Back to vendors

Unikraft

Q: How much does Unikraft cost?

Tiered public plans: a solo tier for builders trying millisecond compute, a team tier, a scale tier for agent fleets and sandboxes, and an enterprise tier with dedicated hosts, on prem, and BYOC. Exact per tier numbers were not retrieved this session. Cost tracks compute usage.

Q: Who is Unikraft for?

developers, AI platforms, enterprise, infrastructure teams

Also known as: Unikraft Cloud

Visit site

Agent infrastructureindependentVerified 2026-07-08

Millisecond compute layer that runs AI agent sandboxes and headless browsers as strongly isolated microVMs, with instant boot, checkpoint and fork, and scale to zero.

Unikraft is a millisecond native compute layer for sandboxes, headless browsers, CI and builds, functions, agents, and databases. Built on unikernels (minimal specialized operating system images that run a single application), it delivers VM grade isolation with boot and scale to zero in under 10 milliseconds and can pack more than 100,000 strongly isolated instances onto a single server, cutting the cost and footprint of cloud workloads by an order of magnitude. Purpose built for agentic workloads, it lets teams fork a live virtual computer (and its children) in 10 milliseconds to spin up parallel sub agents, headless browsers, or debugging sessions, turn any runtime into a template to launch large numbers of instances, and checkpoint, rewind, suspend, and resume running VMs so idle agents cost almost nothing but wake instantly. It transparently transforms a Dockerfile into a minimal Linux image, runs on bare metal, cloud VMs, or as a node inside an existing EKS, GKE, or AKS cluster, and can be deployed on prem, BYOC, or on dedicated hosts. Provisioning is driven by the Unikraft Cloud REST API, and it integrates with existing developer tooling like Kubernetes, Prometheus, and Grafana. Customers running agent fleets on Unikraft include Browser Use, TinyFish, and Prisma.

Vendor details

Canonical URL

https://unikraft.com

Funding status

Raised $6M at launch (Oct 2025) with Vercel Ventures and Heavybit. Founded 2022; CEO Dr. Felipe Huici. Grew out of research backed by NEC Laboratories Europe and a Linux Foundation open source project (unikraft, ~4K GitHub stars).

Company status

independent

Use cases & customers

Primary use cases

agent sandboxesheadless browser infrastructurecode execution isolationephemeral compute at scalecheckpoint and fork of running VMs

Target customers

developersAI platformsenterpriseinfrastructure teams

Deployment options

SaaSon-premBYOCdedicated hostsbare metalAWSGCPAzureKubernetes node

Integrations

Dockerfiles, Kubernetes (EKS, GKE, AKS), Prometheus, Grafana, and the Unikraft Cloud REST API. Runs on bare metal or on AWS EC2, GCP Compute Engine, and Azure VMs. Supports databases such as Postgres, MongoDB, and Redis.

Sources & related URLs

Research sources

https://unikraft.com/ https://unikraft.com/pricing/ https://www.businesswire.com/news/home/20251009046776/en/Unikraft-Launches-With-$6M-to-Drive-Dramatic-New-Efficiencies-in-Scale-and-Cost-for-Cloud-Computing-in-the-AI-Era https://ossstartuppodcast.substack.com/p/episode-198-how-unikraft-launches https://browser-use.com/posts/two-ways-to-sandbox-agents https://www.crunchbase.com/organization/unikraft

Capability coverage

8.0 / 14 capabilities · 57%

Integrations & Tool CallingIntegrates with existing developer tooling including Dockerfiles, Kubernetes (EKS, GKE, AKS), Prometheus, and Grafana, exposes the Unikraft Cloud REST API for provisioning, and runs on bare metal or on AWS EC2, GCP, and Azure VMs.	Full
Workflow OrchestrationSchedules microVMs across metros, scales instances to zero and back, and can fork a running virtual computer and its children to spin up parallel sub agents and headless browsers. This is compute level scheduling and fan out rather than agent workflow orchestration.	Partial
Knowledge Grounding & RAGUnikraft is a compute substrate and does not provide knowledge grounding or retrieval augmented generation. Any RAG is implemented by the workloads it hosts.	Unable to verify
Human Oversight & GuardrailsAs an infrastructure layer, Unikraft does not document human in the loop approval or agent guardrail constructs. Those belong to the agent frameworks running on top of it.	Unable to verify
Security, Identity & GovernanceVM grade, strongly isolated execution is the core value proposition, with unikernels exposing a minimal attack surface. The Browser Use case study shows each agent running in its own isolated microVM with no access to host credentials, and on prem and BYOC options keep workloads in the customer's environment.	Full
Observability & AuditabilityIntegrates with Prometheus and Grafana for metrics and monitoring of workloads. No built in agent action audit log is documented at the platform layer.	Partial
Memory & State PersistenceLets teams checkpoint a running virtual computer at any point, rewind it, or launch a new VM from a checkpoint, and provides pause and resume with scale to zero so an idle sandbox suspends and later resumes its exact state instantly.	Full
Deployment & Data ResidencyDeploys on prem, in the customer's own cloud (BYOC), on dedicated hosts, on bare metal, on AWS, GCP, and Azure VMs, or as a strongly isolated node inside an existing EKS, GKE, or AKS cluster, letting customers own the full stack.	Full
Prebuilt Agents / Templates / PacksAny sandbox, headless browser, or build runtime can be turned into a template to launch large numbers of instances, and the platform supports thousands of such templates. These are runtime templates rather than prebuilt agents.	Partial
Triggers & Channel CoverageSandboxes are provisioned on demand through the REST API and suspended VMs wake instantly on the next request via scale to zero. Triggering is API and request driven rather than broad channel or event coverage.	Partial
Model Flexibility & RoutingUnikraft runs any workload but does not provide model selection or routing itself. In the documented Browser Use architecture, LLM calls are proxied through the customer's own control plane, not Unikraft.	Unable to verify
APIs / SDKs / MCP ExtensibilityExposes the Unikraft Cloud REST API for provisioning and control, is built around standard Dockerfiles and Kubernetes, and the core unikraft kernel is open source, giving broad programmatic extensibility.	Full
Testing, Debugging & OptimizationEnables parallel debugging and build sessions and, per the Browser Use case study, spinning up hundreds of identical sandboxes in parallel for evaluation pipelines, so it supports testing and eval workloads at scale without being an eval harness itself.	Partial
Browser / Computer-useProvides first class support for running headless browsers and computer use style sandboxes as isolated microVMs (including parallel headless browsers to retrieve webpages), and hosts browser agent platforms, though it does not itself perform browser automation.	Partial

Recent platform changes

No recent material changes tracked yet.

View all changes for Unikraft →

Pricing

Tiered public plans: a solo tier for builders, a team tier, a scale tier for agent fleets and sandboxes, and an enterprise tier with dedicated hosts, on prem, and BYOC. Exact per tier numbers not retrieved this session.

compute usage (instances, VM runtime, resources)

Public — partialHigh variable costFree tierTrial available

Cost watchouts

Usage based compute cost grows with fleet size and runtime. Enterprise dedicated, on prem, and BYOC are quote based.

Variable cost rationale

Compute infrastructure billed on usage, so total cost scales with the number of instances, VM runtime, and resource consumption rather than a fixed plan fee. Scale to zero reduces idle cost.

Sales call required

Mixed (some tiers require a call)

Free / trial

Solo tier for trying millisecond compute

Lowest paid plan

Solo tier (numbers not retrieved this session)

Key ambiguities

The public pricing page lists tier names and descriptions but exact per tier prices and usage rates were not retrieved this session.

Verified 2026-07-08

Official pricing page

Data confidence: high