Arcbox

Arcbox

ArcBox gives each AI agent its own hypervisor-isolated virtual machine — a full OS with its own kernel, kept for months, thrown away in seconds, or handed to a

74/100Safe BetFree planFreemium

If you need agent isolation you can defend in a security review, ArcBox is the most literal answer available: one kernel per environment, snapshot restore around 22 ms, warm creates around 50 ms, per-second Linux and Windows metering from $0.0000140 per vCPU-second, and an MIT/Apache-2.0 runtime you can actually read on GitHub. The catch is supply, not capability — Linux and Windows are still waitlist-gated, Apple platforms go through sales with a 24-hour minimum lease, and no GPU option is published. Worth a waitlist spot if you already container-sandbox agents (compare with Firecracker-on-E2B or Fly Machines if you need immediate capacity).

Verified 11d ago · liveness 74/100 · cite: rightaichoice.com/tools/arcbox

Best for
  • Agentic AI platforms executing untrusted or self-generated code
  • Security researchers and red teams containing exploitable binaries
  • CI/CD pipelines wanting a clean VM per job
  • Platform engineers who need to read the runtime source (MIT/Apache-2.0)
Not ideal for
  • Teams that need production capacity immediately — Linux and Windows are waitlist-gated
  • GPU workloads — no GPU option is published
  • Anyone wanting self-serve iOS or Android environments (iOS Sandbox and Runner and Android Runner are not offered)
Visit Website

IntermediateLinux and Windows: minutes after your waitlist invite clears — install abctl or the Python/TypeScript SDK, create a machine, and you have a shell; the docs are public. macOS, iOS, and Android: days to weeks, since capacity is quoted with sales and Apple VMs carry a 24-hour minimum lease. ArcBox Desktop beta on a Mac: an evening's install to a first container.Web · API · CLI · Desktop · MobileAPI availableVerified 11d ago
Pricing
Free plan
FreemiumFree tier3 plans5 hidden costs
Learning curve
Intermediate
Linux and Windows: minutes after your waitlist invite clears — install abctl or the Python/TypeScript SDK, create a machine, and you have a shell; the docs are public. macOS, iOS, and Android: days to weeks, since capacity is quoted with sales and Apple VMs carry a 24-hour minimum lease. ArcBox Desktop beta on a Mac: an evening's install to a first container.
Runs on
WebAPICLIDesktopMobile
API available · 10 integrations
Who it's for
Agent platform engineerCI leadSecurity researcher or red team
Live sentiment
Is Arcbox actually worth it?

We scan live Reddit threads, YouTube comments, X posts, G2 reviews and other communities — and hand you an honest verdict in under a minute.

  • Honest verdict, not marketing
  • Real pros & cons from real users
  • Attributed quotes with receipts
Run a free scan

3 free scans · no card needed

Skip it if

Skip ArcBox if you need production agent capacity this week — Linux and Windows open in waves from a waitlist, and macOS, iOS, and Android are planned with sales rather than self-serve.

The 30-second take
Biggest gripe

Stopped Computers are not free: the disk and any snapshots stay metered at $0.10 and $0.05 per GiB-month, so idle persistent machines keep costing money.

Price reality

Free tier costs nothing to start — usage credits, Linux and Windows machines in all three forms, up to five concurrent environments, no card. Pro is usage-based from $0.0000140 per vCPU-second plus $0.10 per GiB storage-month, which undercuts flat-rate hosted VM plans for bursty agent work but can surprise teams with always-on machines. Enterprise is custom and is the only route to macOS, iOS, Android, and SSO/audit-log controls.

In short

Arcbox — ArcBox gives each AI agent its own hypervisor-isolated virtual machine — a full OS with its own kernel, kept for months, thrown away in seconds, or handed to a. Best for Agentic AI platforms executing untrusted or self-generated code, Security researchers and red teams containing exploitable binaries, CI/CD pipelines wanting a clean VM per job. Free to use.

What's new in Arcbox

Checked 3 days ago

Across the latest 2 updates: 2 news mentions.

What people actually say about Arcbox — is it worth it?

We scanned public community sources for Arcbox on Sep 1, 2026 and could not establish that the discussion we found is about this tool rather than something else sharing its name. Our own analysis of that scan says the posts were off-subject. Rather than publish a sentiment score built on the wrong subject, we publish nothing here and re-run the scan.

