Build a Claude Code Command Center in Obsidian

Most people use Claude Code inside Obsidian to track markdown files and stop there. That's barely scratching the surface. You can turn Obsidian into a command center for your entire Claude OS — a system with a fleet of skills and automations, a memory layer, a customizable visual dashboard, and fully local voice control. I've gone through several iterations of this setup, and the current one combines the best of my old web-based build with the power of Obsidian. Here's exactly how it works, and how you can build it yourself.

The hard part isn't making it look cool. It's making it actually provide value instead of turning into a zero-substance visual spectacle. So I'm going to break it into the three parts that matter: the visual and voice layer, the skills-and-automations backbone, and the Obsidian memory layer.

What Does an Obsidian Claude OS Actually Buy You?

Here's the thing — if this setup doesn't give you something the terminal and the Claude desktop app can't, it's pointless. So that's the bar.

The command center is a custom Obsidian plugin that Claude Code built for me, and it's 100% customizable. Whatever metric you want to see, you put it there. Mine shows my token usage, my social media metrics across every channel, my schedule pulled from Google Calendar, and my morning headlines. There are buttons to run automations and skills with a single click, plus tabs that go deeper into the reports I run daily.

For example, my research tab is fed by a morning intel automation that surfaces GitHub trending repos over the last 7 and 30 days, outlier YouTube videos, content angles, and a running look at what's hot on Hacker News. My audience tab shows how my content is performing across YouTube — what's working, what isn't, and why.

And because it's Obsidian, I'm not boxed in. I can pull up a terminal inside Obsidian, and I can drive the whole thing with voice.

How Does Local Voice Control Work with Claude Code?

The voice layer was inspired by ChatGPT and Codex voice mode, and it runs 100% locally on my GPU. I can hit a hotkey and ask a question even when I'm tabbed out of Obsidian entirely — working in the Claude desktop app, doing something else, doesn't matter. That was the whole point: audibly interacting with Claude Code the way you would with GPT live mode, because the voice mode inside the Claude desktop app just isn't there yet.

When I ask something like "what was the biggest AI news story today?" the response comes back fast — because it's not searching the web live. It's been wired to read reports I've already run, like my research report, and answer from those.

Here's the pipeline behind it:

1. You speak. The voice recording goes to the voice module (call it Jarvis, call it whatever).
2. Transcription. faster-whisper, an open-source repo running locally, turns your voice into text.
3. Routing. That text goes to Haiku 4.5 — the smallest, cheapest, fastest model — which decides where the request goes.
4. Execution. The task runs.
5. Speech. kokoro, another open-source tool, turns the text response back into voice.

I use Haiku for routing because it's a nice standard I can hand to anyone regardless of their hardware. You can swap Haiku for any local model if you have the GPU for it — the whole thing can be made fully local.

What Are the Three Tiers of Voice Routing?

The router breaks every request into one of three tiers. This is what keeps fast things fast and lets slow things run without blocking you.

• Tier 1 — Skills. "Run the morning intel report." This routes straight to a skill. No extra reasoning, no added words — just execute and report back when done.
• Tier 2 — Metrics. "What's today's biggest AI news?" This just reads a report that already exists. I deliberately keep this tier from doing any searching or research on its own, so it stays fast — find what exists and tell me.
• Tier 3 — Headless Claude Code. "Pull up Fable 5, do deep research on X, Y, and Z, and come up with a plan." That's not a skill and not a metric, so it spins up a full headless instance of Claude Code, runs everything, and reports when finished. This is the catch-all — the most powerful and the slowest.

The key design decision is separating these tiers. Quick tasks that pull from information you've already gathered should never trigger a full research run.

Why Is the Skills and Automations Backbone the Real Substance?

The dashboard is the wrapper. The skills and automations are the substance. Without them, the visual layer is just a pretty shell with no information flowing through it.

The concept is simple, but very few people do it correctly. You take everything you do day-to-day and week-to-week and turn it into skills, then decide which skills deserve to become automations. I break my work into domains — memory, productivity, research, content, community, agency, sales. Inside research alone, I have a YouTube pipeline, deep-research skills, a light RAG database, a morning trend scan, and competitor tracking.

The workflow to build these is straightforward:

1. Define the tasks. What do you actually do that can be codified? AI is non-deterministic by default — tell it to do one thing ten times and it'll do it ten different ways. When you specify how you want something done and what the end state looks like, you can turn it into a skill and make it as deterministic as possible.
2. Find the tasks three ways. First, give Claude Code a stream of consciousness — turn on your mic and yap about what you do every day and week, then ask it which parts can become skills. Second, have it read your logs. Every tool call, command, and message you make in Claude Code is logged on your computer, so Claude Code can review the last 30, 60, or 90 days and pull real skills out of your actual behavior. Third — and best — combine both. Your mental model of what you do plus the reality in the logs is where the strongest skill list comes from.
3. Promote skills to automations. If something needs to run every day at a set time, make it a routine that's one command away. Best practice: keep it as a skill first for a while, confirm the outputs are what you want, then turn it into a full automation that just runs on its own.

You can go even further with loop-engineering principles — comparing an automation's output against its past runs and building a self-improving system with a goal and success criteria. That's more advanced, but it's where this can go. It's all built on the skills. Can you codify what you do every day into a skill? Yes — and it's simpler than it looks.

