May 2026 Summaries

boxd’s comparison argues that both boxd and exe.dev provide persistent, SSH-accessible Linux virtual machines with built-in services such as DNS, HTTPS, reverse proxies, persistent storage, and long-lived state, making them suitable for developers and AI agents. It characterizes exe.dev as stronger for individuals seeking a polished, well-documented experience and predictable $20-per-month personal pricing, supported by established search visibility and onboarding resources. Boxd positions itself for agent-oriented, bursty, or regulated workloads through claimed under-200-millisecond copy-on-write forks of live VM state, sub-millisecond hibernation resume, usage-based sleep-to-zero billing, a JSON CLI and gRPC/SDK API for programmatic VM control, larger default disk capacity, and self-hosting through a Rust binary on KVM. The comparison notes that pricing and feature information reflect public documentation and internal benchmarks at specified 2026 dates, and frames the choice as dependent on whether users prioritize personal-use simplicity or automation, parallel experimentation, sovereignty, and infrastructure control.

Self-hosted agent execution refers to running the environments where coding agents read files, edit code, execute commands, and use tools on infrastructure controlled by the customer, while allowing the underlying AI model to remain a hosted API or be deployed separately. The approach is presented as relevant for organizations with regulatory, intellectual-property, data-residency, audit, network-isolation, identity-management, and blast-radius concerns, particularly EU-based teams navigating DSGVO and Schrems II implications. Effective platforms are described as providing isolated persistent workspaces for each developer, rapid provisioning from reusable templates, reproducible environments, detailed activity logs, and low idle costs. The piece promotes boxd as an implementation of this model: a single Rust binary running KVM-isolated virtual machines with SSH-key identity, template-based forks, hibernation, stable HTTPS endpoints, and no external managed control plane. It argues that genuine jurisdictional control depends on where the execution platform and its data path operate rather than simply using a vendor’s EU cloud region, while also acknowledging that hosted services may be more practical for small teams, nonsensitive workloads, or organizations unable to operate their own platform.

Agent harnesses, the orchestration layers that connect AI models to tools and long-running work loops, require production infrastructure that preserves state, resumes quickly, supports branching from live execution, maintains stable network addresses, and survives crashes. The article argues that common options such as ephemeral sandboxes, containers, serverless functions, Codespaces, and long-lived VPSs address only parts of these needs, often requiring developers to build complex checkpointing, recovery, and networking mechanisms themselves. It presents persistent microVMs as a better-suited runtime model for coding agents and similar stateful workloads, highlighting boxd as a KVM-based platform that preserves processes and disk state, hibernates when idle, resumes in under a millisecond, forks running machines through copy-on-write snapshots, and provides each VM with a public HTTPS endpoint. The piece acknowledges that ephemeral platforms remain appropriate for short stateless code execution, high-volume inference, and GPU training, but contends that long-running, branching agent workflows benefit from persistent infrastructure; it identifies exe.dev and Sprites as comparable products and notes that boxd can also be self-hosted.

Cloud development environments are described as entering a renewed phase in 2026 as AI agents increasingly require persistent remote machines for long-running code execution, state retention, experimentation, and collaboration with human developers. The category has evolved from browser-based tools such as GitHub Codespaces, aimed at rapid onboarding and standardized setups, to self-hosted container platforms such as Coder and Gitpod for organizations needing infrastructure control, and toward persistent VM or agent-oriented platforms such as Daytona, Sprites, and boxd. Key capabilities now emphasized include persistent storage and processes, fast wake-up and workspace forking, stable network identities, and SSH access that supports both developers and agents. The text argues that teams should choose platforms according to their operational needs, scale, compliance requirements, preferred infrastructure model, and whether their work depends primarily on browser IDEs, Kubernetes-based environments, or persistent virtual machines. It presents boxd as an SSH-first, self-hostable KVM-based option designed to give humans and AI agents access to the same persistent, forkable workspace.

Anthropic’s computer-use capability enables Claude to operate graphical interfaces by analyzing screenshots and issuing mouse and keyboard actions, but its practical challenge is hosting a desktop that remains fast, persistent, recoverable, and accessible over time. The piece argues that serverless platforms, ordinary containers, Codespaces, local laptops, and always-on cloud instances each lack some combination of a real graphical environment, low-latency screenshots, durable session state, crash recovery, or stable remote access. It presents persistent virtual machines as a better fit, particularly for multi-hour agent tasks that require browsers, logins, files, and desktop state to survive across turns and interruptions. The author promotes boxd as a persistent microVM platform offering hibernation for low idle costs, rapid resume, machine-state forks for parallel experimentation or rollback, public HTTPS domains, and KVM isolation, with an example setup using Anthropic’s Docker-based computer-use demo. The article distinguishes this approach from browser automation tools such as Playwright and Puppeteer, which it considers more appropriate for high-volume, stateless web scraping, while positioning persistent desktops for longer-running AI agents operating arbitrary GUI applications.

Sprites and boxd are presented as persistent-microVM platforms for AI agents, sharing the view that agents need isolated, resumable compute rather than disposable sandboxes, with sleep-to-zero billing, checkpointing, networking, and HTTPS access as core capabilities. Sprites is Fly.io’s managed offering, built on Firecracker and Fly’s established infrastructure, and is positioned as stronger in hosted operational maturity, Claude and native MCP support, broad language SDKs, developer marketing reach, and granular usage-based billing. Boxd, whose team authored the comparison, emphasizes a self-hostable KVM-based architecture distributed as a Rust binary, open-source ambitions, European jurisdiction control, fewer infrastructure dependencies, sub-millisecond hibernation wakeups, and copy-on-write forks of live VM state in under 200 milliseconds. The comparison distinguishes Sprites’ checkpoint-and-restore approach, reportedly taking about a second to restore a running environment, from boxd’s faster wake and memory-based branching features, while noting that the products target different operational preferences rather than fundamentally different architectural goals. It recommends Sprites for developers seeking a polished, Claude-first managed service on Fly.io, and boxd for organizations requiring self-hosting, compliance and data-sovereignty control, or rapid parallel agent execution, while cautioning that the claims reflect one competitor’s assessment and published information current in 2026.

Model Context Protocol (MCP) servers expose tools, resources, and prompts to AI clients and commonly operate as long-lived, stateful, bursty, credential-bearing services that must be reliably reachable by one or more agents. The article outlines four hosting approaches: local stdio processes for individual development, servers bundled into clients for simple stateless integrations, edge or serverless deployments for independent HTTP-based calls, and persistent virtual machines for shared, stateful production workloads. It argues that production deployments require durable storage and processes, stable DNS and TLS, observability, concurrency handling, secure authentication and isolation, low idle costs, and self-hosting options for sensitive or regulated environments. The author presents persistent microVMs as a strong fit for these needs and promotes boxd as a self-hostable platform offering persistent VMs, per-VM HTTPS endpoints, hibernation during inactivity, tenant isolation, and rapid VM forks for testing. While local and serverless hosting remain appropriate for scratch tools and genuinely stateless API wrappers, the article concludes that most teams running production agents will use a mix of patterns and generally need at least one persistent host for credentialed, multi-agent MCP services.