Viability Score

74/100
Safe Bet

How well maintained and how widely used is Arcbox? Built from what the vendor actually publishes (docs, changelog, tutorials, integrations, pricing), whether the site is live, and how much real users discuss it. How we calculate this

Recent activity
90
Traction
100
Site health
95
User sentiment
43
What the vendor publishes
40

Last calculated: October 2026

How we score →

Key Features

  • Hypervisor-isolated virtual machine with its own kernel per environment
  • Computer: persistent machine that snapshots disk and memory and resumes later
  • Sandbox: ephemeral microVM created in milliseconds, destroyed on a TTL
  • Runner: fresh machine for every CI job, torn down after it finishes
  • Snapshot restore in ~22 ms; warm create in ~50 ms
  • Lifecycle event stream (ready, exited, snapshotted, destroyed) instead of polling
  • Ubuntu Linux x86-64 and arm64, any OCI image as root filesystem
  • Windows 11 and Server 2022, full desktop sessions for GUI agents or headless
  • macOS Apple silicon hosts, leased with a 24-hour minimum per Apple licence terms
  • iOS Simulator and device lab as long-lived Computers on macOS hosts
  • Android Emulator with x86-64 images and snapshot restore
  • abctl CLI with create, exec, snapshot, destroy across every OS
  • Python and TypeScript SDKs plus a REST API
  • Per-environment network identity with port publishing and egress rules
  • Open-source ArcBox Core runtime (MIT/Apache-2.0) and ArcBox Desktop public beta

About Arcbox

FreemiumIntermediateAPI availableWeb · API · CLI · Desktop · Mobile

ArcBox is an Agent Computer platform: instead of putting agents in a shared container namespace, it boots each one its own virtual machine with its own kernel, filesystem, and network identity. The isolation boundary is the hypervisor, not a namespace and not a permission prompt. The same isolated machine comes in three lifetimes. Computer is persistent — create it, work, snapshot disk and memory, stop it, resume later on the same disk and the same address. Sandbox is an ephemeral microVM, created in milliseconds and destroyed on a TTL. Runner boots a fresh machine for every CI job and tears it down after. All three are rows in one lifecycle event stream (ready, exited, snapshotted, destroyed), so agents subscribe instead of polling. Five operating systems are on the board. Ubuntu Linux (x86-64 and arm64, any OCI image as root filesystem) and Windows 11 and Server 2022 with full desktop sessions are waitlist-gated and, once in, billed self-serve by the second: $0.0000140 per vCPU-second, $0.0000045 per GiB memory-second, $0.10 per GiB storage-month, $0.05 per GiB-month for snapshots, $0.05 per GiB egress with ingress free, plus the Windows licence passed through at $0.046 per hour. macOS, iOS, and Android are planned with sales because Apple's licence forces a 24-hour minimum lease per VM. macOS and Android are billed per dedicated host-hour or environment-hour; iOS is an environment-hour. iOS Sandbox and Runner and Android Runner are stated as not available. Tooling is the same four verbs everywhere — create, exec, snapshot, destroy — through the abctl CLI, Python and TypeScript SDKs, and a REST API. The runtime underneath (VMM, VirtIO devices, userspace TCP/IP stack) is ArcBox Core, MIT/Apache-2.0 on GitHub, shipped alongside ArcBox Desktop, an open-source macOS Docker Desktop alternative in public beta. If you are container-sandboxing agents today and a security reviewer keeps asking what actually contains the code, this is the literal answer — as long as you can wait for capacity.

