# Unikraft Cloud

Official pricing: https://unikraft.com/pricing  
Category: agent-sandbox · Isolation: firecracker

## Pricing regimes (raw)

- **Self-serve plan pool (Hobby / Team / Pro concurrent running capacity)** (pool)
- **Enterprise: dedicated hosts / on-prem / BYOC (custom price)** (resource) [sales]

## Features

Yes: snapshots, memory snapshots, fork/clone, pause/resume, persistent disk, whole disk survives a stop, volumes, idle auto-stop, ≥24 h sessions, custom image (Docker or snapshot), start from your own snapshot, your own Docker/OCI image, full VM (own kernel), nested virtualization, root, HTTPS preview URLs, custom domains, egress allowlist, private networking, self-hosting, BYOC, SOC 2, HIPAA, GPU, arm64, Python, Node.js, credential injection, wake on request, live resize, live fork (no pause), memory fork, scheduled runs (cron), audit logs, EU data residency, automatic snapshots, snapshots on demand, gVisor or VM (no shared kernel), inbound access rules, firewall inside (nftables), secret proxy, secret proxy for any API, extra volumes, shared volumes, FUSE mounts, kernel features (eBPF, mounts, cgroups)

No: Docker inside, browser, desktop GUI, computer-use API, code interpreter, browser + desktop control, anti-bot stealth, CAPTCHA solving, residential IPs, preinstalled agents, SSH, public IPv4, static egress IP, open source, SSO, Windows, macOS, GPU desktop, HTTP method/path egress rules, MCP server, usage per sandbox (API), agent harness API, hosted agent API (their own agent)

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

## Company

- Ownership: independent
- Stage: seed [source](https://unikraft.com/blog/we-raised-6m-to-build-the-future-of-ai-cloud-infra)
- Accelerators: none found
- Total raised: $6M (estimated: Only the $6M Oct-2025 seed round has a disclosed amount (company blog). An earlier 2022 pre-seed round is confirmed to have happened (unikraft.com/about) but its amount was never disclosed, so true cumulative total is likely somewhat higher than $6M; a third-party aggregator (Startup Intros) estimates $12.0M across 2 rounds but this is unverified.) [source](https://unikraft.com/blog/we-raised-6m-to-build-the-future-of-ai-cloud-infra)
- Last round: seed, $6M, 2025-10, led by Heavybit (confirmed) [source](https://unikraft.com/blog/we-raised-6m-to-build-the-future-of-ai-cloud-infra)
- Investors: Heavybit, Vercel Ventures, Mango Capital, Firestreak Ventures, Fly VC, First Momentum Ventures, Chris Schagen, Adam Christian, Stefano Stabellini, Eric Fourrier, Neil Patel, Seif Lotfy, Mathias Biilmann Christensen, Søren Bramer Schmidt, Zac Smith, Suhit Gupta, Saqib E. Awan
- Revenue: unknown

Unikraft Inc. (brand 'Unikraft Cloud', product formerly launched as 'KraftCloud' via HN Launch HN on 2024-04-02) is a separate commercial company from the open-source Unikraft unikernel project (unikraft.org), which began as NEC Labs Europe research (2017) and became a Linux Foundation/Xen Project hosted project in 2019. Unikraft Inc. was founded 2021/2022 by the same core team (Felipe Huici CEO, Simon Kuenzer CTO, Alexander Jung CPO) to commercialize a cloud platform built on that research; it is not itself a Linux Foundation entity. No evidence of a YC or other accelerator batch was found (a YC directory search for 'unikraft' returns no matching company).

