Mar 19, 2026 · 1h 4m · startup-ideas
I fixed OpenClaw so it actually works (full setup)
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Moritz Kremb joins Greg Isenberg on the Startup Ideas podcast to deliver a comprehensive, 10-step masterclass on configuring OpenClaw into a reliable, autonomous digital employee. The guide covers core architectural setup, memory persistence, fallback routing, security hardening, and real-world pipelines for content production and CRM management.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Greg holds 12% of the talking time here. How this is scored →
Greg as informed peer 3.6 Guest teaching 5.5 Guest disagreement 0.1 Greg pushing back 0.8
speaking balance: gold is Greg, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 26:34 Moritz Addressing Terms of Service Ambiguities
Moritz candidly acknowledges that running OAuth against Anthropic breaches official terms of service and bluntly advises using disposable accounts to survive bans.
Hardest push from Greg ▶ 25:56 Greg Challenging Anthropic Ban PolicyGreg directly confronts Moritz about recommending models that have actively banned OpenClaw users, asking why anyone should build on a restricted backend.
Biggest teaching moment ▶ 12:00 Context7 Documentation Project SolutionMoritz reveals the Context7 markdown repository ingestion method, directly solving Greg's recurring problem of hallucinated support advice and broken Reddit snippets.
Greg holds their own ▶ 7:43 Greg Framing OpenClaw Against Claude DispatchGreg cites Anthropic's Dispatch preview and challenges Moritz on how OpenClaw maintains an edge against rapidly advancing native desktop-to-mobile agent features.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Greg as informed peer | Guest teaching | Guest disagreement | Greg pushing back | Why |
|---|---|---|---|---|---|---|
| Comparing ChatGPT, Claude Code, and OpenClaw Paradigms | 2 | 6 | 0 | 0 | Moritz breaks down the structural evolution from ChatGPT cloud chats to Claude Code local files and OpenClaw autonomous cron loops. Greg remains in passive host mode as Moritz delivers the foundational architectural breakdown. | |
| OpenClaw vs. Proprietary Claude Cowork Systems | 6 | 4 | 1 | 3 | Greg displays strong domain awareness by referencing Anthropic's brand-new Dispatch research preview and pressing Moritz on why users should choose OpenClaw over native proprietary co-work tooling. Moritz explains the open-source customizability advantage. | |
| Step 1: Establishing a Robust Troubleshooting Baseline | 3 | 7 | 0 | 0 | Moritz reveals the Context7 documentation project trick to prevent hallucinated answers when troubleshooting OpenClaw errors. Greg acknowledges he had been struggling with hallucinated prompts and misleading Reddit threads. | |
| Step 2: Configuring Workspace Context and Persona Files | 4 | 5 | 0 | 0 | Moritz explains the workspace folder hierarchy including agents.md, soul.md, and user.md files. Greg reinforces the importance of structured persona files based on previous discussions. | |
| Step 3: Solving Memory Persistence with Auto-Compaction Flushing | 2 | 7 | 0 | 0 | Moritz details how to prevent memory loss during context compaction using compaction flush commands and automated heartbeat persistence logs. Greg listens attentively without interruption. | |
| Step 4: Model Selection, OAuth Setup, and Fallback Routing | 6 | 5 | 1 | 4 | Greg pushes back on OAuth recommendations by pointing out Anthropic's ban on third-party OpenClaw routing. Moritz clarifies the terms-of-service grey area and suggests burner accounts as safe mitigation. | |
| Step 5: Structuring Telegram Topic Channels with Custom Prompts | 2 | 5 | 0 | 0 | Moritz demonstrates how to configure granular Telegram group topics with scoped system prompts to prevent context pollution across distinct tasks. Greg absorbs the operational setup. | |
| Step 6: Navigating Browser Automation and Live Grocery Demo | 3 | 6 | 0 | 0 | Moritz demonstrates the differences between API web fetching, managed browser sessions, and Chrome relay while triggering a live automated grocery checkout. Greg chimes in with supportive banter. | |
| Step 7: Activating Bundled Skills and Auditing ClawHub | 5 | 4 | 0 | 2 | Moritz reviews bundled skills and the ClawHub marketplace. Greg intervenes with informed security context, noting independent analyses showing malicious payloads inside community-submitted skills. | |
| Step 8: Configuring Background Heartbeat and Cron Health Checks | 4 | 4 | 0 | 1 | Moritz demonstrates cron job self-healing routines embedded in heartbeat.md. Greg questions whether viewers should directly copy Moritz's configuration or construct their own from scratch. | |
| Steps 9 & 10: Security Foundations, Injection Defense, and Dedicated Accounts | 4 | 6 | 0 | 1 | Moritz presents prompt injection defenses, .env scoping, and local Mac execution versus cloud VPS vulnerabilities. Greg seeks concrete model recommendations capable of resisting injection attempts. | |
| Real-World Pipeline: Autonomous Short-Form Video Production | 3 | 7 | 0 | 0 | Moritz presents a complete seven-stage autonomous content production pipeline featuring scrapers, planning loops, and editor handoffs. Greg praises the rigorous process-driven architecture. | |
| Real-World Pipeline: Conversational AI CRM Management | 3 | 6 | 0 | 0 | Moritz walks through a conversational CRM integrated with Google Sheets, Gmail drafts, Calendar bookings, and WhatsApp outreach. Greg acknowledges the novelty and practical utility of the implementation. |
Statements from this episode (15)
Moritz Kremb: OpenClaw is the closest system to a truly autonomous agent
“And so it's kind of the first really personal agent that exists and also Currently, I would say it's the closest to what we have of a truly autonomous agent.”
