Boxd's $2M pre-seed backs full computers for coding agents

Why it matters

Coding agents are creating demand for persistent, isolated compute that can branch into parallel tasks. Boxd is betting that machine layer can remain independent of model vendors and large clouds.

A $2 million pre-seed round led by BlueYard Capital is backing Boxd founders Michiel Voortman, Laurentiu Ciobanu and Hidde Kehrer in their bet that AI coding agents need full, persistent computers of their own. Tech.eu reported on September 15th that OVNI Capital, Antler, S20 and Script Capital also joined the financing, alongside angels from General Intuition and other technology businesses. The date when the round closed was not specified.

The founders arrived at the problem from different parts of the software stack. Voortman studied at the University of Amsterdam and earned several Microsoft Azure certifications. Ciobanu previously served as CTO at video-conferencing startup Sessions and at Webmarc. Kehrer studied at Leiden University and worked at Dutch consultancy Boer & Croon before turning his attention to the machines beneath coding agents.

That combination matters because Boxd is making an infrastructure argument, rather than wrapping another model in a developer interface. The founders believe autonomous coding systems will work across several tasks at once, stay active after a developer closes a laptop and execute code that no engineer has reviewed. Those jobs need compute environments that retain their state, can be copied quickly and put a hard boundary around generated code.

A full computer for every agent

Boxd provides hardware-isolated Linux virtual machines with root access, persistent disks, networking and dedicated kernels. Developers can create a prepared machine, snapshot it and fork copies for several agents to pursue different implementations in parallel. A human can later enter the same environment over SSH to inspect or continue the work.

Kehrer has framed the product around a direct thesis: "Every AI coding agent needs a real computer to run on rather than a disposable sandbox," he told Tech.eu. The founders expanded that position in an August 14th manifesto, arguing that a neutral machine layer should support whichever coding agent a developer chooses instead of belonging to a single model provider.

Boxd's architecture uses KVM virtual machines and a control system coordinated through Raft consensus. In an engineering breakdown published in March, Kehrer wrote that Boxd separates control, worker, proxy and DNS responsibilities while avoiding a separate database, message broker and service-discovery layer. Each virtual machine runs as its own operating-system process, an architecture intended to let machines continue running if a supervisor process fails.

The security pitch starts with the dedicated kernel. Boxd says generated code runs behind a hardware boundary instead of sharing a host kernel through a container. Boxd also strips its virtualization layer down to CPU, memory, storage and networking, leaving out emulated peripherals such as webcams, audio devices and USB ports. That reduces the amount of virtualization code exposed to an agent, according to Boxd.

Boxd's speed claims need a narrower reading. Tech.eu reported that live forks, including memory and active connections, can be created in under 100 milliseconds. Boxd's current product material gives a range of 100 to 200 milliseconds. Both figures come from Boxd, and the difference leaves the exact performance of production workloads dependent on machine configuration and operating conditions.

The round buys engineering time

Boxd plans to use the financing to hire and continue building its custom virtualization engine. OVNI Capital, Antler, S20 and Script Capital joined BlueYard in funding a systems-heavy product whose core work sits below the application layer.

That is a capital-intensive place to compete. Infrastructure customers expect isolation, predictable billing, low startup latency and machines that survive failures without losing work. Agent workloads add another problem: a single user may ask several machines to run simultaneously, then leave most of them idle while inference or human review catches up. Boxd's commercial proposition depends on hibernating those machines cheaply while preserving enough state to wake them without rebuilding the environment.

The product starts accounts with capacity for 50 machines and provides 2 vCPU, 8 GiB of memory and 100 GiB of disk by default, according to Boxd's pricing page. Boxd also offers self-hosting, giving teams the option to run the platform on their own hardware. That may prove useful for customers with data-location, security or internal-network requirements that make a shared public service harder to adopt.

A crowded machine room

The founders have picked a market where several rivals already use nearly identical language. Daytona raised a $24 million Series A in February around its pitch to give every agent a composable computer. E2B raised $21 million in July 2025 for cloud sandboxes used by coding, data and browser agents. Runloop raised a $7 million seed round that month for development environments and evaluation tools aimed at enterprise coding agents.

Fly.io brought an existing cloud platform into the same category on July 24th, when it announced $25 million in Series D financing and made "computers for agents" its central strategy. Fly.io said more than 8,000 of its 37,000 customers were agent-native at the time of that announcement, figures supplied by Fly.io.

Boxd enters that contest with a smaller round and a focused technical identity: persistent machines, live memory forks, hardware isolation and software that can run on customer-owned infrastructure. The $2 million gives Voortman, Ciobanu and Kehrer room to turn those primitives into a dependable service. Their larger task is proving that developers will choose an independent machine layer when agent vendors and better-funded cloud providers can bundle similar capabilities into broader platforms.