## Caveats

- Unikraft Cloud publishes NO per-hour, per-GiB or egress price: self-serve pricing is a flat plan fee that caps concurrent running capacity (Hobby 2 vCPU/4 GiB/2 instances, Team 8 vCPU/8 GiB/8, Pro 24 vCPU/32 GiB/24). Anything bigger is Enterprise (custom price). Modelled as a pool mode.
- Idle cost is $0 by design: scale-to-zero puts instances in standby in ~ms (stateful mode snapshots memory to disk) and standby instances don't count toward the running pool; standby storage included (8 GiB / 15 GiB / 1 TiB).
- Self-serve sandbox ergonomics: network shield egress filtering, GPUs, arm64, nested virtualization and live vertical scaling are Enterprise-only; instance forking is preview. The sandbox plugin (command exec + filesystem API, used by the JS SDK Sandbox class and the CLI exec/shell/copy commands) carries no Enterprise label in Release 14 or the plugins doc, which gates plugins only on BYOC/on-prem licences; explicit availability on Hobby/Team/Pro is not stated (re-checked 2026-10-14).
- Per-instance max vCPU/RAM per plan unpublished; max_ram_gib 8 comes from the quotas doc example (1 vCPU max there). The pool tiers may therefore overstate what a single instance can get.
- Whether running above plan capacity is blocked or billed is not documented (no overage prices).
- Egress/bandwidth not priced or capped on the pricing page; network values null.
- snapshot_gib_month 0 = standby storage included within plan quota; no price for going beyond the quota.
- Claimed <10 ms cold start / scale-to-zero is a vendor figure; no independent benchmark (perf null).
- Regions: dal, fra, sfo, sin, was metros; 28 more listed as potential on request.
- open_source=false for the platform; CLI/kraftkit/Terraform provider are open source (and the separate unikraft.org unikernel project).
- Launched 2024-04-02 as KraftCloud (Launch HN, closed beta); rebranded Unikraft Cloud; domain now unikraft.com.
- Re-checked 2026-09-28 (unikraft.com/pricing rendered in headless Chrome, unikraft.com/llms-full.txt): still no pay-as-you-go / per-instance / per-vCPU-hour rate; self-serve is only Hobby $0 / Team $39 / Pro $199 flat plans plus Enterprise custom. No PAYG regime can be carded.
- Release 14 (2026-10-05): live vertical scaling of memory/vCPUs, full-VM templates + scale-to-zero, GPU snapshots (GPU released while stopped/standby) and NVIDIA vGPU are all marked "Enterprise only"; vertical scaling is "coming soon to the hosted platform" (Limited Access). CLI Dockerfile builds, autokill and the CLI/JS-SDK sandbox commands are not Enterprise-labelled. Checked 2026-10-14.

## How this provider charges


Unikraft Cloud (unikraft.com, Unikraft Inc.) runs Dockerfile-built images as Firecracker-based microVMs
(Unikraft kernel or minimal "TinyX" Linux) with scale-to-zero and wake-up in about 10 ms (vendor claim).
Launched **2024-04-02** as KraftCloud in closed beta (Launch HN https://news.ycombinator.com/item?id=39902949;
earlier kraft.cloud HN posts date from 2023-07-21).

**Pricing is capacity-based, not metered.** The public page lists no per-vCPU-hour, per-GiB-hour or per-GB-egress price.
Each flat monthly plan caps the account's concurrent running pool and its standby (scaled-to-zero) pool.
Idle instances go to standby and are free within the plan. All figures below are list prices.

