OpenClaw (Clawdbot / Moltbot) Videos & Summaries

OpenClaw 4.23: Key Features and Updates

• Enhanced Image Generation: OpenClaw 4.23 integrates OpenAI's GPT image model via Codex OAuth or Open Router, enabling image creation and reference editing without separate API keys. It allows for specific quality levels, output format hints, and background settings.
• Sub-Agent Context Management: The update introduces optional forked context for sub-agents, allowing child agents to inherit or isolate conversation history from parent agents, facilitating coordinated multi-agent systems.
• Per-Call Timeout Control: Media generation tools (images, video, music, TTS) now support custom timeouts per call, preventing failures due to slow provider responses for complex tasks.

Technical Improvements and Fixes

• Memory Configuration: Local embeddings feature a configurable context size, adjustable for resource-limited or high-performance machines.
• Bug Fixes: Resolved issues include Slack progress messages leaking into channels, Telegram images appearing as URLs, and Windows Codex model executable file discovery.
• Web Chat Interface: Generated images in the web chat interface now persist after page reloads.
• Dependency Updates: Bundled Pi packages updated to 0.70.0, incorporating the latest GPT 5.5 catalog metadata.

OpenClaw's Core Philosophy and Community

• Personal AI Assistant: OpenClaw runs locally, accessing user files, apps, and browsers, controlled via existing chat apps (WhatsApp, Discord, etc.).
• Proactive Operation: Agents can perform tasks proactively, send briefings, and run on schedules.
• Open-Source Advantage: Users maintain control, data privacy, and flexibility in model choice (Claude, GPT, local models, Open Router).
• Extensibility via Skills: Agents can write and install new skills, enhancing capabilities on demand. ClawHub serves as a community repository for these skills.

Introduction to Agentic Systems

• The video recommends an agentic architectural system as an alternative to OpenClaw.
• The focus is on using Claude Code to achieve similar benefits to OpenClaw while adhering to Anthropic's guidelines.

Problems with OpenClaw and Alternatives

• Anthropic has restricted the use of their subscription in third-party tools like OpenClaw, causing workflow disruptions.
• Other models like GPT 4.7 and 5.4 reportedly lack the "feeling" and "agentic capability" of Anthropic's Opus models.
• OpenClaw, while powerful, had security concerns and a complex codebase that was difficult to understand.

Claude Code Solution

• Claude Code, with its access to the local file system, is presented as a powerful starting point for agentic AI.
• A key benefit is its ability to perform tasks on the user's computer, like organizing files.
• The speaker uses the Claude Code CLI wrapped in a tmux session, accessible via SSH with Termius.
• This setup allows access through the native Claude app and provides a more understandable system than OpenClaw.
• The official Claude Code CLI Discord plugin is used as the primary interface for the agent, named "Toddbot."

Key Features and Use Cases

• Toddbot is configured to act as a "man of God" aligned with the organization's mission.
• The system can connect to local network devices, NAS, and databases (e.g., PostgreSQL on Railway).
• It can record ministry meetings and provide updates from them.
• Claude MD files are used to define agent instructions and context, similar to Obsidian markdown files.
• A local model (Quinn 3) with cloned voice is used for text-to-speech to aid with dyslexia.
• Reminders can be set via Discord, added to Google Calendar, and listed in Obsidian.
• A "dreaming feature" at 2 AM reviews logs and makes improvements to tool usage.

Security and Benefits

• Using Anthropic's maintained systems (Claude Code CLI and Discord plugin) ensures compliance with their terms of use.
• This approach is considered more secure than open-source projects like OpenClaw, which are more susceptible to attacks.
• It allows continued use of subsidized Opus models.

Future and Recommendations

• The speaker plans to release more in-depth tutorials on setting up and using this system.
• It is recommended for ministry leaders and anyone interested in personal agents for productivity and organization.
• The system offers redundancy through access via Termius and the native Claude app.

New Features on Tap Insight

• Introduced Street View on Google Maps and enhanced OpenClaw functionality for X3 Pro AR glasses.

Street View Integration

• Users can search for points of interest (POIs) using the AI assistant.
• Map loading is faster; results are presented in a list.
• Selecting a POI automatically switches to Street View.
• Street View allows camera rotation and navigation (driving/biking simulation).
• "Drone mode" provides an aerial view of the POI.
• Searches can be refined by criteria like price range and ratings.

OpenClaw File Support

• Allows transferring text and media files to the glasses.
• Supports downloading text files from sources like Project Gutenberg for reading.
• Can download and play playlists (e.g., M3Us for radio stations).
• Media browser includes folders for playlists, text files, and videos.
• Downloaded playlists automatically open in the media player.
• Bookmarking functionality saves found stations/files.
• Supports videos with closed captions and a scrub bar.
• Downloaded text files can be read on the glasses, with support for multiple text-to-speech models.
• Example demonstrated downloading and reading the sci-fi story "Flatland."

