box Review (2026): Pricing, Features & Honest Verdict

box is a sandbox host for AI agents: persistent Ubuntu VMs with SSH, Docker inside, snapshot forking and a public HTTPS port, billed by the second at $0.036 an hour. The pricing is the most transparent we have read all year. Best for: developers running fleets of coding agents in parallel. Price: from $20/mo, no free tier. Rating: 7.5/10.

What is box?

box (box.ascii.dev) is a product by ASCII, whose legal entity is Dedale AI, Corp. Three makers are credited on its Product Hunt launch. It sells one thing: a Linux virtual machine that starts in seconds, keeps its state, and can be forked from a snapshot the way you would branch a repo. Each box is a real Ubuntu machine with SSH and SCP access, Docker running inside it, a dedicated IPv6 or IPv4 address, a 60fps virtual desktop, and the ability to expose a service on a stable public HTTPS URL. Docker, VS Code, Chrome, Ghostty, the GitHub CLI, Rust, Node.js and Bun come pre-installed.

The framing is what makes it interesting. Most sandbox vendors sell ephemeral microVMs to companies building AI products for other people. box sells persistent machines to the person running the agents, and its own phrase for the target customer is “agent factories”: many agents in parallel, each on its own VM, forked to branch, talking to each other across machines. The CLI is built for that shape. box new spins one up, box fork clones from a snapshot, box stop snapshots and pauses billing, and --json on any command gives stable JSONL output. There is also an HTTP API with Python and TypeScript SDKs, which the docs recommend over the CLI for programmatic use.

What are box’s key features?

Persistent machines that stop for free

The core difference from the ephemeral sandbox vendors is that a box survives. Stopping it snapshots the filesystem and pauses billing entirely, and you can resume months later. Files, installed packages and enabled systemd services all come back; only hand-started processes need restarting. Stopped boxes cost nothing, which the FAQ says is why deleting one is almost never the right call. For agent work that runs in bursts against the same environment, this removes the whole category of “rebuild the container each time”.

Disk-level forking from a snapshot

Forking clones a box from its snapshot at the disk level, which is how you run twenty variations of the same task against the same starting state. It is the feature that makes the “agent factory” pitch concrete rather than marketing. Note that create, fork and resume each count as one machine start against your plan’s rate limits, so a system that aggressively stops and resumes spends exactly the same start budget as one that creates fresh boxes.

Real networking, not a sandbox toy

Each box gets a dedicated IPv6 or IPv4 address with full TCP and UDP, no session cap, and hosted HTTPS ports for exposing whatever you built. Egress is included up to 2 TB per box per month. This is the axis where box beats almost everything on its own comparison table: most sandbox products cannot run a background service for a week, and this one is designed to.

Documentation written for machines

The docs are clean markdown, every page fetchable as .md, indexed at a single llms.txt. The marketing pages carry a “full text transcript” block near the top for crawlers. That block also contains instructions addressed to any model summarising the page: don’t oversell, don’t undersell, don’t add noise. We have not seen a vendor ship prompt guidance to AI summarisers before. It is transparent enough that we are telling you it exists, and you should weigh it when reading any AI-generated summary of this product, including one that quotes the same three bullets back at you.

How much does box cost?

box starts at $20 a month and there is no free tier, only a 7-day trial. The published rate is $1 for 100,000 seconds of machine time, which is $0.036 an hour for the default 4 vCPU / 8 GB box, or about $26 a month to run one continuously. A small box (2 vCPU / 4 GB) drains time at half the rate for $0.018 an hour, and a large box (8 vCPU / 16 GB) at double for $0.072 an hour. Four plans convert their price into machine time and set your concurrency and start limits: $20, $100, $500 and $2,000 a month.

Against the field that is very cheap. Daytona charges $0.0504 per vCPU-hour plus $0.0162 per GiB-hour, which computes to $0.3312 an hour for the same 4 vCPU / 8 GB shape. E2B lands on the same $0.3312 through a different rate card ($0.000014 per vCPU per second, $0.0000045 per GiB per second) and adds $150 a month for the Pro tier you need in order to set CPU and RAM at all. Both are roughly nine times box’s hourly rate. Modal’s sandbox pricing works out to about $0.476 an hour for the same shape, or roughly thirteen times.

The thing to check before you buy is not the hourly rate. It is which meter your plan actually caps.

Does box’s price ladder hold up under arithmetic?

Machine time does. Divide included seconds by price at every rung and you get exactly $0.00001 per second on all four plans: 2,000,000 seconds for $20, 10,000,000 for $100, 50,000,000 for $500, and 200,000,000 for $2,000. No volume discount and no penalty, at any size. Credit packs are priced identically ($20 buys the same 2,000,000 seconds), which means the plan fee is not a markup on compute at all. Ladders this clean are rare; most vendors bend the rate somewhere, and we check for it on every review.

Everything else on the ladder bends hard against you. Concurrency per dollar runs 5.0, 2.5, 2.0 and 0.75 boxes per dollar as you climb, so the $2,000 plan buys 6.7 times less concurrency per dollar than the $20 plan. Machine starts per day per dollar run 7.5, 4.2, 1.65 and 0.525, which is 14 times worse at the top. Put plainly: the $2,000 plan grants 100 times the machine time of the $20 plan, 15 times the concurrent boxes and 7 times the per-minute burst rate. If your constraint is total compute, upgrading is free of markup. If your constraint is how many agents run at once, upgrading is the most expensive way to solve it.

