# boat.dev

Official pricing: https://docs.boat.dev/pricing  
Category: agent-sandbox · Isolation: vm

## Pricing regimes (raw)

- **On-demand (flat per size, per second while running)** (sizes): small 2 vCPU/4 GiB $0.018/h; default 4 vCPU/8 GiB $0.036/h; large 8 vCPU/16 GiB $0.072/h; xlarge 16 vCPU/32 GiB $0.2/h

## Features

Yes: snapshots, fork/clone, pause/resume, persistent disk, ≥24 h sessions, custom image (Docker or snapshot), start from your own snapshot, full VM (own kernel), Docker inside, nested virtualization, root, systemd, browser, desktop GUI, computer-use API, browser automation API, desktop control API, browser + desktop control, anti-bot stealth, preinstalled agents, SSH, HTTPS preview URLs, raw TCP inbound, open internet, Python, Node.js, Go, Rust, Java, live fork (no pause), webhooks, EU data residency, zero data retention, scoped API keys, spend limits, automatic snapshots, snapshots on demand, agent harness API, gVisor or VM (no shared kernel), inbound access rules, firewall inside (nftables)

No: memory snapshots, volumes, idle auto-stop, your own Docker/OCI image, code interpreter, CAPTCHA solving, residential IPs, public IPv4, egress allowlist, static egress IP, private networking, self-hosting, BYOC, open source, SOC 2, HIPAA, SSO, GPU, arm64, Windows, macOS, GPU desktop, HTTP method/path egress rules, wake on request, live resize, memory fork, audit logs, MCP server, hosted agent API (their own agent), secret proxy, secret proxy for any API, extra volumes, shared volumes

Unknown: everything else. Evidence (source + quote) per feature: https://battleships.dev/data/providers/boat.json → feature_evidence

## Caveats

- Battleships is made by boat.dev; boat.dev's numbers come from its public docs and are held to the same standard as every other provider.
- boat is not in the ComputeSDK benchmarks: cold_start_ms, burst_tti_ms, dax_s unknown.
- Plan fee is 100% usage credit but expires monthly (effective $20/month minimum); overage bought as $20 packs that never expire. Engine fee_is_credit=true models this; pack residue (<$20) carried to next month is not modelled.
- Trial plan (trial_only) is one-time 25 h (~$0.90 of default time; if metered in default-seconds, small would get 50 h), 2 concurrent, small/default, TTL <= 2 h; it cannot be extended with paid usage.
- No idle auto-stop: TTL counts from creation/resume (default 1 h; forks reset to 1 h). Idle-but-running bills full size rate; --no-auto-stop (paid only) runs until stopped. FAQ: "Boat has no idle timer. The auto-stop TTL counts from creation or resume, never from last activity" (https://docs.boat.dev/faq.md, 2026-09-28), so auto_stop_idle=false is correct.
- Organization seats multiply concurrency and start limits (one pooled budget); seat price unpublished (per_seat null).
- Egress included up to 2 TB per sandbox per month (egress_free_gib=2000 is per sandbox, engine treats as account-wide = conservative); price beyond 2 TB unpublished.
