coding-agent tools, or Anthropic's Claude Managed Agents worker already installed. If your workload needs something else, tell us about it.

Enable a daily snapshot schedule for each computer in the dashboard or API; schedules are not enabled by default. Scheduled history is bounded by your plan’s retention window. Scheduled snapshots count toward your snapshot storage but are never stopped by it, so a long scheduled history can leave less room for manual snapshots. You can also take snapshots on demand from the API or dashboard and restore a saved snapshot when needed.

Each computer is a genuine virtual machine behind hardware virtualization — its own kernel, its own network policy, its own credentials. Machines can't see each other, other customers, or our infrastructure.

From the billing page, in one click. Your plan runs to the end of the paid month; no phone calls, no retention flows.