Moritz Kremb: Claude Cowork is just Claude Code with a better UI
“So they kind of built cloud co-work and put that in the app with just a nicer UI. But under the hood, it's basically the same as Cloud Code.”
Moritz Kremb expects major AI providers to build proprietary OpenClaw clones
“And so what I do expect to happen is that it, like Claude code and Claude co-work and so on, they will all build like their own kind of versions of open claw.”
Moritz Kremb: Claude hallucinates OpenClaw solutions instead of checking official documentation
“Because normally Claude will, if you ask it something about OpenClaw, it's like, it can give you the right answer, but it often just makes something up. If it doesn't really know, it doesn't always go and check the docs by itself through its web search feature…”
Moritz Kremb: Preloading OpenClaw documentation into Claude fixed 99% of errors
“And since I have this it's solved like 99% of my problems.”
Moritz Kremb: Instruct OpenClaw to update its context files for behavioral feedback
“When you notice something that you want to happen again, or don't want to happen again, just tell your OpenClaw to update these files.”
Moritz Kremb warns autonomous AI agent API requests can cost $0.20 each
“If you use a model through the API, it can really get very expensive. Like every, every request can be like 20 cents and that can stack up really quickly.”
Moritz Kremb: OpenAI permits OpenClaw OAuth connections while Anthropic officially bans them
“Because OpenAI has stated that they're okay with it. So for them, it's definitely fine to use this OAuth method with Anthropic. It's kind of a gray area a bit. I think officially in their terms of service, it says you're not allowed to do that.”
Moritz Kremb: Anthropic is banning users for connecting OpenClaw via OAuth
“And some people have been getting banned over it. So I do recommend that if you, if you're scared about your account getting banned, just create a new one, create a new account and put the 20 dollar plan on that.”
Moritz Kremb automated his grocery ordering on German service Rewe using OpenClaw
“I've logged into that browser with my, like the, in Germany it's called Rewe. It's like this Instacart basically. And it will start like Ordering my groceries for me essentially and clicking around and ordering my stuff. So I've automated like this whole part …”
Moritz Kremb warns users to inspect ClawHub skills for suspicious community instructions
“One thing you should be, you should always be double checking these skills because anyone can create them and there can be like all kinds of instructions inside of these skills. So what actually the creator has done, they've, Like added this security scan here…”
Moritz Kremb: Hosting autonomous AI agents locally is safer than cloud VPS
“To mitigate that, I actually recommend just like set it up on a local Mac. Because there the risk is much lower than if you're setting it up on a VPS. Like a VPS is this thing that's in the cloud and it's like constantly connected to the internet. And it's jus…”
Moritz Kremb: Stronger frontier AI models natively defend against prompt injection attacks
“Actually the most useful one is to mitigate this risk is just to use a strong model. Because the smarter the model, the better it is actually at like not falling for these prompt injection tricks.”
Moritz Kremb: Automated AI content machines fail because audiences demand authentic trust
“And I would say nowadays it's really not that hard to just create this content machine that like just pumps out AI slop. But it's not useful at all. Right. Especially in an era AI era you know, where, where trust like is getting lower and lower, your content s…”