## Regime table

| Regime | When it applies | How billed | Numbers | Source |
|---|---|---|---|---|
| Hobby | Default, free forever, no card | $0; hard pool caps | 2 running instances, 2 vCPU + 4 GiB total running; 5 standby instances, 8 GiB standby storage; 10 GiB image storage; community support | https://unikraft.com/pricing |
| Team | Small production | $39/month flat | 8 running, 8 vCPU + 8 GiB total; 50 standby, 15 GiB standby storage; 50 GiB images; transparent migration | https://unikraft.com/pricing |
| Pro | "Agent fleets and sandboxes" | $199/month flat | 24 running, 24 vCPU + 32 GiB total; 5,000 standby, 1 TiB standby storage; 100 GiB images; user-defined proxy, shared memory, snapshot history & rollback, live kernel updates, priority support | https://unikraft.com/pricing |
| Enterprise (hosted / dedicated / on-prem / BYOC) | Anything above Pro, plus Enterprise-only features | Custom contract | "Unlimited instances", dedicated host fleet, on-prem/BYOC, SOC2/HIPAA/audit logs, named SRE, 1 h SLA; price not published | https://unikraft.com/pricing, https://unikraft.com/on-prem |
| Running | Instance is serving/awake | Counts against the plan running pool (vCPU, memory, instance count) | No per-hour charge published | https://unikraft.com/pricing, https://unikraft.com/docs/platform/quotas.md |
| Standby (scale-to-zero) | No traffic for the cooldown (policy `on`, or `idle`, which also covers idle TCP connections) | Not charged ("You're not charged for the service in this state"); counts against standby count and standby storage | Stateful mode snapshots memory to disk; wake takes ms | https://unikraft.com/docs/features/scale-to-zero.md |
| Templates / branches / checkpoints | Clone warmed state | Uses instance and standby quotas | Memory + volume state; forking is preview | https://unikraft.com/docs/features/branching.md, https://unikraft.com/docs/features/checkpoints.md |
| GPU ("full" QEMU VMs) | Enterprise preview | Unpublished; GPU stays assigned for the instance's whole lifetime, even while stopped | 1 NVIDIA GPU per instance max, model not selectable | https://unikraft.com/docs/releases/r12-thebe.md |
| Sandbox plugin, network shield, nested virt, arm64 | Enterprise only | Unpublished | Exec/filesystem API, egress filter plus secret injection | https://unikraft.com/docs/features/plugins.md, https://unikraft.com/docs/releases/r13-adrastea.md |
| Egress / bandwidth | All traffic | Not published | null | (none found) |
| Regions | Metro chosen per instance | No regional multiplier published | dal, fra, sfo, sin, was (28 more "potential" metros on request) | https://unikraft.com/docs/platform/metros.md |

## Gotchas

1. **No meter, no overage.** The docs don't say whether going over the plan's running pool is blocked or billed. The quotas API (`hard.live_vcpus`, `hard.live_memory_mb`) suggests hard limits.
2. **Self-serve capacity is small.** Pro allows 32 GiB of RAM in total. That is four 4 vCPU / 8 GiB instances at once (limited by memory), one on Team, and none on Hobby (4 GiB total).
3. **Per-instance size limits aren't published.** The quotas doc example shows 1 vCPU and 16 MiB–8 GiB per instance. Real per-plan limits may be lower than the pool totals suggest.
4. **The "sandbox" API is Enterprise-only.** The sandbox plugin (exec + filesystem, used by the JS SDK `Sandbox` class), egress filtering (network shield), GPUs, nested virt and arm64 all require Enterprise. Self-serve plans deploy Dockerfile services behind TLS/HTTP endpoints.
5. **Idle is genuinely free, but only if traffic stops.** With policy `on`, an open TCP connection keeps the instance running. Policy `idle` fixes this but can drop long-idle connections (tradeoff documented).
6. **"Full" (QEMU) VMs give up the headline features.** They lose scale-to-zero, templates, branching and checkpoints. GPU instances hold their GPU even while stopped.
7. **Customer-reported ~$0.30/h per browser** (Browser Use case study) refers to Browser Use's own bare-metal Enterprise deployment. It is not a list rate.

## Worked example

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

- Peak need is 50 × (4 vCPU / 8 GiB) = **200 vCPU / 400 GiB running at once**. The largest self-serve plan (Pro) holds 24 vCPU / 32 GiB, so **only Enterprise fits**, and its price is unpublished.
- Hourly utilisation (30% CPU) and hours don't change the self-serve bill. It is a flat fee sized to peak concurrency.
- Snapshots: standby memory snapshots are included within plan quota (Pro: 1 TiB), so 50 GiB costs $0 on Pro.
- Egress: not published (null).

| Regime | Feasible? | Monthly total |
|---|---|---|
| Hobby | No (4 GiB total < 8 GiB for one instance) | n/a |
| Team | 1 concurrent 4/8 instance only | $39 flat |
| Pro | 4 concurrent 4/8 instances (memory-bound), i.e. up to 704 h in this schedule | **$199 flat**. It covers 4/50 of the workload. |
| Enterprise | Yes | **unknown** (sales) |
| Hypothetical linear extrapolation (NOT an offered price) | 400 GiB / 32 GiB = 12.5 → 13 Pro-sized pools | 13 × $199 = $2,587/month. Running multiple Pro accounts is not documented as allowed; shown only as an order-of-magnitude reference. |

Closest honest figure: **unknown (Enterprise)**. At small scale (≤4 concurrent 4/8 instances), Pro is a flat **$199/month** however many hours they run.