- Public IP is 'dedicated IPv6 or IPv4' - IPv4 not guaranteed (ipv4_available null) and it changes on resume; no charge either way.
- Snapshots: latest per sandbox free for life of sandbox (FAQ: "Stopped sandboxes are free, and their snapshot is kept for the life of the sandbox." https://docs.boat.dev/faq.md), bounded by the size's data disk; up to 10 named snapshots ("You can keep up to 10"), no expiry (https://docs.boat.dev/snapshots.md). Named-snapshot pricing is not stated beyond that; snapshot_gib_month=0 rests on the FAQ line. Filesystem only (no memory), so resume = reboot.
- No extra-disk purchase: disk is fixed per size (12/50/125/251 GB), so disk needs can force upsizing. included_disk_gib=50 refers to default size.
- Homepage says 'dedicated 4 vCPU / 8 GB VM-time' while FAQ says 'Sandboxes have shared vCPUs'; treat vCPUs as shared.
- EU only (Germany, Finland, France); no US/Asia option; no region multiplier.
- Custom environments via named-snapshot templates and setup scripts, not Dockerfile/OCI images (custom_image=false).
- Out of balance without auto-pay: 24 h grace, then snapshot+stop. Upgrades prorated immediately; downgrade terms unpublished.
- Per-start minimum and whether boot/hydration time is billed are not documented (assumed 0).
- Performance: standard machines (AMD Ryzen 9 9950X, "baremetal" class) are the default. A temporary Hetzner fallback (older AMD, 8.88 vs 27.67 runs/s, no /dev/kvm) is used only when standard capacity runs out. Same price either way.
- Billing scope: running time only; stopped time is never counted and time past a refused stop is excluded (meter paused), per https://docs.boat.dev/billing.md (verified 2026-09-28).
- Start limits per plan (creates, forks and resumes all count): 12/min · 60/h · 200/day on $20, 30 · 210 · 840 on $100, 65 · 420 · 1,680 on $500, 90 · 600 · 2,400 on $2,000 (https://docs.boat.dev/pricing, 2026-09-28); multiplied by seats on an organization. Per-minute and per-day caps are enforced by the estimator; the hourly cap (max_starts_per_hour) is recorded but not yet enforced.
- Organization plan fee is per member (plan price x members). Whether the included machine time also scales x members is not stated (concurrency and start ceilings do). If an org subscription is cancelled, leftover packs still run but limits drop to one seat.
- Start limits can be lifted for committed accounts (sales); per-member caps (usage, concurrency) available to org owners.
- Prepaid balance ('funds') and auto-refill exist besides $20 packs: auto-pay charges the card 'in small automatic increments'. Grace-period overuse is carried as a shortfall and settled on the next payment.
- Snapshot retention conflict: docs say the latest snapshot is kept for the life of the sandbox and named snapshots have no expiry, but the Terms of Service (https://boat.dev/terms) say snapshots are 'generally retained for a limited period, currently up to approximately 30 days'. Treat free indefinite retention of stopped sandboxes as documented but not contractual.

