# Run agent code on a GPU, Windows or desktop machine

> Run agent code on a GPU, Windows or desktop machine: 0 companies fully do it, 10 with a limit, 24 carry the tag. Every listing in Code execution sandboxes with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/code-sandboxes.gpu-or-special-machine/
- Markdown: https://www.anchorterminal.com/niches/code-sandboxes.gpu-or-special-machine/index.md (~1,550 tokens)
- Slim: https://www.anchorterminal.com/niches/code-sandboxes.gpu-or-special-machine/index.min.md (~330 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/code-sandboxes.gpu-or-special-machine/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

No company's docs show a full fit yet. The listings below do the job with a limit, or not at all.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 0 | 10 | 24 | 6 | 3 |

## What counts as done

A sandbox of the requested class starts, the code sees the GPU or desktop (nvidia-smi or a screenshot succeeds), a mounted volume or snapshot of at least 10 GB restores without error, and the hourly price is published.

Measured, for our own tests: the sandbox starts within 60 seconds.

How agents ask for it: run this training script on a GPU; test the Windows build; emulate the Android app off my laptop; run the iOS simulator somewhere else

## Does the job, with a limit

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

- **Agent 37 Cloud (Agent 37 Inc.)**: The agent37-android template boots an Android 14 emulator on a VM instance with a noVNC screen, and desktop templates exist, with prices published. No GPU, Windows or macOS class.
- **Celesto (formerly SmolVM) (Celesto AI, Inc.)**: Windows 11 from your own ISO image and a macOS desktop preview boot locally with a display, and Linux desktops with Chromium. No GPU passthrough and no hosted hourly price for them.
- **Daytona (Daytona Platforms Inc.)**: NVIDIA and AMD GPU sandboxes with published hourly prices and S3-backed volumes, Windows (Tier 3, contact support) and macOS classes. GPU sandboxes are ephemeral and spot needs Tier 2.
- **E2B (FoundryLabs, Inc.)**: Desktop sandboxes with VNC and an Android 16 template exist with published compute rates; no GPU, Windows or macOS class.
- **Freestyle (Freestyle Cloud Inc.)**: Nested KVM runs an Android emulator inside a Linux VM per the guides; no GPU, Windows or macOS class, and no display price.
- **InstaVM (Checksum Labs, Inc.)**: A Linux desktop with noVNC and a headless browser are offered at the published compute rate; no GPU, Windows or macOS.
- **Maritime (Maritime AI, Inc.)**: Linux desktops at $0.12 an hour, Windows at $0.18 (Growth plan or above) and Android phones at $0.15 with screenshots over MCP, sleeping with disk at $0.08 a GB-month. No GPU class.
- **Modal Sandboxes (Modal Labs, Inc.)**: GPU sandboxes at Modal's per-second GPU rates, but only with the gVisor runtime and never with memory snapshots; issue 4139 hit a 256 GiB changed-data limit on filesystem snapshots.
- **Offscreen (Mark Phelps (individual))**: A fresh macOS VM with Xcode runs XCUITest, the iOS Simulator and built apps and saves screenshots as artifacts, free. A simulator boot takes minutes, only two VMs run, and there is no GPU or Windows.
- **Sandbox0 (Lougao, LLC)**: A chrome template gives a shared Chrome desktop; no GPU, Windows or macOS class.

## Couldn't check

The docs were closed to us, or silent. These carry no mark either way.

- **Runloop Devboxes (Runloop AI, Inc.)**: The docs index lists an optional GPU on devboxes, but no GPU page was read and the service is winding down.

- not this job: Amazon Bedrock AgentCore Code Interpreter (Amazon Web Services, Inc.): No GPU, Windows or desktop class; every session is a Linux microVM capped at 2 vCPU.
- not this job: Blaxel Sandboxes (Blaxel, Inc.): No GPU, Windows or desktop class; an Android-compatible kernel option runs Android in a container, not a device with a screen.
- not this job: Cloudflare Sandbox SDK (Cloudflare, Inc.): Linux containers and Dynamic Workers only; no GPU, Windows or desktop class.
- not this job: DeltaBox Lite (DeltaBox (delta-box)): Runs on the host it is installed on; no GPU, Windows guest or display class.
- not this job: Deno Sandbox (Deno Land Inc.): Linux microVMs of 768 to 4096 MB only; no GPU, Windows or desktop class.
- not this job: Microsoft Execution Containers (Microsoft Corporation): Contains code on the host it runs on; no GPU, remote Windows guest or display class.
- not this job: Morph Cloud (Morph Labs, Inc.): No GPU, Windows or desktop class in the docs text.
- not this job: SPACE (Perplexity AI): Firecracker Linux microVMs only; no GPU, Windows or desktop execution environment.
- not this job: Sprites (Fly.io, Inc.): Linux microVMs with 8 vCPUs and 100 GB; no GPU, Windows or desktop class.
- not this job: Together Code Sandbox (Together Computer, Inc.): No GPU, Windows or desktop class in the sandbox docs.
- not this job: Vercel Sandbox (Vercel Inc.): Linux microVMs up to 32 vCPUs; no GPU, Windows or desktop class.

## What stops agents finishing

- GPU sandboxes only with the gVisor runtime (Modal) or ephemeral and gated to Tier 2 (Daytona spot GPUs)
- filesystem snapshots fail above a 256 GiB changed-data limit (Modal issue 4139)
- Windows and macOS classes are early access

## Around it

Before: open-weights.pick-model, code-sandboxes.coding-agent-workspace. After: inference.serve-model, code-sandboxes.dev-server-preview.

## Where this comes from

Weight 6 of 100 in Code execution sandboxes: 3 of 53 mentions (an A100 Modal exec issue, a Modal snapshot size limit issue, a laptop user offloading Android emulation). Only Modal and Daytona list GPUs, so the count is small by construction.

As data: https://www.anchorterminal.com/api/v1/use-cases/code-sandboxes.gpu-or-special-machine.json, or the MCP tool get_use_case with id code-sandboxes.gpu-or-special-machine. Draft, read 2026-11-12 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
