In the world of coding agents—software that writes, tests, and deploys code autonomously—half a second might as well be an eternity. Or at least that's the wager Freestyle is making.
The Y Combinator-backed startup, which emerged from the accelerator's Summer 2026 cohort, launched a product in early April that spins up full Linux virtual machines in roughly 500 milliseconds. More striking still: the system can clone a running VM—memory, state, and all—in under 400 milliseconds, a feature the company calls "Live Forking." For developers building agents that need to test parallel branches of code or fuzz a live database without disrupting the original, that speed could matter.
The timing is deliberate. Freestyle enters a market that has become unexpectedly crowded in recent months, with LangChain, Modal, Docker, and a handful of well-funded competitors all launching sandbox environments tailored to autonomous coding tools. The question isn't whether agent infrastructure is hot—it clearly is. The question is whether Freestyle's technical bet, prioritizing full virtualization over the lighter-weight container approach most rivals favor, will resonate with the developers actually building these systems.
Not Just Containers in Disguise
Co-founders Ben Swerdlow and Jacob Zwang are emphatic about one point: these are not containers dressed up as VMs. They're actual Linux virtual machines with root access, per-guest kernels, and support for nested virtualization—VMs running inside VMs. That architectural choice unlocks features like eBPF instrumentation, systemd multi-user isolation, and the ability to run Docker itself inside a Freestyle environment, capabilities that don't translate cleanly to containerized setups.
"We're giving agents real computers," Swerdlow explained in the Hacker News thread announcing the launch, which drew 263 upvotes and 145 comments within 15 hours. Perhaps more revealing than the product pitch was the discussion that followed: developers probing the edges of what "forking" a live VM actually means, debating security trade-offs, and questioning when the overhead of a managed service justifies itself versus self-hosting.
The fork feature, in particular, seems to be Freestyle's sharpest differentiator. According to the founders' launch post, the system can pause a running virtual machine, clone it horizontally with no more than a 400-millisecond interruption, and hand you a duplicate environment—complete with captured memory state. One commenter noted interest in using it for UI fuzzing; another asked about forking a running Postgres database mid-transaction to test failure scenarios. The founders acknowledged the feature works best for deterministic workloads, where replaying state makes sense, rather than services with heavy external dependencies or non-reproducible side effects.
That nuance matters. In infrastructure, the devil lives in the edge cases.
Speed as a Product Feature
Freestyle advertises VM provisioning times under 700 milliseconds on its homepage, though the founders cited approximately 500ms in the Hacker News thread—a discrepancy that suggests either ongoing optimization or the messiness of real-world performance claims. Either way, the speed rivals what many developers expect from containers, traditionally the go-to for fast, ephemeral compute.
But speed isn't just about boot times. The platform also supports pausing VMs with zero cost while idle, a pricing model designed for teams that need burst capacity—think hundreds of concurrent agents spinning up to compile code, run test suites, or interact with live services—but don't want to hemorrhage money when those agents sit dormant.
The company offers SDKs in TypeScript and Python, with built-in Git repositories that sync bidirectionally with GitHub, granular webhooks, and integration guides for Bun, OpenCode, and the Mastra AI SDK. That set of integrations hints at where Freestyle sees early traction: teams building coding assistants and autonomous developer tools that need more than a sandbox—they need something closer to a full development environment, spun up on demand.
Pricing for the Burst Economy
Freestyle's free tier allows up to 10 concurrent VMs, 5,000 repositories, and 500 runs per month, with a default timeout of 15 seconds. The Hobby plan runs $50 monthly plus usage, expanding to 40 concurrent VMs and a 600-second timeout. The Pro tier costs $500 per month, supporting 400 concurrent VMs, 50,000 repos, and half a million runs. Enterprise customers get SAML, SCIM, and a 99.99% uptime SLA—though pricing at that level isn't public.
Usage billing sits at roughly four cents per vCPU-hour and just over a penny per GiB of memory per hour, with storage clocking in at fractions of a cent per GiB-hour. In the launch thread, one of the founders suggested the service makes sense when teams need to burst to hundreds of VMs or want advanced features like forking and GPU support. For smaller, steadier workloads, self-hosting might still pencil out.
It's a candid admission, the kind that makes you wonder if Freestyle is targeting a narrower market than its pitch might suggest.
The Sandbox Wars
The competitive landscape has shifted noticeably in the past few months. LangChain introduced LangSmith Sandboxes in mid-March, offering secure remote environments with deep hooks into its agent framework. Modal launched a solution page for coding agents, emphasizing sub-second spin-up and the ability to scale beyond 50,000 concurrent containers. Docker released Docker Sandboxes for desktop, positioning microVM-isolated environments with network controls for local agent development.
Daytona, which raised $24 million in a Series A round in early February, pitches the same core idea: "give every agent a computer." The company emphasizes instant sandboxes and snapshotting. E2B, using Firecracker-based microVMs, updated its billing and concurrency documentation in March. Signadot offers Kubernetes-native sandboxes for agents that need to interact with live clusters.
The technical distinctions are subtle but significant. Freestyle's VM-first architecture trades some of the container ecosystem's portability and simplicity for kernel-level isolation and access to features that agents increasingly need as they grow more autonomous. Whether that trade resonates depends on what developers are building—and how much control their agents require when they're compiling dependencies, managing system services, or interacting with running databases.
What Comes Next
Freestyle's homepage lists logos for Onlook, Wordware, Anything, and Vly, though the nature of these relationships hasn't been independently verified. The company deprecated an earlier "Dev Server" API in February in favor of the VM product, signaling a strategic pivot toward the agent execution market—a shift that suggests the founders saw which way the wind was blowing.
The Hacker News discussion revealed some of the questions still hanging over the product. Security trade-offs versus rootless containers. Fork and snapshot use cases that go beyond simple state cloning. When to self-host versus when to pay for managed infrastructure. One developer asked about forking for parallel testing scenarios; the founders clarified that the feature works best for deterministic workloads and services that benefit from instant state duplication.
Freestyle ships with SDKs for Node, Bun, PNPM, Yarn, and Python. The documentation emphasizes that teams can run any Linux software—not just containerized workloads. As AI coding agents move from proof-of-concept demos to production systems that need to compile code, manage dependencies, and interact with running services, the infrastructure layer is starting to look less like a container platform. It's starting to look like a fleet of ephemeral computers, spun up and discarded as needed.
Whether Freestyle's bet on speed and stateful forking defines that next generation, or whether it becomes a footnote in a market already thick with competitors, remains to be seen. For now, though, the company has a product, early customers, and a technical differentiation that at least sounds compelling. In the world of developer tools, that's enough to be worth watching.
