Show HN: AgentVM – Safe, Sandboxed Linux VM for OpenClaw and AI Agents

Show HN: AgentVM – Safe, Sandboxed Linux VM for OpenClaw and AI Agents(agentvm.deepclause.ai)

3 points by phunterlau 101 days ago | 2 comments

# Show HN: AgentVM – Safe, Sandboxed Linux VM for OpenClaw & AI Agents

Hi HN,

We built AgentVM (WASM + Alpine Linux) because Docker is too heavy for per-session AI agents, and `exec()` by an OpenClaw agent on the host is too dangerous.

It runs a full Alpine Linux VM inside a Node.js Worker Thread. It boots in milliseconds, is completely isolated, and supports Python, networking, and standard Linux commands out of the box, perfect for local AI agents like OpenClaw and others.

Website: [https://agentvm.deepclause.ai/]

Repo: [https://github.com/deepclause/agentvm]

Core Value

* Instant Boot: No container orchestration delay.

* Sandboxed: Runs in a worker thread. If it crashes or gets malicious, your host is safe.

* Full Environment: Real Linux userspace. Install packages, run Python scripts, grep files.

* Agent Ready: Drop-in integration for *OpenClaw* and Vercel AI SDK.

Why we are sharing this:

We want this to be the standard "execution unit" for local AI agents.

We need your feedback on:

1. The security model (WASM/Worker isolation).

2. API ergonomics for your agent frameworks.

PRs and issues are very welcome! We'd love to see what you build with it.

verdverm 101 days ago |

> Docker is too heavy for per-session

Why do you believe this to be true?

I don't agree, I use containers via Dagger

> Full network access

Not really a sandbox if the agent can make POST or GET requests to exfiltrate

schmuhblaster 101 days ago | |

Network can be closed down. It also uses a userland network stack, so future iterations might include being able to define rules for ingress and egress.