boxd vs exe.dev

Both give you a persistent Linux VM over SSH. Where exe.dev is the better choice, where boxd is, and how to pick.

Hidde KehrerHidde Kehrer6 min readComparisons
A creature measures one of two similar machines with a large pair of callipers on a workbench under a bright lamp, notebook open.
On this page · 6 min

boxd and exe.dev both give you a persistent Linux VM you reach over SSH, with HTTPS and coding agents ready to use. Choose exe.dev if you want more regions or a pool of capacity shared across many VMs. Choose boxd if you need to run it in your own cloud, want a paused machine to keep its memory, or are building on typed SDKs.

At a glance

Everything about exe.dev here comes from its public docs, checked on 10 November 2026. If something has changed, tell us at hello@boxd.sh and we'll fix it.

exe.devboxd
What you getPersistent VM over SSHPersistent KVM microVM over SSH
RegionsNine listed, with Oregon closed to new accountsTwo: Frankfurt and New York
CapacityVMs share a pool of CPU and RAM; standalone VMs get their ownEach machine has its own size
Largest machineStandalone preset up to 16 vCPU, 32 vCPU on Work plans, up to 2 GiB of memory per vCPUUp to 8 vCPU and 32 GiB as standard; up to 128 vCPU and 1.5 TB of RAM on request
DiskLocal NVMe, with changes copied off-host within the region100 GB copy-on-write disk per machine
BackupsDisk backups to your own S3 bucketDisk backups to object storage, plus up to 10 memory and disk checkpoints per machine
APISSH commands, and an HTTPS API that mirrors themCLI with --json, plus TypeScript and Python SDKs
Outbound network controlAll traffic, or integrations onlyAllowlist per machine, by host
Run it in your own cloudNo option in its docsSelf-host or BYOC

Where exe.dev is the better choice

You need a region near your users. exe.dev lists Oregon, Los Angeles, New York, Dallas, Frankfurt, Tokyo, Sydney, Singapore and London, although Oregon is closed to new accounts. boxd runs in Frankfurt and New York. If your team or your users are in Asia or Australia, exe.dev will be closer today.

You run many small VMs that are rarely busy at once. exe.dev VMs draw on a shared pool of CPU and RAM, so a quiet VM leaves its share for the others. On boxd, each machine has its own size.

You want your disk copied off the host by default. exe.dev copies disk changes to another part of the regional cluster as they happen. boxd protects data with backups to object storage and checkpoints, so restoring goes back to your last save point.

Where boxd is the better choice

You need it in your own cloud

We found no self-hosted or bring-your-own-cloud option in exe.dev's docs. boxd is a single Rust binary with no Kubernetes, external Postgres or Redis to run. A cluster deploys in 30 to 60 minutes on Google Cloud, AWS or bare metal, is free up to 16 worker vCPUs, and runs the same release as boxd cloud. The installer is open source. If you'd rather not operate it, boxd can also run inside your cloud account for you.

Your agent should come back to the same running machine

A boxd machine in standby keeps its memory, processes and open sockets, and resumes in under a millisecond. Your dev server, your test watcher and your agent's process are all still there. A hibernated machine wakes in about 85 ms. There's no session cap, so a machine can run for as long as the work takes.

exe.dev has pause and resume commands, but we couldn't find docs saying whether a paused VM keeps its memory. If you depend on this, test it.

You're building a product on top

Both products can be scripted. exe.dev's HTTPS API runs the same commands as its SSH interface through POST /exec, with a 30-second timeout and no stdin or pty.

boxd has open-source (MIT) TypeScript and Python SDKs, and every CLI command takes --json:

Terminal
boxd machine new my-agent --json
boxd machine exec my-agent -- npm test
boxd machine remove my-agent --confirm --json

You want to control exactly what a machine can reach

exe.dev's egress policy has two settings: all, or integrations, which blocks the public internet and keeps attached integrations working. boxd takes an allowlist of hosts for each machine, so an agent can reach your package registry and your API and nothing else. Setting it to none cuts the machine off from the internet, DNS included.

