Run your own host: security, networking, and clustering basics
If you have a capable box at home or in a lab, you can run it well yourself. This isn't a full threat model. It's the checklist we'd want a friend to work through before sharing hardware, and it builds on clustering and sharing models.
Lock the box down
- Keep it updated. Apply OS, driver, and firmware updates on a schedule. Most break-ins use old, known bugs.
- Close what you don't use. Use a default-deny firewall for incoming traffic, and open only the ports you actually need. Don't expose remote-admin or model-server ports to the internet.
- Use keys, not passwords, for remote access. Turn off password SSH, and use a VPN or overlay network instead of forwarding ports.
- Separate accounts. Your own login, admin access, and anything that runs other people's work should all use different accounts.
This guide is general host practice. Hypermesh's own isolation of renter workloads is separate and is not claimed here.
Keep renter work away from your stuff
Treat shared work as somebody else's code running on your machine:
- Run it in a container or VM with no access to your home directory, your keys, or your personal files.
- Put the host on its own network segment (a VLAN or a separate subnet) so it can't reach your NAS, cameras, or laptops.
- Don't put secrets you care about on the same disk without encryption.
- Watch for disk, memory, and temperature problems. A healthy box is part of what renters pay for.
Plan the network
For one box, ordinary gigabit Ethernet and a stable uplink are fine for most inference. When you split one model across machines, the links between those machines become the bottleneck:
- Data path vs control path. Give the heavy traffic between nodes its own fast link. Keep management, updates, and discovery on your normal LAN.
- 10GbE is a sensible step up for small clusters and costs a reasonable amount now.
- 25/100GbE and RDMA-capable NICs matter when you split tensors across nodes and need low latency. They add cost and setup work, so only buy them when your workload needs them.
- On Apple silicon, Thunderbolt 5 links between Macs can carry this traffic directly. See our clustered memory survey.
Some cluster tools are explicitly not hardened. For example, llama.cpp warns that its RPC backend should never run on an open network. Keep that traffic on a private segment.
Grow into a cluster
Start with one box that is stable and patched. Add a second node over a dedicated link, test that a model splits across both and recovers when a node restarts, and only then add more. Write down your setup. Future you, and any renters, will appreciate it.
Try it
Hypermesh connects hosts who have enrolled, checked hardware with renters who need time or tokens on it. Renters pay in USDC (or buy USDC through Stripe’s onramp), and hosts are paid through Stripe Connect. When you're ready to share, the hosts page walks you through enrollment, which is a single install flow. Weekly bank payouts start once you reach the $5 minimum.
Sources: llama.cpp, “RPC backend README”. This is general guidance, not a security audit of your setup.