OpenClaw vs Cowork vs Obsidian + Claude Code — Which Setup Works?

I run OpenClaw on a VPS, use Claude Cowork on my Mac, and have an Obsidian vault wired up to Claude Code. Every week, someone asks me which one they should use. The answer depends on what you actually want to do with it.

This isn't a feature matrix. I'm not going to score them out of ten and declare a winner. Each one does something the others can't, and each one has a limitation that will annoy you eventually. I know because all three have annoyed me.

So I'll just tell you what each one actually does well, and where each one has let me down.

OpenClaw: the one that never sleeps

OpenClaw is the AI agent that lives in your messaging apps. Telegram, WhatsApp, Signal. You send it a message, it does the thing. It doesn't need you to open a laptop or click anything. It's persistent, proactive, and — once you get the config right — surprisingly capable.

I'm currently using it to build niche coaching funnels and a small SaaS product. Not just prototyping. Actual building. The first funnel is focused on AI training for property management — specifically small landlords in the UK. I messaged my OpenClaw agent with the brief and it built a full marketing landing page and a functional demo app. Not a wireframe. A working thing I could show to someone. It also drafted the content for the page, which needed editing but was a real starting point, not the generic padding you get from ChatGPT.

It did all of that on GPT-5.4. And I need to talk about GPT-5.4 for a second because it's hard to explain how good it is inside OpenClaw. I've routed tasks through Gemini, Kimi K2, and various other models. GPT-5.4 just seems to understand the OpenClaw architecture better than anything else. It knows the file structure, it knows the skill system, it writes config that works. I don't have a technical explanation for why. It just does.

Johann Sathianathen posted his workspace structure after three weeks with OpenClaw and it looked nothing like the default install. That tracks with my experience. Day one is SOUL.md and IDENTITY.md. Week three is a custom workspace with specialised skills, agent coordination, and model routing that stops you overpaying for simple tasks.

The thing most people get wrong is running OpenClaw in its default sandbox mode and wondering why it feels limited. You need to allow exec and enable multiple functions. You probably want elevated mode turned on for your own user. That's where it goes from "interesting chatbot" to "actual assistant." But that comes with a trade-off: elevated mode means the agent can run commands directly on your gateway host. /elevated full skips approval prompts entirely. You're trusting the AI not to do something stupid with real access to your machine.

The token question

This is the thing nobody talks about enough. OpenClaw is token-heavy. Every interaction, every skill execution, every background task — it all burns tokens. And the model you route through determines the cost.

OpenAI lets you use Codex OAuth, which means you can run OpenClaw on your existing OpenAI subscription without paying separately per million tokens. That's a big deal when your agent is running tasks all day. Anthropic's approach is less clear. Paying per million tokens on Opus 4.6 gets expensive fast if you're using it as your primary model. I've found myself routing most OpenClaw tasks through OpenAI for exactly this reason, even though I use Claude for almost everything else.

Alex Lieberman interviewed Nate Liason about his agent Felix, which runs an entire business. X account, Stripe, bank account. That's the ceiling. Most people don't need that. But the fact that the ceiling is that high tells you what this architecture can do.

OpenClaw is the right tool when you need always-on automation. Team use via group chats. Messaging-native workflows. Persistent memory across sessions. Something that runs while you sleep.

The trade-offs are real. The config is fiddly. The security model is basically "trust yourself not to misconfigure it." There's no sandbox in elevated mode. Browser use is weak if you're running it on a VPS — I've got Appify set up but the GUI refreshes every time I reconnect, so I can't hold a persistent Chrome session. And if something goes wrong, it goes wrong on your actual machine. I wrote a configuration guide that covers the things I wish someone had told me.

Cowork: the one with guardrails

Claude Cowork is the opposite philosophy. It's a sandboxed desktop tool. One person, working in a folder, with browser access. Anthropic runs it inside a lightweight Linux VM on your machine. It can read your files, write code, build documents, browse the web. But it can't escape the sandbox. It can't touch your system files. It can't send an email on its own without asking you first.

That's the point. The security model is "the AI can't wreck anything even if it hallucinates."

For individual productivity work — writing, research, building documents, analysing data — Cowork is genuinely strong. I use it for content production, SEO work, and building out the site. The scheduled tasks feature means you can set it to run things automatically. I have a research funnel that runs at 6am every morning, pulling AI news and compiling a rundown. The output is actually solid. Genuinely useful briefing.

But here's what I've hit up against.