Both keep credentials away from the code running on the VM. exe.dev's HTTP proxy integration adds a header to requests made through it. With boxd's host-bound secrets, the machine holds only a placeholder, and boxd swaps in the real value on HTTPS requests to the hosts you name.

You need a big machine

boxd's standard sizes run from 1 vCPU and 4 GiB to 8 vCPU and 32 GiB, always at 4 GiB per vCPU. Above that, you can ask for any size up to 128 vCPU and 1.5 TB of RAM. A standalone exe.dev VM has up to 2 GiB of memory per vCPU, which is tight for builds and large test suites.

Speed

We ran the open-source aisandboxbench harness on boxd in November 2026, 100 runs per test, using a caller near each region. Medians, lower is better:

Testboxd
New machine to first command0.36 s
Run a command32 to 37 ms
Reconnect to an existing machine0.13 to 0.16 s
Pause and resume, total68 to 80 ms

exe.dev isn't in the published results for this benchmark, so we can't put a number beside these. boxd is mid-pack on raw CPU, memory and disk throughput against the providers that are.

Which should you pick?

Pick exe.dev if you want a region in Asia-Pacific or Australia, you run many small VMs from one shared pool, or you want disk changes copied off-host without setting anything up.

Pick boxd if you have to run on your own infrastructure, your agents need to resume with memory and processes intact, you're building on an SDK, or you need per-host egress rules or machines above 32 vCPU.

Both will give you a good Linux machine over SSH. To try boxd:

Terminal
curl -fsSL https://boxd.sh/downloads/install.sh | sh

Also comparing Fly.io's product? Read boxd vs Sprites.


Last checked 10 November 2026. exe.dev details come from its docs. boxd's speed numbers come from the aisandboxbench harness, which anyone can run. This is one vendor's comparison of a competitor, so please check anything your decision depends on.

Frequently asked

Should I use boxd or exe.dev?
exe.dev if you want more regions, including Asia-Pacific and Australia, or a pool of CPU and RAM shared across many VMs. boxd if you need to self-host, want paused machines to keep their memory, or are building on TypeScript or Python SDKs.
What is a good exe.dev alternative?
boxd is the closest alternative if you want the same persistent Linux VM over SSH. It adds self-hosting and BYOC, standby that keeps memory and running processes, per-host egress allowlists, and machines up to 128 vCPU and 1.5 TB of RAM on request.
Can I self-host exe.dev?
We found no self-hosted or BYOC option in exe.dev's docs as of November 2026. boxd self-hosts as a single Rust binary on Google Cloud, AWS or bare metal, deploys in 30 to 60 minutes, and is free up to 16 worker vCPUs.
Does exe.dev have an API?
Yes. exe.dev publishes an HTTPS API that runs the same commands as its SSH interface through POST /exec, with a 30-second timeout and no stdin or pty. boxd has a CLI with --json on every command, plus open-source TypeScript and Python SDKs.
Where does exe.dev run compared with boxd?
exe.dev lists nine regions: Oregon (closed to new accounts), Los Angeles, New York, Dallas, Frankfurt, Tokyo, Sydney, Singapore and London. boxd runs in Frankfurt and New York, or in your own cloud if you self-host.
Can I run coding agents on exe.dev and boxd?
Yes, on both. boxd machines come with Claude Code, Codex and OpenCode preinstalled, and each machine is a KVM microVM with its own kernel. A per-host egress allowlist limits what an agent can reach.
Hidde KehrerHidde Kehrer
PostShare
Published
May 4, 2026
Reading time
6 min
Words
1,173
Last verified
Nov 10, 2026

Read next

Field notes

Subscribe for release notes and architecture write-ups

No spam, ever. Unsubscribe anytime.

Your inbox