Summary
Highlights
Moritz introduces OpenClaw as a personal, autonomous agent that remembers, gets better over time, and automates tasks with access to built-in tools. He highlights its flexibility across chat tools and its status as the closest to a truly autonomous agent currently available. The discussion differentiates OpenClaw from Chat GPT and Claude Code, emphasizing OpenClaw's local operation, enhanced tool flexibility, and unique features like heartbeat and cron jobs.
The conversation touches on Claude Co-work's evolution towards OpenClaw's features, like Anthropic's 'Dispatch' preview. Moritz explains that Claude Co-work is essentially a nicer UI for Claude Code, while OpenClaw offers more powerful and open-source features. He anticipates that major players will develop their versions of OpenClaw, but OpenClaw will remain the flexible, community-backed open-source option.
Moritz outlines a 10-step guide to optimize an OpenClaw setup. The first step involves establishing a troubleshooting baseline by creating an 'OpenClaw support' project in Claude (or ChatGPT) and uploading the OpenClaw documentation. This ensures that when errors occur, the AI can reference the official documentation for accurate solutions, preventing reliance on potentially inaccurate web searches.
The second essential step is personalization, involving configuring files like 'agents.md' (agent behavior), 'soul.md' (personality), 'identity.md', and 'user.md' (user info) within the OpenClaw workspace. This provides context for the agent. The third step addresses memory persistence, a common user struggle. It involves ensuring 'memory.mmd' (long-term memory) and daily memory files within the 'memory' folder are created and maintained. A crucial setting is enabling 'compaction memory flash' and 'memory search experimental session memory' to prevent data loss during session compaction, along with an autosave feature in the heartbeat file.
Configuring models and fallbacks is the fourth step. Moritz recommends using the OAUTH method with an existing ChatGPT Plus subscription for cost-effective usage. He also advises setting up backup models, such as Anthropic's offerings, and additional fallbacks via services like OpenRouter or Kilo Gateway. This creates a robust 'backup chain' to ensure continuous operation even if a primary model becomes unavailable. A discussion on Anthropic's stance on OpenClaw usage and the importance of strong models for security is included.
The fifth step focuses on optimizing chat interactions, particularly with Telegram. Moritz suggests using Telegram groups and topics to separate conversations by subject (e.g., general, to-dos, journaling, content), preventing information overload and confusion. Each group/topic can have a specific system prompt, ensuring OpenClaw understands the context of the conversation.
The sixth and seventh steps involve leveraging OpenClaw's browser capabilities and skills. There are three ways OpenClaw can use the browser: regular web search/fetch, the OpenClaw managed browser (for logged-in actions like grocery ordering), and Chrome Relay (for temporary access to logged-in sessions on your main browser). The seventh step emphasizes utilizing skills, both built-in (e.g., summarize, Notion, OpenAI Whisper) and custom-built, found on marketplaces like Clawhub.ai. Users are cautioned to verify the security of third-party skills.
The eighth step covers the heartbeat file, which runs every 30 minutes, executing instructions like memory maintenance and to-do list updates. Users should carefully select contents to avoid excessive usage. The ninth step addresses security basics, distinguishing between backend access risk (mitigated by local Mac installation) and prompt injection. A simple prompt injection defense is shared, along with recommendations for storing API keys outside the workspace and using strong AI models. The tenth step highlights the 'principle of least access' and advises creating 'agent-owned accounts' for enhanced security and organization, treating the AI as a new employee with dedicated accounts.
Moritz presents a 'no AI slop short form video content system' he built with OpenClaw. This seven-step system generates authentic short-form video content by minimizing AI-generated 'slop.' Steps include: idea capture (automated from YouTube, Twitter, and manual input), weekly planning (using logged ideas and analytics feedback), script writing (based on a custom library of past scripts and styles), filming, editing (with automated uploading to an editor), posting across platforms, and analytics tracking.
Another use case is a flexible CRM system. This OpenClaw-powered CRM can be interacted with via chat (e.g., Telegram). It accesses Google Sheets for lead storage, Gmail for fetching information and updating the CRM, and the calendar for meeting details. The system can also be hooked up to WhatsApp and Telegram for automated follow-ups, including drafting messages based on existing templates.
Moritz concludes by emphasizing that OpenClaw is still in its early stages, akin to ChatGPT's initial release. Despite rough edges and bugs, it offers 'magical moments' that hint at its future potential. He believes that everyone will eventually have personal AI agents, whether OpenClaw-based or from other companies. He encourages early adoption and experimentation, seeing it as a significant opportunity to get ahead and learn how to manage these emerging AI employees effectively. Jensen Huang's statement about every company needing an OpenClaw-like agentic system is cited, underscoring the importance of this technology.