machine0 – persistent VMs for AI agents

TL;DR

TL;DR: machine0 is a CLI that provisions persistent cloud VMs for AI agents — from a small instance to 60 vCPU/240 GB plus 8xH100 GPUs — billed by the minute, driven entirely by the agent via CLI or MCP.

Source and Accuracy Notes

⚠️ This section is MANDATORY. All links must be verified from actual source, not guessed.

• Project page: machine0.io — verified on 2026-08-22
• Source repository: No public GitHub repo — this is a proprietary CLI product
• YC S26 launch: news.ycombinator.com/item?id=49348136 — verified batch S26
• YC company page: ycombinator.com/companies/machine0 — verified Summer 2026 batch
• HN object ID: 49348136

What Is machine0?

machine0 is a CLI tool that gives AI agents their own persistent cloud virtual machines. The core command:

curl -LsSf https://machine0.io/install.sh | sh

One command spins up a full VM with SSH access, a static IP, and an HTTPS endpoint at <vm>.mac0.io. The VM stays alive (99.99% uptime) until you explicitly tear it down — no laptop required.

Agents drive it themselves via the CLI or through an MCP server. A human or orchestrator sets up a profile (bundles of MCP connections, credentials, prompts, and environment variables), then hands the CLI to the agent. The agent manages its own fleet: creates machines, snapshots them, clones them, suspends idle ones, and tears them down when done.

The product description on the landing page reads:

“Up to 60 vCPUs, 240 GB RAM and GPUs. Designed for agents using CLI or MCP. Billed by the minute.”

The Problem with Ephemeral Agent Compute

The founder, Barnaby Malet (YC W20, previously CTO of Upflow), describes the pain in the YC launch post:

• Resources — a few agents working on a large codebase saturate your laptop’s RAM and CPU. Auto-research agents need GPUs you do not have.
• Security — running —yolo on your personal machine is one prompt injection away from exfiltrated credentials.
• Availability — close your laptop and the agent dies mid-task.
• Isolation — there is no clean separation between everything on your machine and what the agent actually needs.

How It Works

Installation

curl -LsSf https://machine0.io/install.sh | sh

Creating a VM

machine0 new mybox --profile default

This creates an SSH-ready VM with a static IP and HTTPS endpoint. The VM is always-on at 99.99% uptime until you switch it off.

MCP Integration

An MCP server ships with the CLI. Hand it to Claude, Codex, or OpenCode and the agent can manage its own fleet without reading documentation:

# Agent creates a VM for a build task
machine0 new buildbox --profile claude-coding

# Agent SSHes in and runs its task
machine0 ssh buildbox

# Agent snapshots the configured machine
machine0 snapshot buildbox

# Agent clones 10 identical workers from the snapshot
machine0 clone buildbox --count 10

# Agent suspends an idle machine (stops billing)
machine0 suspend buildbox

# Agent resumes when needed
machine0 resume buildbox

Profiles

Profiles bundle MCP connections, credentials, prompts, and environment variables. They are injected at VM creation. Each agent gets exactly the capabilities you choose — and nothing else.

OS Options

NixOS and Ubuntu are available as first-class options. NixOS provides declarative, reproducible machines; Ubuntu provides a familiar server environment.

Practical Evaluation Checklist

• [ ] Install the CLI: curl -LsSf https://machine0.io/install.sh | sh
• [ ] Create a small VM: machine0 new testbox --profile small
• [ ] SSH into it: machine0 ssh testbox
• [ ] Suspend and resume: machine0 suspend testbox && machine0 resume testbox
• [ ] Set up an MCP profile for your agent
• [ ] Tear down when done: machine0 delete testbox

Pricing

Billed by the minute. CPU-only plans from $0.013/hr (1 vCPU / 1 GB). GPU plans start at $0.836/hr (1x RTX 4000 Ada) and scale up to $39.336/hr (8x H200).

Key tiers from the pricing page:

| Size | vCPU | RAM | Disk | Hourly | Monthly (est.) | |------|------|-----|------|--------|---------------| | xs | 1 | 1 GB | 10 GB | $0.013 | $9 | | small | 2 | 4 GB | 50 GB | $0.052 | $38 | | medium | 4 | 8 GB | 100 GB | $0.104 | $76 | | xl | 8 | 16 GB | 320 GB | $0.208 | $152 | | 6xl | 60 | 240 GB | 900 GB | $3.714 | $2,711 |

GPU tiers include RTX 4000 Ada, L40S, RTX 6000 Ada, MI300X, H100, H200, and 8-GPU clusters. GPU sizes are available in US East, UK, EU, and Asia.

Security Notes

• VMs provide isolation between agent workloads and your local machine
• Profiles allow fine-grained credential scoping per agent
• Static IPs and HTTPS endpoints give each agent a dedicated network identity
• The founder’s background at Upflow (handling sensitive accounting data for companies like Postman and ProductBoard) informs the security design

FAQ

Q: Is this open source? A: No. machine0 is a proprietary CLI product. There is no public source repository.

Q: What happens to my VM if I close my laptop? A: Nothing — the VM runs in the cloud, not on your machine. It stays up at 99.99% uptime until you suspend or delete it.

Q: Can I use this without an AI agent? A: Yes. You can interact with machine0 entirely from your own terminal. The CLI is designed for humans and agents alike.

Q: What OS images are available? A: NixOS and Ubuntu. NixOS is a first-class citizen for declarative, reproducible machine configurations.

Q: How does billing work for suspended VMs? A: Suspended VMs do not bill for compute. You only pay for storage while a VM is suspended.

Conclusion

machine0 targets a specific and growing pain: running AI agents that need persistent, isolated, and powerful compute — without tying that compute to your laptop. The CLI-first design means both humans and agents can use it without context-switching. With per-minute billing, profiles for capability scoping, and MCP support for agent self-service, it is purpose-built for fleet management of autonomous agents.

If you are running coding agents, research agents, or agent swarms that need more than a laptop can offer, machine0 is worth evaluating. The entry price of $0.013/hr makes it accessible to individual developers, while the top-tier GPU configurations support serious workloads.

HN launch thread: news.ycombinator.com/item?id=49348136