DigitalOcean MicroVMs

Isolated virtual machines for your AI agent infrastructure. Each MicroVM runs your container image with its own kernel, typically starts in under a second, pauses when idle, and resumes where it left off. You pay for compute only while it runs.

Three problems with running agent infrastructure today

Abstract Circle

Always-on machines for bursty work

Agents spend most of a session waiting on a model, a test run, or a person. A long-lived VM bills for every idle minute.

Abstract Circles

Shared kernels for untrusted code

Agents run code nobody has reviewed. Containers share the host kernel, which is a weaker boundary for code you don’t trust.

Abstract Cylinder

Building your own virtualization layer

Doing it yourself means fleet management, snapshots, pause and resume, per-tenant isolation, placement, and recovery. That is time not spent on your product.

What you get

MicroVMs give you the virtual machine and its full lifecycle, through the Control Panel, doctl, or the API. Because these environments can instantly pause when the AI is thinking or waiting, you only pay for the seconds when the system is actually working. Plus, you can bring your own custom harness and keep granular, kernel-level sandbox control without stitching hypervisors together.

Your image, its own VM

Bring your own container image and run it in its own VM, with its own kernel and a hardware-virtualized boundary.

Pause when idle, resume where you left off

Set an idle timeout. A paused MicroVM stops compute billing and typically resumes in under a second on the next request, with memory, files, and running processes preserved.

Start warm with checkpoints

Set up an environment once, checkpoint it, and start new MicroVMs from that state.

Authenticated HTTPS endpoint

Each MicroVM gets its own HTTPS endpoint, configured to require a valid DigitalOcean API token for access. Only requests with a valid DigitalOcean API token can reach it.

Private networking

Run MicroVMs inside your VPC to reach databases and internal services over private networking rather than the public internet. If you don’t have a VPC, DigitalOcean can create one for you. Add a NAT gateway to the VPC if your MicroVMs need internet access; NAT gateways are billed separately at standard rates.

Built-in recovery

If a workload crashes after it starts, the MicroVM restarts it automatically, up to the retry limit described in the Documentation, after which retries stop.

Commands and console

Run a command or open an interactive terminal inside a running MicroVM.

What you can build

Coding-agent platforms

One MicroVM per session, paused while the agent waits.

Sandboxes for untrusted code

Run AI-generated or user-submitted code in its own VM, alongside your own controls and review.

Evals and RL at scale

Start every run from the same checkpoint, with the same memory and disk.

Ephemeral CI/CD runners

Give every pipeline job a fresh VM with its own kernel. Each job starts from a fresh VM, and the VM is terminated when the job ends.

How it works

  1. 1. Bring your image

    Package your agent, toolchain, and dependencies as a container image.

  2. 2. Create a MicroVM

    From the Control Panel, doctl, or the API.

  3. 3. Send requests

    Reach your workload through its authenticated HTTPS endpoint.

  4. 4. Pause and resume automatically

    Idle MicroVMs pause and resume on the next request.

Create a MicroVM
doctl compute microvm create my-agent-session \  --oci-ref registry.digitalocean.com/your-team/agent:latest \  --region mkc1 \  --cpu 2 --memory 4096

Pricing summary

Pay for compute only while it runs

Compute is billed per second while a MicroVM is running. A paused MicroVM is billed for its storage.

Need more information?

See MicroVMs pricing
  • $0.022

    vCPU · per vCPU-hour

  • $0.0085

    Memory · per GiB-hour

  • $0.05

    Storage · per GiB-month, on what you use

  • $0.01

    Outbound data transfer · per GiB on public networking

Example: AI agent workspace. A 2 vCPU / 4 GiB MicroVM that’s active 2 hours each workday and paused the rest of the time costs about $4.18/month: $3.43 compute, $0.70 storage, $0.05 data transfer. The same MicroVM running continuously would cost $56.94/month in compute. This is an illustrative example based on the assumptions stated; actual costs vary with your usage.

See MicroVMs pricing

Part of your stack

Run MicroVMs next to the rest of DigitalOcean.

Managed Agents

MicroVMs power DigitalOcean Managed Agents. Use Managed Agents if you want the agent runtime managed for you.

Inference Engine

Keep agent compute close to the models it calls.

Container Registry

Store the images your MicroVMs run.

VPC

Connect MicroVMs privately to your databases and services.

Frequently Asked Questions (FAQ)

Start building with MicroVMs

Create your first MicroVM in minutes from the Control Panel, doctl, or the API.