Agent setups
Claude Code, Codex, Hermes or OMP on your machine, asking before it sends, spends or deletes.
Agent SetupAI that runs on machines you own.
I am Tanner Osterkamp. I set up AI agents and open models on your hardware, and I build apps.
- Nodes
- 2 × GB10
- Link
- 200 Gb/s
- Context
- 262,144
$ homelab status 2026-10-05 22:18 PDT
NODECHIPMEMORYGPUTEMP
sparkNVIDIA GB10121 GB96 %81 C
ascentNVIDIA GB10121 GB96 %84 C
workstationRTX 408016 GB1 %48 C
SERVING
modelglm-5.3-flash (open weights)
context262,144 tokens
enginevLLM, tensor parallel across spark + ascent
linkspark <-> ascent, 200 Gb/s
stateup running 4 waiting 5 kv cache 77 %
answered/v1/models in 28 ms
My own machines are the test bench.
Two NVIDIA GB10 machines, a DGX Spark and an ASUS Ascent GX10, joined by a 200 Gb/s link, serve one open model to every computer at home. An RTX 4080 workstation sits beside them.
- Engine
- vLLM, tensor parallelOne model split across both GB10 machines.
- Model
- Open weightsglm-5.3-flash with a 262,144-token context.
- Used by
- Four agent harnessesClaude Code, Codex, Hermes and OMP do real work against it every day.
Three kinds of work, each with a fixed price.
Each one has a written scope before I start. The prices are on the packages page.
Models at home
Open models on hardware you own, tested on your tasks, so your prompts stay on your network.
Homelab AIWorkflows and apps
One real workflow automated on your accounts, or a web or iPhone app from written scope to launch.
Agent Workflow and App BuildYour hardware. Your models. Your data.
This is the shape of a homelab setup. The model runs on a machine you own, your agents use it, and a person approves anything that sends, spends or deletes.
- You keep the keys
I work on your machines and accounts. You keep the keys, the code and the hardware.
- Cloud is a choice
Nothing you type leaves the machine unless you choose a cloud model.
- Written down
Agent and homelab setups end with a written runbook, so someone else can run them.
Things I built that you can open today.
See the work page for how each one is built and what it runs on.
Protocol Guide
Source-cited EMS protocol search. Every result names the protocol book it came from.
protocol-guide.com
MeshVault
A private AI server you own, with Hermes Desktop and an iPhone app.
meshvault.aiMy homelab
Two GB10 machines linked at 200 Gb/s serve an open model to every computer at home.
How it is builtThe Protocol Guide screen is a sample search. Education and reference only.
Written scope first. Your machines, your keys.
Talk it through
A short note or a strategy call that ends with a written plan.
Fix the scope
A fixed price and a written scope before work starts.
Build on your side
Set up on your machines and accounts, with a person approving anything irreversible.
Hand it over
A walkthrough, a written runbook and email support afterwards.
Plain answers, from running it myself.
All posts on AI agent setup, local models, MCP servers and what they cost.
How to stop an AI agent sending or deleting
Five controls to set before an agent touches your inbox, files or payment accounts, with the order to set them in.
ReadWhat is an MCP server, and who needs one?
What the Model Context Protocol does, how it differs from a plain API, and two live examples you can open.
ReadClaude Code, Codex, Hermes or OMP: which first?
What each agent harness is built for, a short decision guide, and why most people need only one.
ReadHave something in mind?
Write to me at tanner@meshvault.ai. Fixed prices are on the packages page.