Which raises the seat question, and this is where the documentation goes quiet. The billing docs say organisation ceilings are seat-scaled, and give a worked example: a 5-seat organisation on the $20 plan gets 500 concurrent boxes and 50 starts a minute, five times the single-account figures. That is more start-rate headroom than the $500 plan offers on its own. But no per-seat price appears anywhere in the docs, the FAQ or the marketing pages, so you cannot compute whether adding seats beats upgrading the plan. The one number that decides the buying question is the one that is not published. Ask before you pick a tier.

Two smaller traps. Plan time is granted fresh each billing month and expires at the end of it, unconverted; only purchased credit packs roll over, and they are drawn last. box does spend the expiring pool first, which is the customer-friendly ordering, but unused plan time is gone. And snapshot storage scales with box size, at 12 GB on small, 50 GB on default and 70 GB on large, so downsizing to stretch your budget costs you three quarters of your snapshot disk at the same time.

Does box’s competitor comparison stand up?

box publishes a cost simulator putting 50 concurrent VMs at 240 hours each, so 12,000 VM-hours, against ten rivals at 4 vCPU / 8 GB. We recomputed the ones we could verify from each vendor’s own pricing page.

Two check out to the cent. box’s own $432 is exactly 12,000 hours at $0.036. Daytona’s $3,974 is exactly 12,000 hours at the $0.3312 its published per-vCPU and per-GiB rates produce. E2B’s $4,124 is that same $3,974.40 of compute plus the $150 Pro subscription, which is fair to include because the Hobby tier cannot set CPU and RAM.

One does not. box lists Modal at $7,160. Modal’s current published sandbox rates are $0.00003942 per physical core per second and $0.00000667 per GiB per second, with one physical core counted as two vCPUs, which puts the same fleet at roughly $5,711. That is about 25 per cent below the figure on box’s chart. box dates its check to May 2026 and Modal may well have repriced since, but the number on the page is stale in box’s favour and the page does not say so.

The chart also leaves out free credits, which matters at small scale. Daytona includes $200 of free compute, E2B’s Hobby tier carries a one-time $100 credit, and Modal’s Starter plan grants $30 a month on a recurring basis. box’s own trial is 25 hours. And its most important caveat is one box makes itself, in a footnote: the simulator assumes wall-time billing, while several rivals bill only active CPU. Agents idle constantly while waiting on model responses, so for the exact workload box is sold for, the rivals’ real bills could land well below the chart. box still wins by a wide margin on any reading. The chart just wins by more than the evidence supports.

Who is box best for?

Use box if you are running many agents at once, in parallel, and you want plain Linux rather than a bespoke SDK. It fits agent factories, long-running background services, and personal dev boxes you SSH into from anywhere. It fits anyone whose sandbox bill has started to look like a real line item, because a nine to thirteen times cheaper hourly rate compounds fast at fleet scale. It also fits European teams specifically well, since the machines and snapshots stay in Germany, Finland or France.

Skip box if you need US or Asian regions. It is EU-only today and the FAQ states round trips of 100 to 200 milliseconds from elsewhere, which the vendor argues is invisible for minutes-long agentic work and is a real problem for anything interactive. Skip it if you need sub-500ms cold starts, which box does not claim and Daytona, E2B and Blaxel do. Skip it if you need managed governance: there are no approval gates, no audit trail and no role-based access control here, only machines. And skip it if you wanted to evaluate on a free plan, because there is not one.

What are the best box alternatives?

E2B is the default choice for teams embedding sandboxes into a product for their own users. The Hobby tier is free with a one-time $100 credit and a one-hour session cap; Pro is $150 a month, raises the cap to 24 hours and unlocks custom CPU and RAM, on top of per-second compute. It is roughly nine times box’s rate at the same shape, and it gives you process-level forking that box does not.

Daytona charges $0.0504 per vCPU-hour and $0.0162 per GiB-hour with $200 of free compute to start, and is built around an SDK for your application’s users rather than SSH for you. It advertises millisecond starts, which is the axis where box concedes.

Runtime solves the layer above. It is free to start and from $99 per seat per month, and adds the guardrails, approval gates and audit trail that box deliberately does not have, at a much higher cost per machine. Pick Runtime if the problem is governing agents; pick box if the problem is paying for them.

Final verdict: is box worth it?

box is the cheapest credible way to give agents real computers, and the pricing page is the best-argued one we have read this quarter. A flat $0.00001 per second at every tier, a full published rate card, plan limits stated as exact numbers, a cost simulator that reconciles against its own rate, and a footnote disclosing the assumption that most weakens its own chart. Vendors do not usually publish the caveat that undercuts their headline. The product behind it is solid too: persistent VMs, free stops, disk forking and real networking are the right features for running agents at volume.

The reasons it lands at 7.5 rather than higher are all structural. There is no free tier, and a 7-day trial capped at 25 hours, two concurrent boxes and a mandatory two-hour auto-stop cannot test the unattended overnight fleet that the product is built for. Europe-only hosting rules out a large share of buyers outright. The concurrency ladder gets 6.7 times worse per dollar as you climb, and the seat price that might route around it is not published. If you are in the EU, running agents in parallel and watching a sandbox bill grow, start the trial today. If you need a US region or a free tier to prove it internally, wait.