Skip to main content

For agents

CLI for any AI workload in any cloud account

Guardrails let agents safely run jobs, create Kubernetes Deployments, and launch VMs on CPU and GPU infrastructure.

# Paste into Claude Code, Codex, or another coding agent
›set up https://anycloud.sh/SKILL.md
$anycloud jobModel training, evaluations, batch inference, and data processing.
$anycloud deploymentWarm model servers and inference workers that run many short Jobs.
$anycloud vmInteractive exploration, debugging, and custom research workflows.
Get Started →Read the Docs

Your agent's next H100 run, on any of them.

anycloud: CLI for any AI workload in any cloud account. Tell your coding agent set up https://anycloud.sh/SKILL.md.

Choose the account and hardware: anycloud job --gpu-type h100 --spot. Illustrative spot prices: AWS $1.69/hr, Azure $2.05/hr, GCP $2.31/hr, Lambda $1.84/hr.

Track and limit spend by agent session: anycloud budget set 4000 --per day --agent-session. Bars show spend per session, and a Slack alert in #training reads "anycloud: workloads waiting on spend cap".

Spot saves up to 90%. A chart shows a preempted run recovering and reaching 100%.

Your cloud accounts stay yours. Get Started.

Spot prices are illustrative. AWS and Azure list spot capacity at up to 90% off on-demand prices.
One CLI across the cloud accounts you already use
AWSAzureGCPLambda

Agent guardrails

Track and limit spend by agent session.

Your team sets the research direction. Claude Code, Codex, Cursor, and Aider operate between decisions, while analytics, throttles, budgets, and alerts keep each agent session visible and bounded.

Keep the loop moving

Agents can launch batches, read status and cost as JSON, retrieve results, and prepare outputs for the next decision using predictable commands.

Budgets, throttles and Slack alerts for each Claude Code and Codex session

Scope budgets, burn-rate throttles, and alerts to each detected agent session. When a cap is reached, new jobs wait while running jobs finish. Explore spend controls.

Launch, monitor, repeat—with controls

$ anycloud budget set 4000 --per day --agent-session

Budget  day  (resets in 19h)
  per agent-session  cap $4,000 each
    train-llama-sft     ██████████  96%    $3,848 / $4,000      $152 left
    lr-sweep            ████████░░  78%    $3,136 / $4,000      $864 left
    eval-harness        █░░░░░░░░░   9%      $368 / $4,000      $3,632 left

Agents in production

Agent safeguards that accelerate the research loop.

At Ångstrom AI, researchers chose the experiments and interpreted the results. Agents launched batches of jobs, monitored status, downloaded outputs, and generated plots and summaries for the next decision—all under per-session spend controls. The work spanned 100,000+ GPU hours.

“We have been using anycloud at @Angstrom_ai to make the gpus go brrrrrr!”
Laurence Midgley·@SilkyDogfish·Angstrom AI
Read the Angstrom AI case study

Execution modes

Choose the right primitive for any workload

Run AI research as finite jobs, turn results into long-lived Kubernetes Deployments, or drop down to a VM for direct compute access—all behind the same cloud-account workflow.

anycloud job

Job

Run finite work to completion.

Good for
Model training, evaluations, batch inference, and data processing.
Explore Jobs
anycloud deployment

Deployment

Keep an application running on Kubernetes for Jobs to target.

Good for
Warm model servers and inference workers that run many short Jobs.
Explore Deployments
anycloud vm

VM

Work directly with provisioned compute.

Good for
Interactive exploration, debugging, and custom research workflows.
Explore VMs

Prefer Python? Jobs and Kubernetes Deployments are also available through the Python SDK.

Your cloud accounts

Your cloud accounts stay yours.

Choose where work runs while infrastructure, credentials, data, credits, and billing stay in your accounts.

Choose the account and hardware

Pin a cloud account, region, or hardware profile—or let anycloud make a cost-aware choice from matching compute.

Bring your own container

Use the same container image across workload modes, with the environment, secrets, and runtime settings each one needs.

Keep ownership in your accounts

Your infrastructure, credentials, data, credits, and provider billing stay in the cloud accounts you control.

anycloud job --gpu-type h100 --spot

For one job, the lowest price across every region in every account wins. Spot prices are illustrative.

How it works

Any workflow, any cloud, any agent

Run it yourself or hand it to an agent; anycloud follows the same path from cloud-account selection through provisioning, operation, and cleanup.

  1. 01

    Choose a mode

    Start from the workload abstraction that fits.

  2. 02

    Select an account

    Choose credentials, region, and hardware.

  3. 03

    Provision

    anycloud starts matching compute in that cloud account.

  4. 04

    Operate

    Follow the workload through its mode-specific lifecycle.

  5. 05

    Clean up

    Release compute at completion or when you terminate the workload.

Optional cost optimization

Spot saves up to 90%.

AWS and Azure list spot capacity at up to 90% off on-demand prices. For checkpointed Jobs, that lowers the cost of long runs. If capacity is reclaimed, anycloud first tries to restart the VM in place; otherwise it cleans up the lost VM, provisions a replacement, restores /mnt/checkpoint, and restarts the container.

Preemption recovery
VM 1VM 2VM 3checkpoint⚡preemptionresume on a new VM
See how spot recovery works

Give agents reliable primitives for working in your cloud accounts

Your team sets the direction and limits. anycloud handles execution through the same CLI.

Get Started →Read the Docs