How to build your own Agent OS.

"Do I need OpenClaw if I have Hermes? Or is Hermes just a better OpenClaw?"

Karim — great question, because the framing matters. Honest answer: they do different jobs. Run both if you want, start with Hermes if you have to choose.

Here's how I think about it:

• Hermes is the agent harness — kanban-aware, tool-using, skill-extensible, runs goals autonomously. Better for jobs with defined outputs (SEO posts, videos, research, outreach, scheduled tasks).
• OpenClaw is local-first and ambient — lives across WhatsApp, Slack, Discord, Telegram. Better as the assistant that's just there when you need to ask something or capture a thought.

They're not replacements for each other — they're complementary surfaces. Same vault, same dashboard, different access patterns.

That said: if you're brand new to building your Agent OS and you have to pick one to start, start with Hermes. Get the harness running, learn how skills work, build your first goal. Once Hermes is solid for you, add OpenClaw later for the always-on ambient surface. Setup is much smoother in that order than trying to wire both together from scratch.

Full Hermes setup is in The Goldie Hermes Agent OS →. Full OpenClaw walk-through is in The Goldie OpenClaw 5.20 →. Both live next to each other in The 7-Layer Blueprint →.

"Now that I can communicate with Hermes via Claude, should I still prompt Claude Code directly?"

Nate — you've hit the moment where the OS starts paying off. Honest answer: use Claude Code less, use the OS more, but keep Claude Code for one specific thing.

Here's the rule of thumb I run by:

• Inside the OS (talking to Hermes via Claude): for everything that has a defined output — write me an SEO post, build me a landing page, generate a video script, refactor this file. Hermes orchestrates the right agents, pulls from your vault, saves the output to Workspace. You get the result in your dashboard, end-to-end.
• Direct Claude Code: for deep exploratory coding sessions where you're learning a new codebase, debugging something tricky, or pair-programming through a hard refactor. Claude Code is more "two of us in a room together"; Hermes-via-Claude-in-OS is more "set the goal, get the output."

So your build-out goes through Hermes inside the OS. That's the right move. Faster, better organised, every output saved to Workspace + auto-logged to your Obsidian vault. You're not losing anything by stopping the direct Claude Code prompts for the build — you're upgrading from a single-thread chat to an orchestrated system.

Claude Code stays in your toolkit for the days you want to deep-dive on one specific codebase question. Most days, you won't open it.

"I set up Agent OS on my Nvidia Spark. How do I add a local LLM?"

Ken — the Nvidia Spark setup is sweet. Adding a local LLM is two parts: install the model server, then point Agent OS at it.

Step 1 — Install Ollama (or LM Studio) and pull a model.

• Install ollama on your Spark machine: curl -fsSL https://ollama.com/install.sh | sh
• Pull a model that fits your VRAM. On the Spark you've got plenty of headroom — try ollama pull qwen2.5-coder:32b or ollama pull llama3.3:70b if you want max quality.
• Start the Ollama server: ollama serve (usually auto-starts on install).
• Confirm it's running: curl http://localhost:11434/api/tags — should return your model list.

Step 2 — Wire Ollama into Agent OS via the openclaw.json or hermes config.

For OpenClaw, edit ~/.openclaw/openclaw.json and add Ollama as a provider:

"providers": {
  "ollama": {
    "baseUrl": "http://localhost:11434/v1",
    "api": "openai-completions",
    "models": [{ "id": "qwen2.5-coder:32b", "name": "qwen2.5-coder" }]
  }
}

For Hermes, run hermes config set model.provider ollama followed by hermes config set model.endpoint http://localhost:11434/v1.

The reason the Model tab only shows cloud options is that the local provider lives in your config file, not the UI's default catalog. Once you add it to the config, the tab will show your local model alongside the cloud ones — and you can pick which to use per chat or per goal.

Heads-up: local models work great for "fast and free" — Bonsai 8B (what Brad's running), Qwen2.5-Coder, Llama 3.3. But for hard reasoning and tool use at the frontier, you'll still want to fall back to qwen3.7-max or Claude. Best to wire in both — local for routine, frontier for demanding. That's the Build-Anything Stack → default.

"Is there a free AI agent that builds high-quality video using my laptop hardware?"

Arpiet — yes, and we've already built it inside the Agent OS. It's the Hermes Video Agent, and it runs locally for $0 per render.

How it works:

• HyperFrames CLI — open-source tool that turns HTML compositions into MP4 videos. Renders deterministically on your local machine using your CPU + GPU. No cloud render farm. No per-minute fee.
• Hermes drives the HyperFrames CLI — you give Hermes a prompt ("build me a 60-second cinematic intro about [topic]"), it scaffolds the HTML composition + queues the render + saves the MP4 to your Workspace tab.
• AI Avatar layer (optional) — for talking-head videos, you set up your avatar clone once + your voice clone once. Every render after that uses them.

Output quality depends on your hardware — on a base M-series Mac mini, you'll happily render 1080p at deterministic frame rates. On an Nvidia Spark or similar dedicated GPU, you'll fly. On a 5-year-old laptop, it'll work but slowly.

Full walk-through with the five-station forge (Spark → Form → Frame → Forge → Filing) is in The Goldie Vision Forge →. The video agent is one of the most-used surfaces in Agent OS for a reason — once you've shipped one render at $0, you stop thinking about HeyGen or Synthesia or any of the per-render SaaS tools.

"How does the Obsidian 3-file system work for storing context across LLMs?"

Lawrence — great instinct. The 3-file system is the cleanest way to make your Obsidian vault feed every LLM in your dashboard without duplication. Here's how I structure mine:

File 1 — personal-profile.md · the "who I am" file

• Your name + business + role + voice
• Your customers + their language
• Your goals for the quarter + year
• Your non-negotiables (don't say X, always say Y)

Every agent reads this at the start of any session. It's what makes every output sound like you, not generic AI.

File 2 — skills-index.md · the "what each agent does" file

• One row per skill (SEO writer, video script, outreach drafter, content multiplier, etc.)
• For each skill: what it does, when to use it, which agent calls it, where the output lands
• Links to the individual skill.md files in ~/.hermes/skills/

This is the "operating manual" for your own dashboard. When an agent needs to know which skill applies to a request, this is the first thing it reads.

File 3 — knowledge-base.md · the "everything else" file (or folder)

• Your real numbers (MRR, subs, member counts, win stories)
• Your case studies + testimonials
• Your brand voice doc + writing rhythm rules
• Your past content links + topic positioning

For most operators, this one is a folder with sub-files rather than a single document — but it works the same way. Agents query it for grounded specifics whenever they need real numbers or examples.

The key move: link the three files at the top of every agent's system prompt. Hermes, Claude, OpenClaw, Codex — they all read the same three files. Same context. Different agent. Same brand voice every time.

And yes, Obsidian beats Notion for this because Obsidian is local-first (your files live on your machine, not in someone else's cloud) and uses plain Markdown (every AI tool reads it natively, no API needed).

Full pattern is in The Goldie Brain Loop →. The new Knowledge Studio → guide shows how the Notebook tab inside Agent OS auto-syncs to the same vault for completeness.