## How this provider charges


As of 2026-09-28. boat is the user's own product. It is held to the same standard as the others: limits and inconsistencies are listed below.

Sources:
- P = https://docs.boat.dev/pricing (raw: `raw/docs_boat_dev_pricing.txt`)
- B = https://docs.boat.dev/billing.md
- F = https://docs.boat.dev/faq.md
- M = https://docs.boat.dev/machines.md
- S = https://docs.boat.dev/snapshots.md
- L = https://docs.boat.dev/long-running-tasks.md
- D = https://docs.boat.dev/data-retention.md
- H = https://boat.dev (raw: `raw/boat_dev.txt`)
- BM = `benchmarks.json` (hpc 2026-08-24)

## Core model in one line

A flat price per size, per second, **only while running**. CPU and RAM are not metered separately. Utilisation has no effect on the bill. Stopped sandboxes, snapshots, IPv4/IPv6 and up to 2 TB/sandbox/month of egress are free. The monthly plan fee is 100% usage credit.

## Regime table

| # | Regime | When it applies | How billed | Numbers | Source |
|---|---|---|---|---|---|
| 1 | **Size: small** | `--type small` | Flat $/h, per second while running | $0.018/h: 2 vCPU / 4 GB / 12 GB disk (= $0.009/vCPU-h, RAM included) | P |
| 2 | **Size: default** | Default size | Same | $0.036/h: 4 vCPU / 8 GB / 50 GB | P |
| 3 | **Size: large** | `--type large`. Not available on the trial | Same | $0.072/h: 8 vCPU / 16 GB / 125 GB | P, L |
| 4 | **Size: xlarge** (plan-gated) | Needs the $100+ plan **and** an operator to allocate capacity on request | Same | $0.200/h: 16 vCPU / 32 GB / 251 GB. This is **5.56x** default for 4x the resources, so $0.0125/vCPU-h, a **39% per-vCPU premium** | P |
| 5 | **Trial** | New account, until the first payment | One-time free hours | 25 h, 2 concurrent, small/default only, TTL ≤ 2 h, auto-stop cannot be disabled (about $0.90 of default time) | P, L |
| 6 | **Plan $20/mo** | Paid, 100 concurrent | Fee = usage credit that expires at month end; overage from packs/auto-refill | $20 credit (555 h default); starts 12/min, 60/h, 200/day | P |
| 7 | **Plan $100/mo** | 101–300 concurrent, or xlarge | Same | $100 credit; 300 concurrent; 30/210/840 starts | P |
| 8 | **Plan $500/mo** | 301–1,000 concurrent | Same | $500 credit; 1,000 concurrent; 65/420/1,680 starts | P |
| 9 | **Plan $2000/mo** | 1,001–2,000 concurrent | Same | $2,000 credit; 2,000 concurrent; 90/600/2,400 starts | P |
| 10 | **> 2,000 concurrent** | Enterprise | "By agreement" | unpublished | H |
| 11 | **Credit packs** | Usage beyond monthly plan time | Prepaid $20 packs at list rates, drawn after plan time | $20 = 2,000,000 default-seconds. Packs **never expire**. Auto-refill is optional | P, B |
| 12 | **Organization seats** | Org wallet with N seats | Concurrency and start limits × seats, as one pooled budget | Seat price **not published**. Per-member usage and concurrency caps are available | P, B |
| 13 | **Running (busy or idle)** | Any time the VM is up | Full size rate, regardless of CPU/RAM use | e.g. default 24/7 = 730 h × $0.036 = $26.28 | P |
| 14 | **Stopped** | After `boat stop`, TTL expiry, or balance cutoff | $0 compute, $0 disk, $0 snapshot | Latest snapshot kept for the life of the sandbox, no expiry | F, S |
| 15 | **Always-on** | `--no-auto-stop` (paid plans only), or TTL up to 30 d | Same as running. **No commit discount and no monthly cap** | $26.28/month per default sandbox | L, P |
| 16 | **Named snapshots (templates)** | `boat snapshot save` | Free | Up to 10, no expiry. Backing data is freed at least 6 h after deletion | S, D |
| 17 | **Standard machine (default)** | Where sandboxes normally run: dedicated AMD Ryzen 9 9950X (Zen 5) hosts. "baremetal" is only the internal class name; each sandbox is still a full VM with its own kernel | Same price | Node build loop 27.67 runs/s ($0.36/1M runs); `/dev/kvm` present, so nested virtualisation works | M, BM |
| 18 | **Capacity fallback (temporary)** | Only when standard capacity is temporarily exhausted: a new sandbox falls back to an older Hetzner cloud VM (older AMD CPU) | Same price | 8.88 runs/s (about 3.1x slower, $1.13/1M runs); **no `/dev/kvm`** | M, BM |
| 19 | **Egress ≤ 2 TB/sandbox/month** | Normal use | Included | 2 TB per sandbox per month. Ingress is free | F |
| 20 | **Egress > 2 TB/sandbox/month** | Heavy transfer | **Unpublished** | null | F |
| 21 | **IP address** | Every sandbox | Included | "Dedicated IPv6 **or** IPv4". IPv4 is not guaranteed and the IP changes on resume | F, M |
| 22 | **Out of balance** | Balance reaches 0 without auto-pay | 24 h grace, still running. Then snapshot and stop | No data loss | B |
| 23 | **Refused stop** | Final snapshot fails | Billing is paused while the stop is refused | "time past a refused stop is excluded" | llms-full |
| 24 | **Plan change** | Upgrade | Prorated, immediate. Moving a personal plan into an org refunds the unused share | Downgrade terms not stated | P, B |
| 25 | **Account closure** | Close account | 30-day recoverable window, then purge | $0 | D |
| 26 | **Region** | EU only (DE/FI/FR) | No regional multiplier. No US or Asia option | 100–200 ms RTT from the Americas | F |

Things **not** offered (so no regime exists): GPU, Windows/macOS, spot/preemptible, reserved/commit discounts, memory snapshots (pause with RAM), extra disk beyond the size preset, separate volumes, region choice.

