Andrew Gazelka's résumé includes a stint at xAI—confirmed on Y Combinator's website—where he built high-performance systems in Rust—the kind of low-level engineering work that rarely makes headlines but keeps AI infrastructure humming. His new project is different. It's not another chatbot wrapper or enterprise analytics dashboard.
Instead, Gazelka is tackling something that sounds deceptively simple: What if you could version control an entire virtual machine the same way developers version control code?
That question led to Indexable, a Y Combinator-backed company (Spring 2026 batch) that's attempting to make VM sandboxes behave like git repositories. Fork a running virtual machine in milliseconds. Snapshot everything—files, active processes, memory, even database states. Roll back to any earlier point in the commit graph. It's version control, but for entire computing environments rather than just source files.
The timing is no accident. As AI agents grow more sophisticated, they increasingly need infrastructure that can rewind, not merely restart.
Milliseconds, Not Minutes
At the heart of Indexable's offering is a product called "ix," which models sandboxes as immutable commit graphs. Every change generates a commit automatically. Each one captures the full state of the VM at that moment: disk contents, process memory, the works. Branches point to the latest tips. Fork from any commit, and the system gives you copy-on-write behavior across the entire state.
"We wrote our own VMM on top of KVM," the team noted in a recent job posting. "It can fork a running sandbox in milliseconds."
That speed matters—a lot—for what Gazelka sees as the primary use case. AI agents often need to explore multiple solution paths simultaneously. With ix, an agent can fork its environment, run an experiment, then either merge the outcome or discard the branch entirely. No need to restart from scratch or lose context. The API exposes functions like ix.spawn, ix.fork, ix.exec, with a planned ix.eval that would enable cached execution and syscall tracing.
It's the kind of infrastructure that didn't exist because, until recently, nobody really needed it.
More Than Filesystem Tricks
Most sandbox providers focus on fast cold starts or filesystem snapshots. Modal, for instance, added memory snapshots to their offering earlier this year. Vercel published engineering notes about optimizing filesystem snapshots. Daytona and E2B both have forking APIs aimed at sub-200-millisecond startup times for code execution workloads.
Indexable's pitch goes further, though how much further remains somewhat unclear. According to their Y Combinator profile, they promise "full-environment forks and snapshots across files, processes, memory, and databases." That means not just the filesystem, but running process trees, in-memory state, and database content—all captured atomically. The system's documentation describes pre- and post-commit states for individual commands, theoretically enabling surgical rollbacks and replay.
Gazelka's background shows up in the implementation. The documentation describes content-addressed storage for deduplication, a custom Linux kernel build, and a global /nix store with packages compiled specifically for AMD EPYC Gen 5 processors. VMs can boot any OCI or Docker image. Shell sessions persist across disconnects, exposed as raw PTYs that can render via WebAssembly in browsers using libghostty—a detail that feels both clever and slightly overengineered.
Whether all this technical sophistication translates into a sustainable business remains an open question.
A Crowded Field, Suddenly
The competitive landscape has gotten noisy fast. OpenAI added native sandbox execution to their Agents SDK in mid-April, including what they described as a "more capable model-native harness." Their roadmap mentioned support for "more sandbox providers," a signal that demand for agent runtimes is outpacing OpenAI's willingness to own the entire stack.
Recent research papers have surfaced similar themes. One, titled "DeltaBox," highlighted the need for millisecond-level checkpoint and rollback for stateful agents. Another, "Sandlock," explored lightweight Linux process sandboxes with granular confinement models. Academic interest doesn't always predict commercial viability, but it suggests the problem space is real.
Beyond Modal and E2B, newer entrants include Sandbox0 (open-source with a managed tier), Podflare (hardware-isolated with an emphasis on latency), and AgentSphere (positioning itself as agent-native infrastructure). Community discussions on Reddit and Hacker News recently have centered on the distinction between "untrusted code execution" sandboxes and "computer-use" agents—with recurring emphasis on snapshot, pause, resume, and fork capabilities.
Indexable is entering a market that didn't quite exist 18 months ago and now has at least half a dozen serious players. That's either validation or a red flag, depending on how differentiated the technology actually is.
What It's Built For
The documentation frames three primary use cases, though with varying degrees of specificity. First: agent evaluation at scale. Fork a production-like VM, run variations against identical starting states, compare results, roll back. Straightforward enough.
Second: long-running coding or data agents that need persistent shells, process trees, and databases with recoverable checkpoints. This feels like the real target—agents that aren't just executing scripts but managing stateful, multi-step workflows over hours or days.
Third, more nebulously: "anti-captcha scenarios" where agents route outbound traffic via residential proxies to sidestep rate limits. This use case appears in the docs but isn't elaborated on, and it raises the obvious question of what else might fall under the umbrella of evasion-oriented tooling.
Indexable currently operates data centers in Hillsboro, Oregon and Silicon Valley, with US East flagged as planned. SDKs exist for TypeScript (Node and browser via WASM), Python, Rust, Swift, plus a CLI that installs via a curl-to-shell command. The Python SDK went live on PyPI recently under the package name "ix-sdk," maintained by Gazelka's GitHub account.
Pricing? "Too early to publish fixed pricing," according to the documentation. That's honest, at least. The status page showed a roughly eight-hour outage for long-running VMs in late May, though API uptime held above 99.9% over the prior week. For a YC company still in formation, that's not disastrous—but it's also not reassuring.
The Road Ahead
Planned features include cached execution with syscall tracing and a content-addressed global cache—infrastructure that would let agents avoid redundant work across invocations. There's also talk of three-way filesystem merging with extended attributes for conflict resolution, which sounds ambitious and potentially fragile.
The team is three people in San Francisco, hiring for what they describe as a "founding engineer" role. The job posting asks for someone "anti-tech-debt obsessed" with expertise spanning Rust, KVM, networking, and distributed systems. That's a narrow candidate pool, even in the Bay Area.
Third-party write-ups have mentioned a future compliance roadmap—SOC 2, HIPAA, ISO 27001—though Indexable hasn't confirmed any of this publicly. Cross-host snapshot replication shows up in aggregator summaries as part of an enterprise track, but again, not in first-party materials. It's unclear whether these are real roadmap items or speculative projections.
For now, the bet is simple: As AI agents grow more capable, they'll need infrastructure that can rewind time, not just restart it. Whether that's a $100 million market or a feature that larger players eventually absorb is anyone's guess.
Gazelka and his co-founders are placing their chips on the former. The rest of us will find out soon enough.