The automation side of that same 6am task falls apart. I wanted it to commit the research files to my git repo — partly because I have an OpenClaw agent sharing the same repo, and I was trying to get the best of both tools working together. Cowork can't do it. The git auth doesn't persist inside the sandbox. I also had it creating email drafts to send me the briefing. It could create the drafts in Gmail, but couldn't actually send them. So I ended up with a pile of unsent drafts accumulating in my inbox. I've switched that part of the scheduled job off since.

The other limitation: scheduled tasks only run while your Mac is awake and the desktop app is open. Close the lid and Cowork doesn't run. It's not a server. It's a desktop app pretending to be an agent.

Simon Høiberg described OpenClaw running his life from the airport. Cowork can't do that. It needs you to be at your desk.

Where Cowork does something the others can't is browser use. It can log into Google Analytics, Google Search Console, keyword planning tools, GA4 — and actually work with what it sees. My SEO experience is limited. But using Cowork to browse these tools and act on what it finds, I've generated over 1,000 clicks of organic traffic to the site by just focusing on blog content. For someone who didn't know what a SERP was a year ago, that's not nothing.

Cowork is the right tool when security matters more than autonomy. File-based work, documents, spreadsheets, browsing the web safely. Anything where you want the AI to produce something but you don't want to worry about what it might accidentally do.

The limits are real though. No persistent background execution. Scheduled tasks need the machine awake. Limited integration with authenticated services like git, email, and external APIs. Single-user only. No team or group chat use case. You get safety. You give up autonomy.

Obsidian + Claude Code: the one I'm still figuring out

I'm going to be honest about this one. The idea is great. The reality, for me, is still early.

Ronin posted a video of building a "JARVIS" with Obsidian + Claude Code in an hour and it hit 88K views because people want this: your personal knowledge base, wired up to an AI that can actually use it.

The concept is simple. Your Obsidian vault is your second brain — notes, journals, projects, meeting notes, ideas. Claude Code gets access to the vault and can search, create, organise, and extract information from it. You write a CLAUDE.md file that tells the agent how your vault is structured. Then you talk to it.

The real power is context. OpenClaw and Cowork work with what you give them in the moment. The Obsidian setup starts with everything you've ever written down. You ask it to prep for a meeting and it pulls from your meeting notes, project files, and daily logs without you telling it where to look.

I set up my vault in the same repo as my Cowork folder and OpenClaw workspace. The idea was to use it for managing drafts — this site has a lot of them. But I stopped going to it. I'm not sure yet whether that's because the setup wasn't right or because the habit hasn't formed. The tool works. I just haven't made it part of my workflow yet.

Where I'm headed instead: repurposing one of my OpenClaw agents as a chief of staff. Sitting on top of all my files and folders, finding patterns, surfacing things I've forgotten about, and executing tasks. Maybe with calendar access, or a sandboxed version of it. That's closer to what I actually want from a "second brain" — something that acts on the information, not just searches it.

Dave Drach wrote about building a personal operating system with this stack, and the privacy angle is real. Everything stays local. The vault is on your machine. The AI processes it without sending your notes to a third-party agent framework.

The Obsidian + Claude Code setup wins on one thing nothing else touches: deep context from your own notes. Privacy is fully local. It's the only tool that can answer "what did I think about X six months ago?"

But it's not an agent. It doesn't run in the background. It doesn't live in your Telegram. It can't take action on your behalf unless you're sitting in front of it. It's also the most technical to set up. You need Claude Code installed, an MCP plugin configured, and a CLAUDE.md file that actually maps your vault structure. If you're not comfortable in a terminal, this isn't for you. And if you're like me and you want the AI to do things with the information rather than just retrieve it — you might end up building that on top of OpenClaw anyway.

So which one should you use?

It depends on what you need the AI to do.

If you want an AI that works while you don't — that monitors channels, responds to messages, handles tasks in the background, runs 24/7 — that's OpenClaw. Nothing else does this. The trade-off is security and config time. I offer a setup service for teams who want this without the pain.

If you want a safe assistant for desktop work — writing, research, data analysis, browsing your analytics tools — that's Cowork. The sandbox is a genuine advantage, not a limitation. For most individual productivity work, it's the right choice. I wrote about the Cowork setup if you want to see what it can do.

If you want an AI that knows everything you know — that can search your notes, connect ideas across months of writing, and help you think — that's the Obsidian + Claude Code setup. It's the most niche of the three. I'm not yet convinced it beats a well-configured OpenClaw for my use case, but for people who live in their notes, it might be exactly right.

I use all three. They don't compete with each other. OpenClaw handles the things I shouldn't have to be present for. Cowork handles the things I need to produce. Obsidian + Claude Code handles the things I need to think about.

If you're not sure which setup makes sense for what you're trying to do, message me. I'll tell you honestly — even if the answer is "just use ChatGPT, you don't need any of this."