Behind the Verdict

ArcBox's strongest claim is architectural honesty. Every Computer, Sandbox, and Runner is a virtual machine with its own kernel; the boundary is the hypervisor rather than a namespace or a permission prompt. If your agent writes and then executes its own code, that distinction is the whole product. The second-strongest claim is lifecycle: ready, exited, snapshotted, destroyed arrive as an event stream you subscribe to, so an agent knows the machine is up because the machine said so rather than because it polled. Combine that with snapshot-based warm creates — a fleet of identical machines is one snapshot away — and you get the scaling story agent platforms actually need. The three lifetimes map cleanly onto real work. Computers for stateful agents that should keep their disk between tasks (stopped machines keep the disk and are metered for storage only, no vCPU or memory until resumed). Sandboxes for untrusted code with a TTL and no minimum charge — a sandbox that runs four seconds costs four seconds. Runners for CI, one clean image per job. The unified verb surface (create, exec, snapshot, destroy) across CLI, Python, TypeScript, and REST means you are not maintaining three mental models. Weaknesses are mostly about access and pricing shape. Linux and Windows open in waves via waitlist; macOS, iOS, and Android are quoted with sales, with iOS available only as long-lived Computers (no disposable form) and Android CI runners not offered at all. No GPU option is published, so training or GPU-bound inference is not a fit. Metered per-second billing is excellent for bursty agent loads and awkward for budget planners who want a flat monthly number. Apple pricing is per dedicated host-hour or environment-hour with a 24-hour minimum lease, which makes short experiments on macOS expensive relative to Linux. And if you need production capacity this week, the waitlist is a real blocker rather than a formality. Where it fits: agentic platforms executing untrusted or self-generated code, red teams containing exploitable binaries, and CI pipelines that want a clean VM per job. Also of note for Mac developers — ArcBox Desktop is a public beta, open-source Docker Desktop alternative, and the team publishes its measurements, including a benchmark showing Apple's balloon device does not release pages host-side in Linux VMs, which Docker, Podman, and OrbStack are all working around.

Researching Arcbox? Get your full AI stack in 60 seconds.

Free, no signup — tell us your goal and get tools matched to your budget & existing stack.

Real-world workflow fit

Concrete scenarios for the personas Arcbox actually fits — and what changes day-one when you adopt it.

Agent platform engineer

You run agents that generate and execute their own Python. You call abctl sandbox create --from-image python:3.12 --ttl 900 per task, subscribe to the lifecycle event stream for ready and destroyed, and never mount host files — files are copied in and results copied out.

Outcome: Each task runs inside its own kernel and network identity; a compromised agent cannot reach the host or a sibling environment, and a four-second task costs four seconds.

CI lead

You replace a shared CI runner pool with ArcBox Runners, booting a fresh Ubuntu 24.04 image per job and tearing it down after, driven from GitHub Actions or GitLab.

Outcome: Every job starts from exactly the image you chose with a clean kernel, and no state leaks between builds.

Security researcher or red team

You detonate an exploit or untrusted binary in a Sandbox with its own network identity, publishing only the port you mean to and denying egress by default.

Outcome: Containment is a hypervisor boundary you can point an auditor at, and the TTL destroys the environment without a manual cleanup step.

Use Cases

Limitations

  • Ubuntu Linux and Windows are waitlist-gated at launch, while macOS, iOS, and Android are only available by contacting sales. iOS is offered solely as a long-lived Computer on macOS hosts — there is no Sandbox or Runner form — and Android has no Runner form at all. macOS hosts carry a 24-hour minimum lease per Apple licence terms, and Android is billed per environment-hour, so short jobs on those platforms cost far more than the equivalent Linux sandbox.
  • Stopped Computers keep their disk and are metered for storage only, meaning idle persistent machines still accrue charges.

as of 2026-09-27

Verification history

We have re-verified Arcbox 7 times since . Each pass re-reads the vendor's own pages and re-checks every listed field against that evidence; passes where nothing had changed are marked as such.

  1. — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
  2. — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
  3. — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
  4. — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
  5. — re-checked, vendor evidence unchanged
  6. — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it

Showing the 6 most recent of 7 verification passes.

Free to cite with attribution — this page re-verifies continuously.

12-month cost

Project the real annual outlay, including the implied monthly cost when only an annual tier is published.

Annual total
Free
Over 12 months
Effective monthly
Free
Billed monthly

Vendor list price only. Add-on usage, seat overages, and contract minimums are surfaced under Hidden costs & gotchas.

Plans compared

For each published Arcbox tier: who it actually fits, and what it adds vs. the previous tier. Cross-reference the cost calculator above for projected annual outlay.

Free

$0 to start

Ideal for

A solo developer or small team testing whether per-environment kernel isolation fits an agent workload, with no card required to start.

What this tier adds

Starting tier — $0 to start with usage credits, Linux and Windows machines in all three forms, up to five concurrent environments, and community support.

Pro

Usage-based, billed by the second

Ideal for

Teams running agents in production on Linux and Windows who have cleared the waitlist and need more concurrency and bigger machines.

What this tier adds

Adds higher concurrency and larger machines, team workspaces and API keys, snapshot storage and warm pools, and priority email support on top of Free — billed by the second at $0.0000140 per vCPU-second.

