This website uses cookies

Read our Privacy policy and Terms of use for more information.

// AI Lessons

Run Your Own Local AI Agent

Stand up a personal agent like OpenClaw or Hermes on your own machine in an afternoon — so the agent, its memory, and its keys belong to you, not a vendor.

A few weeks ago I wrote that we are renting this AI boom instead of owning it, with every prompt, file, and memory living on somebody else's servers. Then I went looking at what my "assistant" actually knew about me, and where it kept that knowledge. The answer was: a lot, and none of it on a machine I control. I spent two decades in open source precisely so no vendor could hold my stack hostage — and here I was handing the most personal software I use, the thing that reads my email and remembers my projects, to a subscription I don't own.

So I did the obvious thing. I stood up my own agent on a little box on my desk. It took an afternoon. This is that lesson.

Run one personal agent on your own machine before you trust a hosted one with your life.

// The Takeaway: You can now self-host a capable personal AI agent — OpenClaw or Nous Research's Hermes Agent, both MIT-licensed — on a $5 VPS or a $599 Mac Mini, point it at a local model through Ollama so nothing leaves your network, and message it from Telegram or Signal like any other contact. It installs with a single command and runs for the price of the electricity. The catch is security, not difficulty: these agents get shell and file access, so you lock them down before you turn them loose. Own the agent, then constrain it.

Install your own agent and run your first local chat → About an afternoon. Free to run local, a few dollars a month for a rented box. Needs Ollama and a machine with 24GB+ of RAM.

For two years, "using AI" has meant renting a seat on someone else's model, with your context and history sitting in their account. That was the only option. It isn't anymore. A pair of open source agents crossed the line from hobby project to genuinely useful this spring, and both are built to run on hardware you own.

They point in slightly different directions. OpenClaw is the one that does things — it lives in your chat apps and can run commands, control a browser, read and write files, and manage your calendar from a text message. Hermes Agent is the one that grows with you — its whole pitch is a learning loop that turns repeated tasks into reusable skills and keeps three layers of memory across sessions. Pick OpenClaw if you want an assistant that acts today. Pick Hermes if you want one that compounds over months.

// The real shift: You stop renting an assistant and start owning one. The model, the memory of who you are, and the keys to your accounts all sit on a machine you can unplug. It is the same move open source gave us in the 2000s — you never needed to own the expensive proprietary stack, you needed to control your own. The difference this time is that the software in question knows your calendar, your inbox, and your projects. Ownership stops being ideology and becomes basic hygiene.

Know what you are trading before you install.

Honest limits first, because these are new and still rough. OpenClaw has had a hard security year: researchers cataloged six CVEs in 2026, including a one-click remote-code-execution flaw that worked even on a localhost-bound instance, and Koi Security found 341 malicious skills in OpenClaw's ClawHub marketplace, a number that later grew past 800. Hermes is quieter but greener: its headline memory feature is off by default, and users regularly report it silently failing to persist until it is configured correctly. Neither is a drop-in replacement for a coding assistant like Claude Code. Self-host for control and privacy, not because it is less work.

The afternoon setup.

This is the OpenClaw path because it is the most turnkey. The Hermes variation is one section below.

  1. Pick a box. A Mac Mini M4 with 24GB of RAM ($599) runs 13B-parameter models through Ollama with the agent on top, draws about 10–15 watts idle — roughly $15 a year in electricity — and sits silently on a shelf. No hardware to spare? A $5–10 per month Linux VPS works too, you will just point at a cloud model instead of a local one.

  2. Install Ollama and pull a model. Run the Ollama installer, then pull a capable local model. For a local agent, plan on a 64k-token context window, which is why the RAM matters.

  3. Launch the agent with one command. Type ollama launch openclaw. Ollama installs OpenClaw if it is missing, shows the security notice, lets you pick your model, and starts the gateway for you. On the first run, it explains the tool-access risks — read that screen, do not skip it.

  4. Point it at your local model. Set the model with openclaw models set ollama/<model>, or put it in the config, so your agent runs against the model on your machine, and nothing is sent to an outside API.

  5. Connect one channel. Use the gateway wizard to link a single messaging app — Telegram or Signal is the easiest start — so you can text your agent from your phone. Stop at one channel until you trust it.

  6. Lock it down before you give it real work. Keep the gateway bound to localhost or behind a VPN and never expose it to the open internet — researchers found more than 42,000 exposed instances in the wild. Install skills only from sources you have vetted. Then hand it one real job, like "summarize the files in this folder," and confirm it works end to end.

Bonus: make it a Hermes agent that remembers you.

Want the version that compounds? Install Hermes Agent instead, wire it to the same local Ollama endpoint, and — this is the step everyone misses — explicitly turn on memory in the config with memory_enabled and the user profile, because it ships disabled. Once it is on, Hermes keeps short-term, long-term, and skill memory, so the report you ask for every Monday quietly becomes a saved skill it runs better each time.

Why this matters now.

The same week I write this, Meta is turning its own supercomputers into a rental business, and the whole industry is racing to be your landlord. Owning the agent is the other side of that trade. You do not have to win the argument about open versus closed, and you do not have to give up frontier models for the hard 20% of your work. You just have to make sure the software that knows the most about you runs somewhere you can reach the power switch. Stand up one agent this weekend. Even if you go back to a hosted tool on Monday, you will negotiate with it differently once you know you don't have to.

Your AI Sherpa,

Mark R. Hinkle
Founding Publisher, The AIE Network
Follow me on LinkedIn

If you want to get in contact or give me feedback, reply to this email. I read every single one of them.

Reply

Avatar

or to participate

Keep Reading