A project called SmolVM surfaced on Hacker News yesterday with a straightforward pitch: give coding agents and OpenClaw proper microVM isolation without the weight of Docker. Built by @theaniketmaurya at CelestoAI, it's already picking up interest from people running parallel agent workflows.
What SmolVM Does
SmolVM is an open-source abstraction layer over microVMs, purpose-built for creating sandboxes for coding agents. The pitch is simple: when you're running multiple OpenClaw Pi agents in parallel, you want each one isolated, fast to start, and as lightweight as possible. Docker gets the job done but brings overhead — image layers, daemon startup, container networking — that adds up fast when you're spinning up dozens of sandboxes.
MicroVMs (think Firecracker, the tech behind AWS Lambda) boot in milliseconds, use kernel-level isolation, and have a much smaller attack surface than a full container runtime. SmolVM wraps this in a simple CLI aimed directly at agent workflows.
The OpenClaw Angle
OpenClaw already supports sandboxed execution for agentic runtimes — the coding-agent skill spawns subprocesses and containers to let agents work without touching your host system. SmolVM offers a tighter alternative: true VM isolation per agent, at near-container startup speeds.
For anyone running a local OpenClaw setup with Pi agents doing parallel coding tasks, SmolVM could meaningfully reduce overhead while improving isolation. Each Pi agent gets its own microVM boundary — no shared process space, no Docker layer caching concerns, no daemon to keep healthy.
The ability to run parallel Pi agents on a local sandbox is the headline use case. If you've ever had Pi agents clobber each other's working directories or leave behind conflicting state, VM-level isolation solves that cleanly.
Getting Started
Install SmolVM with a single command:
curl -sSL https://celesto.ai/install.sh | bash
Then start a Pi agent sandbox:
smolvm pi start
That's it. The abstraction handles the microVM lifecycle, mounts, and networking. You get an isolated Pi agent environment without configuring anything at the hypervisor level.
Why It Matters
The tension between isolation and overhead is real for anyone running agent-heavy workflows. Docker became the default because it's good enough and well-understood. But "good enough" has a cost:
- Docker daemon adds ~100MB+ memory baseline before you run anything
- Container startup involves layer extraction and network setup
- Shared kernel namespaces mean containers aren't true VM-level isolation
- Running 10+ parallel agent containers on a laptop starts to feel it
MicroVMs flip the equation: VM-level isolation, ~50ms boot time, minimal memory footprint per instance. The tradeoff is less ecosystem tooling — you can't just docker pull an image — but for purpose-built agent sandboxes, SmolVM's curated environment is enough.
For OpenClaw users specifically, the smolvm pi start workflow maps cleanly onto the existing Pi agent model. You're not changing how you write tasks or prompts — you're just changing what the sandbox underneath looks like.
Community and Status
SmolVM is early-stage — the GitHub repo is active and the HN thread generated solid discussion about the microVM-vs-Docker tradeoff. @theaniketmaurya has been responsive to questions about roadmap and integration.
If you're running parallel OpenClaw agents and hitting the edges of what Docker handles comfortably, it's worth a look.
- GitHub: CelestoAI/SmolVM
- HN Thread: news.ycombinator.com/item?id=47992937
- Install:
curl -sSL https://celesto.ai/install.sh | bash