Start for $20 — Run Your First Private AI Agent on Any Computer

You do not need a GPU. You do not need a dedicated machine. You need $20 and a computer you already own.

Most people land on the hardware page, see $3,000 GPU setups and "out of stock" badges, and assume private AI is out of reach. It is not. The hardware is the *end goal* — something to grow into once you know what you are building. This guide gets you running a real, private AI agent today, on whatever computer is in front of you, for the cost of a streaming subscription.

A dedicated machine like the MacBook Neo or MacBook Pro M5 is a nice-to-have for keeping agent workloads off your personal laptop — but it is absolutely not required to start.

Let's Be Honest: This Setup Goes Against Our Whole Point

OpenClaw and Hermes Agent exist because we believe your AI should run on hardware you own and control — not on someone else's server. Sending your prompts to a cloud provider, even a privacy-focused one like Ollama Cloud, is exactly the kind of arrangement we built this community to move away from. Your data leaves your machine. You are trusting a third party's infrastructure, their uptime, their billing, and their policies. That is a real trade-off and we are not going to pretend otherwise.

So why are we writing this guide?

Because the hardware situation right now is genuinely difficult. The GPUs that make local inference practical — the cards that let you run a 70B model at a speed that is actually useful — are either out of stock, priced at a premium, or backordered for months. Waiting for hardware to become available should not stop you from learning how Hermes Agent works, building your first skills, understanding what SOUL.md does, or figuring out what kind of agent you actually want to run. The orchestration layer — the part that makes an agent useful — has nothing to do with GPU performance. It runs fine on any modern laptop.

Ollama Cloud is a temporary bridge, not a destination. It is the least bad option for getting started while the hardware catches up. The models are open-weight, the API is OpenAI-compatible (meaning your config migrates directly to a local setup when you are ready), and Ollama's stated policy is zero data retention. That does not make it equivalent to running locally — it just makes it the most honest cloud option we are aware of for this use case.

When your hardware arrives, you point Hermes at http://localhost:11434/v1 instead of https://api.ollama.com/v1, update the model name, and everything else stays the same. The skills you wrote, the memories you built, the SOUL.md you refined — all of it transfers. The cloud subscription becomes optional the day your GPU is in the machine.

Build now. Run local when you can.

What You Are Building

By the end of this guide you will have Hermes Agent running on your computer, connected to a private cloud inference endpoint; Ollama Cloud handling the actual model inference — no GPU required on your end, no data logging, no training on your prompts; Claude Code installed as your build environment for writing skills and customizing your agent (optional but recommended); and a working first task to verify everything is connected.

This is not a toy demo. Hermes Agent is a production-grade autonomous agent with tool use, persistent memory, a skill system, and six terminal backends. Ollama Cloud runs the same open-weight models you would run locally — the only difference is the GPU is theirs, not yours.

What You Need

| Requirement | Details | |---|---| | Any computer | macOS, Linux, or Windows (via WSL2). No GPU needed. | | Git installed | git --version to verify. Install from git-scm.com if missing. | | Ollama Cloud account | Free to sign up. Pro plan is $20/month or $200/year. | | 15 minutes | The installer handles everything else automatically. |

Step 1 — Sign Up for Ollama Cloud ($20/mo)

Go to ollama.com and create an account. The Free plan gives you light cloud usage to test the setup. The Pro plan at $20/month (or $200/year) gives you 50x more cloud usage — enough for day-to-day agent work, coding automation, and deep research sessions.

A few things worth knowing about Ollama Cloud before you hand over a credit card: usage is measured by GPU time, not tokens — there is no fixed token cap that cuts you off mid-task; your prompt and response data is never logged or trained on; models run on NVIDIA Cloud Provider infrastructure with zero data retention; and session limits reset every 5 hours, with weekly limits resetting every 7 days.

Once you have an account, go to your account settings and generate an API key. You will need this in Step 3.

Step 2 — Install Hermes Agent

If you already have Ollama installed, the fastest path is the new official one-command integration:

ollama launch hermes

Alternative: standalone installer (use this if you are not running Ollama locally — for example, if you are connecting to Ollama Cloud only):

Open a terminal and run:

curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash

The standalone installer handles everything automatically: Python 3.11, Node.js, ripgrep, ffmpeg, the repository clone, a virtual environment, and the global hermes command. The only prerequisite is git.

When the installer finishes, reload your shell:

source ~/.bashrc   # or: source ~/.zshrc on macOS

Verify the install worked:

hermes version
hermes doctor

hermes doctor will tell you exactly what is configured and what is missing. At this point it will flag that no LLM provider is set — that is what Step 3 fixes.

Step 3 — Connect Hermes to Ollama Cloud

Ollama Cloud uses an OpenAI-compatible API, which means Hermes can connect to it as a custom endpoint. Run the model setup wizard:

hermes model

When prompted, select "Custom endpoint (self-hosted / VLLM / etc.)" and enter the following:

• URL: https://api.ollama.com/v1
• API key: your Ollama Cloud API key from Step 1
• Model name: qwen3.5:35b (recommended starting point)

Hermes saves this to ~/.hermes/config.yaml. You can also edit it directly:

model:
  default: qwen3.5:35b
  provider: custom
  base_url: https://api.ollama.com/v1
  api_key: YOUR_OLLAMA_API_KEY
  context_length: 32768

Why set `context_length: 32768`? Ollama defaults to a very low context window (4,096 tokens on hardware with less than 24GB VRAM). For agent use with tool schemas and conversation history, you need at least 16k–32k. Setting this explicitly ensures Hermes does not run out of context mid-task.

