On Tuesday, a Y Combinator-backed company called InsForge launched Instabox with a simple pitch: AI agents need somewhere to run code, and waiting around for cloud infrastructure defeats the purpose. The product promises to provision a fully loaded Linux machine in under a second, complete with 4 virtual CPUs, 8GB of RAM, and persistent storage.
Hang Huang, the startup's CEO, announced an invite-only beta on X that gives early users up to five machines at no cost. The timing reflects a broader scramble among infrastructure providers to solve what developers describe as an emerging bottleneck. When an AI agent needs to execute code in isolation, browse the web, or juggle long-running tasks, spinning up a traditional virtual private server can take minutes. That lag compounds quickly in agentic workflows where speed determines whether the system feels responsive or sluggish.
"Stop using your laptop / Mac mini / VPS," Huang wrote in the October 14 launch announcement. Each agent gets its own Linux environment, accessible through a command-line shell or a graphical desktop interface. InsForge describes the infrastructure as "designed from the ground up for agents," though the six-person team has shared few details about the underlying architecture.
The headline metric is that sub-second spin-up time. Instabox claims users land "in a new machine's shell in under a second," a number aimed squarely at cutting latency. An independent tester from RecodeX clocked a machine reaching "Running" status in roughly 27 seconds on October 15, though the tester noted the measurement might not align with the vendor's claim depending on how "ready" is defined. The gap between marketing and real-world performance is familiar territory in cloud infrastructure, where benchmarks often hinge on fine print.
What Instabox does deliver, according to documentation and early tests, is a set of features tailored for programmatic control. Snapshots preserve machine state for cloning or rollback. A built-in secrets vault injects environment variables at boot, avoiding the need to hard-code credentials. Public HTTPS URLs let agents expose local ports without manual firewall wrangling. A pause-and-resume function, the documentation says, "brings back a paused machine's memory and processes," allowing an agent to pick up mid-task. Files and packages persist across reboots, a detail validated in the RecodeX test that confirmed a separate /data volume maintained file integrity through stop-start cycles.
Setup runs through a single command: npx instabox@latest. That signs users in via browser and installs the CLI, along with agent "skills" deposited into directories like ~/.claude/skills, according to the quick-start guide. The TypeScript SDK exposes lifecycle methods for creating, pausing, resuming, and deleting machines, plus file operations, snapshot handling, and desktop control for browser automation. Code samples show machines tagged with metadata for programmatic lookup, with errors surfaced through typed exceptions like InstaboxAuthError.
The built-in desktop is where things get more visual. Agents can fire up a browser or other graphical tools, and humans can "take control" when needed before handing execution back to the agent. SSH and mosh access are also documented for developers who prefer more traditional remote workflows, though the company clearly expects most interaction to happen through the SDK or CLI.
Pricing remains unresolved. "Free until we figure it out," the pricing page says, with a promise of "the most competitive pricing, with a generous free tier" once the beta ends. No timeline for paid tiers has been disclosed. Early adopters in the invite program receive 20 vCPU and 40GB of RAM at no charge, Huang's post indicated, though neither duration nor usage caps were specified. Default quotas include the five-machine limit and around 10 snapshots per account, according to third-party summaries from launch week and a limits reference in the documentation. No GPU instances surfaced in independent testing on October 15, and no GPU SKUs appeared in the public docs as of October 16.
InsForge is far from alone in this space. InstaVM advertises Firecracker-based microVMs that boot sandboxes "in under 200ms," per documentation. Instaclaw and similar services position "your agent's computer" with always-on VMs, typically 2-4 cores and 4-8GB of RAM, emphasizing filesystem persistence. Other services also target the agent infrastructure layer, and third-party comparisons published around the Instabox launch note that Runloop and Freestyle offer different SKU options with published pricing. Third-party observers have speculated that Instabox uses Firecracker microVMs, the isolation technology AWS open-sourced, though the company has not confirmed that architecture in any first-party materials.
The company graduated from Y Combinator's Spring 2026 cohort. CEO Hang Huang and CTO Tony Chang lead the team, according to the YC directory. YC partner Andrew Miklas appears on Crunchbase's investor list, but funding amounts remain undisclosed. InsForge's positioning hinges on the claim that it is "the first cloud infrastructure platform designed from the ground up for agents," language that invites skepticism given how many startups are chasing the same market.
Instabox launched on October 14 with documentation, an SDK, and a CLI already in place. Early access remains gated by invite, and the company has not announced when it will open general availability or publish paid tiers. Whether the product can hold up under real-world load, or whether that sub-second promise translates to practical advantage, remains to be seen as the beta expands and competitors respond.