Once the skills exist, integrating them into the command center is trivial. The voice mode already knows what skills exist because Haiku has a list of everything in the background. And the skills form the basis of what the dashboard even shows — the schedule, the morning headlines, the social metrics, the research section are all just outputs and data from automations. Yours will look different, and that's the point.

How Does the Obsidian Memory Layer Actually Help Claude Code?

This is the part people get confused about. The Obsidian memory layer is not a graph RAG, and it's not supercharging Claude Code's memory. The cool-looking graph is for you, the human.

What Obsidian actually does is organize information in markdown files in a way that makes sense. The tangential benefit for Claude Code kicks in when you have thousands of files and you've structured the vault well: it gives Claude Code a clear map to find information, which returns accurate answers faster and burns fewer tokens. That's the whole benefit — not more memory, a better map.

The system everyone saw came from a Karpathy tweet in April that pulled nearly 22 million views. The template is simple: a vault folder (just a folder on your computer) with three subfolders:

• raw — raw source material. Ask Claude to research RAG systems and the sources and raw notes land here.
• wiki — synthesized, Wikipedia-style articles built from the raw data. A RAG-systems subfolder might hold articles on vector databases and chunking strategies.
• output — the actual deliverables. Turn those wiki articles into a slide deck and it goes here.

Raw information, synthesized data, and the deliverables made from that synthesized data. But the folders aren't the magic. The magic is that every level has a master index — an index markdown file that acts as a table of contents. Go into the wiki folder and its index tells you the subfolders. Go into RAG systems and its index tells you the articles inside.

Do you need a table of contents for three subfolders? No. But what about 3,000 files? 30,000? 3 million? A map gives Claude Code a clear path, which means quicker answers and fewer tokens burned.

Do You Have to Copy Karpathy's Exact Structure?

No. You need a structure that makes sense for you. Mine is a bit more convoluted — folders for content, daily notes, inbox, ops, projects, systems, and wiki. So how do you still get the benefits of a map with a custom layout?

You spell it out in your claude.md. The biggest job of the claude.md in your vault is to lay out the vault structure and how Claude should navigate it — the navigation pattern plus the structure is the map and the instructions to read it.

And if all of this feels like too much, you can literally point Claude at your vault, paste in the Karpathy tweet, and say "set this up in accordance with the principles here."

How Do You Build the Custom Dashboard UI?

The dashboard is a plugin Claude Code creates and installs into Obsidian on its own — mine shows up under community plugins as the Chase Command Center. One plugin you definitely want enabled first is hot reload. You won't find it in the official Obsidian community plugin browser, but it's on GitHub — search "hot reload," point Claude Code at it, and it'll install it for you.

From there, have a conversation with Claude Code about building a custom Obsidian plugin that acts as a command center, and spell out what you want it to show. My process:

1. Mock it up in Claude design. Grab inspiration from Pinterest — search "dashboard" or "command center" — and drop the screenshots in. Say what metrics you want (calendar, skills, token burn, whatever) and ask for five very distinct variations on one page.
2. Double down on the winner. Pick the direction you like and generate multiple variations of that. I landed on a glassmorphism look with clouds showing through the back, a subtle star-twinkle effect, and a little motion — numbers ticking up when I click into an overview.
3. Ship the zip to Claude Code. Claude design bundles all the mock-up code into a zip. Drop it into Claude Code and say "build the plugin based on this." It builds, installs, and you'll see it in Obsidian.

Want it even more hands-off? Enable computer use inside Claude Code and let it run the design iterations while screenshotting itself.

Frequently Asked Questions

Do I need a technical background to build a Claude OS in Obsidian?

No. The command center is a custom Obsidian plugin that Claude Code builds and installs itself — you drive it through conversation. The hardest part is deciding what you want to see and codifying your own workflows into skills, and even that can start as a stream-of-consciousness voice dump to Claude Code.

Is the voice mode actually 100% local?

The transcription (faster-whisper) and text-to-speech (kokoro) run locally on your GPU. The routing uses Haiku 4.5 by default because it's a reliable standard across different hardware, but you can swap it for a local model of your choice to make the entire pipeline local.

Does the Obsidian memory layer make Claude Code smarter?

Not in the way people assume. It's not a graph RAG and it doesn't add memory. It organizes your files with index-file tables of contents so Claude Code has a clear navigation map — which produces faster, more accurate answers and burns fewer tokens as your vault grows.

What's the difference between a skill and an automation?

A skill is a codified task you trigger manually — by button, command, or voice. An automation is a skill promoted to run on its own, usually on a schedule. Best practice is to keep something as a skill until you're happy with the outputs, then turn it into an automation.

How do I figure out which workflows to turn into skills?

Use three inputs: tell Claude Code what you think you do (stream of consciousness), have it read your Claude Code logs from the last 30-90 days to see what you actually do, and combine the two. The overlap between your mental model and your real behavior is where the best skills come from.

---

If you want to go deeper into building your own Claude OS, join the free Chase AI community for templates, prompts, and live breakdowns. And if you're serious about building with AI, check out the paid community, Chase AI+, for hands-on guidance on how to make money with AI.