## Gotchas

1. **The $20 plan is really a $20/month minimum.** The fee is fully credited, but the credit **expires monthly**. A light month (e.g. $6 of usage) still costs $20. Credit packs, by contrast, never expire.
2. **TTL counts from creation or resume, not from idle, and the default is 1 h.** A sandbox stops mid-work at 1 h unless you pass `--ttl`/`--no-auto-stop`. Forks do **not** inherit the source TTL; they default to 1 h again. This is safe for the bill but surprising for long jobs.
3. **There is no idle auto-stop.** An idle but running sandbox bills the full rate. With `--no-auto-stop`, a forgotten sandbox costs $26.28/month (default) until you stop it. Utilisation-based providers would charge less for an idle box; boat charges the same.
4. **Start-rate ceilings can force a higher plan before concurrency does.** Create, fork and resume each count. Example: 50 sandboxes resumed 8× a day = 400 starts/day, above the $20 plan's 200/day. That needs the $100 plan (840/day) even though 50 concurrent fits in $20. A burst of 50 creates takes about 5 min at 12/min.
5. **xlarge is poor value per vCPU.** It is 5.56x the default price for 4x the vCPU/RAM, 39% more per vCPU than small/default/large. It is also gated behind the $100 plan plus manual allocation. Two `large` sandboxes cost $0.144/h against $0.200/h for one xlarge.
6. **Disk only grows by upsizing.** No extra-disk purchase exists. Needing 100 GB forces `large` ($0.072/h) even if 4 vCPU is enough.
7. **The capacity fallback is slower and has no KVM.** Sandboxes normally run on boat.dev's standard machines (dedicated AMD Ryzen 9 9950X hosts; "baremetal" is only the internal class name, every sandbox is still a full VM): 27.67 runs/s, `/dev/kvm` present, $0.36 per 1M Node tooling runs. That was 85% of prod assignments over the 7 days to 2026-09-28 and ~95% over the last 3. Only when standard capacity is temporarily exhausted does a new sandbox fall back to an older Hetzner cloud VM at the same price: 8.88 runs/s (about 3.1x slower, $1.13/1M runs) and no `/dev/kvm`, so nested virtualisation (QEMU/KVM, Firecracker, Android emulator) is unavailable there. During a capacity crunch two weeks earlier the fallback took up to ~60% of assignments. The public docs are being corrected in PR #1489.
8. **"Dedicated" vs "shared" wording conflicts.** The homepage says "dedicated 4 vCPU / 8 GB VM-time". The FAQ says "Sandboxes have shared vCPUs" and advises dedicated compute for sustained 100% CPU.
9. **IPv4 is not guaranteed** ("IPv6 or IPv4"), and the IP changes on resume. Anything that needs a stable public IPv4 cannot rely on it.
10. **Egress beyond 2 TB/sandbox/month is unpublished.** It is also unclear whether heavy egress is throttled.
11. **Snapshot size is bounded by the size's data disk** ("up to your sandbox type's data size"). Snapshots are filesystem-only, so resume means a reboot and running processes are lost.
12. **Seat price is unpublished.** Seats multiply the concurrency and start ceilings, and the whole org shares one pool.
13. **EU-only.** There is no US region at any price.
14. **Trial is hours, not dollars.** It gives 25 h with 2 concurrent and a 2 h TTL cap. You cannot run the worked example below on it.

## Worked example

Workload: 4 vCPU/8 GiB, 50 concurrent × 8 h/day × 22 days = **8,800 sandbox-hours**, 30% CPU util, 50 GiB snapshots retained, 100 GiB egress.

The size is `default` (exact fit). CPU utilisation is irrelevant because billing is flat. The 50 GiB of snapshots is free (≈1 GiB per sandbox, well under the 50 GB data size). The 100 GiB of egress is free (≪ 2 TB/sandbox).

| Regime | Compute | Snapshots | Egress | Plan / packs | **Monthly total** |
|---|---|---|---|---|---|
| $20 plan + packs, stop between sessions (TTL 8 h) | 8,800 × $0.036 = $316.80 | $0 | $0 | $20 credit + 15 packs ($300); $3.20 of pack credit carries over | **$316.80** effective ($320 cash) |
| $100 plan + packs | $316.80 | $0 | $0 | $100 + 11 packs ($220), $3.20 carries over | **$316.80** effective ($320 cash) |
| $500 plan (no packs) | $316.80 | $0 | $0 | $500 fee, $183.20 of credit expires unused | **$500** |
| Same fleet but never stopped (always-on / forgotten `--no-auto-stop`) | 50 × 730 × $0.036 = $1,314 | $0 | $0 | $100 plan + 61 packs, or $2000 plan | **$1,314** |
| Upsized to `large` (e.g. to get >50 GB disk) | 8,800 × $0.072 = $633.60 | $0 | $0 | $20 + packs | **$633.60** |
| `xlarge` (overkill; needs $100 plan + allocation) | 8,800 × $0.200 = $1,760 | $0 | $0 | $100 + packs | **$1,760** |
| Performance-adjusted, CPU-bound, whole fleet on the temporary capacity fallback (3.1x runtime; only during a standard-capacity shortage) | ≈ $316.80 × 27.67/8.88 ≈ $987 (upper bound, same work) | $0 | $0 | | **≈ $987** (fallback worst case, not typical) |
| Trial | Not applicable: 2 concurrent, 2 h TTL, 25 h total | | | | n/a |

Plan choice: 50 concurrent fits the $20 plan (100). Daily resumes are 50 starts/day, under 200/day, but 12/min means about 5 min to bring the fleet up. If each sandbox restarts more than 4× a day, the $100 plan is needed for its start-rate ceiling; the dollar total is unchanged.