September 2026 Summaries

Boxd introduces a Linux system call, `import_cow_range`, to rapidly clone a running virtual machine’s guest memory into an independently running VMM process using fork-style copy-on-write semantics. Rather than copying gigabytes of RAM, writing snapshots to disk, forking and then losing memory through `exec()`, or handling page faults in userspace via `userfaultfd`, the new mechanism shares the parent’s memory pages while copying only page-table entries and write-protecting both mappings. When either machine modifies a shared page, the kernel handles the fault and creates a private page copy, while untouched pages remain shared. The design uses pidfds to avoid process-ID reuse races, requires guest RAM at the same fixed virtual address in both processes, and relies on existing ptrace permission checks. Cloning requires the source VM to remain alive and briefly paused while its mapping is imported, but can create a running machine in under 200 milliseconds without disk I/O or initial RAM copying, enabling live machines to serve as branchable templates for parallel tasks and experiments.

Boxd has raised €2 million in pre-seed funding led by BlueYard Capital, with support from Antler, OVNI, s20, Script Capital, and developer-infrastructure angels, to build a cloud platform for remote software development environments. The company argues that local development setups are poorly suited to AI coding agents and increasingly parallel workflows, creating limits around scalability, security, collaboration, and the ability to run multiple branches or tasks simultaneously. Boxd provides self-service remote environments designed to start in milliseconds, support sub-second forks of running virtual machines, offer hardware-level isolation and SSH access, and run either in Boxd’s cloud or on customer-controlled KVM infrastructure. Built from the hypervisor upward rather than on third-party sandboxes, the platform aims to make development environments fast, reproducible, shareable, and disposable. The funding will support improvements to startup and fork speeds, expansion of its self-serve offering, and development of a self-hosted product as the company positions its infrastructure for agent-driven software verification and execution workloads.

Boxd’s VM-forking system creates a byte-identical, already-running child machine in roughly a tenth of a second, but the restored guest initially retains invalid parent-specific state such as its IP address, hostname, random-number pool, clock calibration, and timer setup. Because all virtual CPUs resume halted, the system injects a non-maskable interrupt after placing a validated mailbox in guest memory, allowing a kernel handler to schedule repair work safely in normal execution context. The guest refreshes KVM clock parameters by rewriting an MSR, revives an AMD local APIC timer by rewriting its initial count register, and signals PID 1 so userspace can apply the child’s identity and configuration. Reseeding the kernel entropy pool with host-generated random bytes is prioritized to prevent separately forked children from producing identical tokens, keys, nonces, and UUIDs. Network settings and the hostname are then updated through raw netlink calls rather than cold-starting command-line tools, preserving the system’s sub-200-millisecond performance target while allowing the original workload to resume without observing the migration.

Boxd’s Tailscale integration allows organizations to move their machines, including development workspaces, preview environments, and agents, from public internet access into their existing private tailnet while retaining the same hostnames, TLS certificates, SSH access patterns, and subdomain routing. Organizations provide Boxd with a reusable Tailscale auth key, after which Boxd’s edge servers join the tailnet as persistent boxd-proxy devices and route HTTPS, SSH, and exposed ports only through private 100.x addresses. Public access is blocked with nonrevealing 404 responses, while Boxd’s management API, CLI, console, and SDKs remain accessible from anywhere. The implementation isolates each organization’s traffic through separate Tailscale nodes and origin-aware routing, while preserving standard ownership and sharing controls. Teams must approve Boxd proxy devices, configure ACL access, disable device key expiry, ensure users are connected to the tailnet, and account for limitations involving per-machine custom domains, expiring auth keys, and corporate DNS resolvers that may reject Tailscale address ranges.

Boxd operates persistent, fast-starting Linux virtual machines for remote development and lightweight services, and its infrastructure must balance zero trust toward guest software with useful application-level visibility while avoiding intrusive inspection of users’ files or processes. To automatically route each machine’s public URL to the active web service without requiring configuration or polling, Boxd patches its guest kernel to emit small signals before and after successful bind() and listen() calls, sending an eight-byte message through x86 PIO ports to the virtual machine monitor. The VMM converts these notifications into port events and selects a likely public service through a simple priority list of common web ports, ignoring interface-specific bindings and preventing later, lower-priority listeners from replacing a better choice. Although guests can technically forge these unprivileged signals, the system treats them only as hints: a false claim can at worst misroute that guest’s own URL, not grant access, alter firewall rules, or affect another tenant. Detected ports are stored in replicated cluster state, shared with proxies, and preserved across snapshots and forks, while the same guest-to-host signaling channel has since been extended to communicate readiness, exit codes, and file events.

Automations typically combine triggers, integrations, logic, and persistent state, but conventional cron jobs, webhook listeners, and polling loops keep machines running continuously even when work is infrequent, increasing idle costs and preventing suspension. Boxd proposes moving schedules and event subscriptions onto its platform, which wakes hibernated machines only when an alarm or webhook is due, allowing automation environments to retain their memory, installed dependencies, open processes, files, and agent context while billing primarily for active runtime or stored disk. Its platform-managed scheduler stores the next scheduled run, while platform-level event handling logs incoming webhooks and delivers them after waking the relevant machine. This model is presented as an alternative to serverless containers and workflow runners, which avoid idle costs but repeatedly rebuild environments and lose accumulated state. Boxd also integrates automation creation into persistent machines, where agents can generate TypeScript jobs, connect services such as Linear and Slack, test them against real data, and run them without a separate deployment target, while credentials remain outside the guest machine. The approach argues that externally managed triggers combined with millisecond-scale wake times could offer serverless-like economics alongside the continuity of a persistent server, particularly for stateful AI agents and automation workloads.