Choosing a Model

| Model | Best For | Size | |---|---|---| | qwen3.5:35b | General agent work, coding, research | 35B | | qwen3.5:14b | Faster responses, lighter tasks | 14B | | qwen3-coder-next | Code-heavy tasks, writing skills | varies | | minimax-m2.7 | Agentic workflows, longer tasks | large | | glm-5.1 | Engineering tasks, tool use | varies |

Start with qwen3.5:35b. If responses feel slow, drop to qwen3.5:14b. You can switch models at any time with /model inside a session or hermes model outside one.

Step 4 — Run Your First Task

Start Hermes:

hermes

Try a simple task to confirm the connection is working:

`` What tools do you have available? List them briefly. ``

Hermes will respond with its tool inventory — file system access, terminal execution, web search, memory, and more. If you see the tool list, everything is connected correctly.

Now try something that uses a tool:

`` Search the web for the latest Hermes Agent release notes and summarize what changed. ``

Hermes will use its web tool, fetch the information, and summarize it. You are running a private AI agent with no data leaving your control.

Step 5 — Migrate from OpenClaw (If Applicable)

If you are already using OpenClaw, Hermes can import your configuration automatically:

hermes claw migrate

This migrates your SOUL.md (agent identity), memories, skills, command allowlist, messaging settings, API keys, and TTS assets. Run hermes claw migrate --dry-run first to preview exactly what will be moved.

The setup wizard (hermes setup) also auto-detects ~/.openclaw and offers migration before you configure anything else.

Step 6 — Install Claude Code (Optional but Recommended)

Claude Code is the fastest way to go from "Hermes is running" to "Hermes is doing something useful for me." It is a terminal-based coding assistant that reads your entire project, writes code across multiple files, and runs commands — making it the natural tool for writing Hermes skills, editing SOUL.md, and debugging agent configurations.

Install it:

curl -fsSL https://claude.ai/install.sh | bash

Then open your Hermes configuration directory and start customizing:

cd ~/.hermes
claude

From here you can ask Claude Code to write a new skill, modify your SOUL.md persona, or add a new tool configuration. Claude Code requires a Claude Pro subscription ($20/month) — it is a separate cost from Ollama Cloud, but the two tools serve different roles: Ollama Cloud is your agent runtime, Claude Code is your build environment.

The MacBook Neo: The Cheapest Dedicated Host — Cloud Only

Running Hermes on your personal laptop works fine for getting started. The reason people eventually want a dedicated machine is straightforward: agent tasks can run for minutes or hours, they consume CPU and memory in the background, and having them on your primary machine means competing with everything else you are doing.

The MacBook Neo ($599) is the cheapest Mac you can use as a dedicated Hermes Agent host. It runs macOS, it is completely fanless and silent, and its 16-hour battery means it can sit on your desk running agent tasks around the clock without issue. For the use case in this guide — routing requests to Ollama Cloud, ChatGPT, or Anthropic's Claude API — it is entirely sufficient.

What the MacBook Neo cannot do: run local inference models. It uses the A18 Pro chip — the same chip as the iPhone 16 Pro, not the M-series found in the MacBook Pro or Mac Studio. With 8GB of unified memory, it cannot load any LLM locally at a useful size. If you connect it to Hermes and point Hermes at http://localhost:11434/v1, you will get errors or unusably slow responses. The MacBook Neo is a cloud-inference host, not a local inference machine. That is not a criticism — it is just what it is, and for this guide's use case it is the right tool at the right price.

If you eventually want to run local models — 7B to 14B comfortably, or 30B–70B with the right config — the MacBook Pro M5 Pro (starting at $2,199 with 24GB unified memory) is the next step up. That is a different guide. For now, the MacBook Neo gets you a dedicated, silent, always-on agent host for $599.

When you are ready to explore the full hardware range, the [hardware page](/hardware) covers everything from the MacBook Neo all the way up to multi-GPU inference servers.

What's Next

Once Hermes is running and you have completed a few tasks, the natural next steps are:

1. Write your first skill — skills are Markdown files in ~/.hermes/skills/ that give Hermes persistent, reusable knowledge. Claude Code makes this straightforward. 2. Edit your SOUL.md — customize the agent's identity, working style, and default behaviors to match how you actually want it to operate. 3. Explore the skill ecosystem — the community maintains a growing library of skills for common tasks. Run hermes skills to browse what is available. 4. Upgrade to local hardware — when you want to run larger models, reduce latency, or eliminate the cloud dependency entirely, the [hardware page](/hardware) is your next stop.

The $20/month Ollama Cloud subscription is not a permanent ceiling — it is a starting point that lets you learn what you actually need before committing to hardware.