Enterprise

Custom priced with you

Ideal for

Organizations that need macOS, iOS, or Android environments, private capacity, or SSO and audit logs for a security review.

What this tier adds

Adds macOS, iOS, and Android environments, dedicated hosts and private networking, SSO, audit logs, and an SLA, plus a named engineer — custom priced with you.

Hidden costs & gotchas

What the public pricing page doesn't put in bold. Captured from pricing-page footnotes, contract terms, and recurring complaints.

  • Stopped Computers are not free: the disk and any snapshots stay metered at $0.10 and $0.05 per GiB-month, so idle persistent machines keep costing money.
  • Windows machines add a $0.046 per hour licence fee passed through at cost on top of vCPU, memory, and storage metering.
  • Snapshot and egress traffic are billed separately at $0.05 per GiB, so pushing large images or results out of an environment adds to the compute line.
  • Apple platforms are billed per dedicated host-hour or environment-hour with a 24-hour minimum lease, so a two-hour macOS experiment costs a full day.
  • Past the Free tier's five concurrent environments you are on usage-based Pro, where cost scales with every vCPU-second and GiB-second instead of a flat monthly fee.

Where the pricing makes sense

The company stage and team size where Arcbox's pricing actually pencils out — and where peers do it cheaper.

Free tier costs nothing to start — usage credits, Linux and Windows machines in all three forms, up to five concurrent environments, no card. Pro is usage-based from $0.0000140 per vCPU-second plus $0.10 per GiB storage-month, which undercuts flat-rate hosted VM plans for bursty agent work but can surprise teams with always-on machines. Enterprise is custom and is the only route to macOS, iOS, Android, and SSO/audit-log controls.

Setup time & first value

How long it actually takes to get something useful out of Arcbox — broken out by persona, not the marketing-page minute.

Linux and Windows: minutes after your waitlist invite clears — install abctl or the Python/TypeScript SDK, create a machine, and you have a shell; the docs are public. macOS, iOS, and Android: days to weeks, since capacity is quoted with sales and Apple VMs carry a 24-hour minimum lease. ArcBox Desktop beta on a Mac: an evening's install to a first container.

Switching to or from Arcbox

How to bring data in from common predecessors and how to get it back out — written for the switcher, not the buyer.

Migrating in
  • →From Docker Desktop on macOS: install the ArcBox Desktop public beta and keep running your OCI images on Apple silicon.
  • →From a container-sandbox agent stack: swap the namespace sandbox for abctl sandbox create with a TTL and the same four verbs.
  • →From GitHub Actions hosted runners: point jobs at ArcBox Runners to get a fresh machine per job.
  • →From Firecracker/microVM tooling you maintain yourself: adopt ArcBox Core (MIT/Apache-2.0) or consume the hosted control plane instead of operating the VMM.
Migrating out
  • ↗To a flat-rate hosted VM provider: less to learn, but you trade per-second metering and per-environment kernel isolation for pooled instances.
  • ↗To a container-namespace sandbox: cheaper and instant, but the isolation boundary becomes a namespace rather than a hypervisor.
  • ↗To self-hosted ArcBox Core: if you would rather run the MIT/Apache-2.0 runtime on your own hardware than wait for capacity.

Integrations

DockerUbuntuDebianKubernetesGitHub ActionsGitLabPythonTypeScriptClaudecURL

Resources & Guides

Tutorials & Learning

YouTube returned 6 videos for “Arcbox”, and we withheld 6: 6 could not be judged, because “Arcbox” is a single word that other videos use for other things. We are showing none, because we could not prove any of them are about Arcbox.

Official links

Tools that pair well with Arcbox

Common stack mates teams adopt alongside Arcbox, with the specific reason each pairing earns its keep.

Featured Head-to-Head Comparisons

Alternatives to Arcbox

View all
Mem0

Mem0

AI memory layer that gives agents and apps persistent, cross-session context

FreemiumTry
Amazon Sage Maker

Amazon Sage Maker

AWS's integrated platform for data, analytics, and machine learning

PaidTry
Granica AI

Granica AI

Granica cuts data lake storage and compute costs and keeps long-running AI agents from losing their place — all inside your own cloud.

Contact SalesTry

Frequently Asked Questions

Used Arcbox? Help shape our editorial sentiment research.