Open Claw Explained

• Open Claw is an AI agent framework allowing LLMs to access local device data (emails, calendar, files) to act as personal assistants.
• It's an open-source project, enabling integration with various LLMs like ChatGPT, Claude, Gemini, and local models via Ollama.
• The developer created it to give LLMs access to personal data for more helpful assistance, running locally rather than in the cloud.

Smartphone Integration

• Chinese brands like Xiaomi and Honor are developing Open Claw versions for smartphones.
• This aims to overcome current AI assistant limitations on phones, such as restricted access to local data like calendars.
• Unlike cloud-based assistants, Open Claw on phones could function offline and keep sensitive data private.
• Potential applications include complex file searches (e.g., photos from specific people/dates) and adding events from unstructured emails (like flight tickets) to calendars.

Comparison to Existing Solutions

• Current phone AI assistants often have limited access or rely on cloud syncing (e.g., Google Gemini accessing cloud calendar).
• Some devices like Oppo and Huawei show partial success in accessing data or performing actions on apps, while others like Xiaomi and iOS struggle.
• Open Claw's approach allows for deeper integration and offline capabilities, making it a potential game-changer.

Future Potential

• Open Claw could enable advanced features like automated email filtering and response based on channel content and user preferences.
• It allows using multiple LLMs, including local on-device models, for a truly personalized and usable AI experience on phones.
• This technology represents a significant step towards phones having truly intelligent personal assistants.

Platform Overview: AI on Edge

• Introduces "AI on Edge," a new platform designed for efficient online business operations.
• Author has 15 years of experience in online schooling (Chi Gong, Tai Chi, Kung Fu).
• The platform aims to solve common struggles with hosting and distributing online content.

Key Features

• Multi-language Real-time Translation: Comments appear in the user's preferred language, while admins see them in English.
• Automated Site Maps: Generates a single sitemap with all languages for better SEO.
• Efficient Video/Audio Processing: Uses efficient codecs for smooth content delivery on mobile devices.
• Document Security: Documents are destroyed and recreated to prevent script injection.
• AI Agent: Features a distinct AI agent approach with 22 "trees" for focused and efficient results.
• Memory Management: AI agent has short-term, mid-term, and long-term storage, understanding context shifts (e.g., changing cars).
• GDPR Compliance: 100% GDPR compliant, using only EU-based servers and no external APIs.

User Interface and Tools

• Admin Dashboard: Includes AI-generated analytics to help achieve business goals.
• AI Assistant: Can be toggled on/off for interaction, generates posts, summaries, and content.
• Content Creation: Supports product pages (displayed as cards/pricing tables) and Clickfunnels (infinite scrolling steps).
• Customization: Allows custom fonts, colors, reaction icons, and branding.
• External AI Integration: Option to integrate with external APIs like Open Claw or Hermes.

Promotional Offer

• Offers a 50% discount with a provided code.
• Professional plan includes unlimited multi-language features.
• Encourages users to try the platform despite ongoing development.
• Support available in the "HQ section" of the admin panel.

OpenClaw 4.23 & GPT-5.5 Integration

• OpenClaw 4.23 now includes GPT-5.5, accessible via OpenI or Codeex.
• Older models are automatically cleaned out and replaced with GPT-5.5 for a current model list.
• Utilize the latest AI capabilities for content writing, customer service, and task automation.

Enhanced Image Generation

• Agents can now create and edit images without needing separate API keys if logged into Codeex.
• Open Routter also supports image generation through OpenClaw 2 with a single key.
• Specify image quality, transparency, file format, and compression levels during generation.
• Improved editing capabilities for existing images, including fixing issues with multiple reference images.

Improved Multi-Agent Communication

• Sub-agents now share context with main agents, retaining conversation history for better task understanding.
• Helper agents no longer start blank, reducing redundant questions and improving workflow efficiency.
• Agents can set custom timeouts per task, preventing failures for complex generations.

Messaging & Browser Fixes

• Images now display correctly in Telegram group chats instead of just links.
• Internal progress messages are hidden in Slack channels, keeping them clean.
• WhatsApp media sending is now consistent across direct sends and automatic replies.
• Browser interface now displays error messages and suggests alternative models when issues arise.
• Generated images are now saved in the chat history and persist after page refreshes.

System & Security Enhancements

• Local search resource usage can be controlled for better performance on smaller machines.
• The "Dreaming" feature now functions independently of the "Heartbeat" system.
• Security tweaks for WhatsApp contact cards and group chat visibility prevent instruction injection.
• Discord commands and Telegram model buttons have been secured against unauthorized access.
• Android app auto-sending messages is now fixed.
• Plugin installation is more stable, with better error reporting and reduced issues on Windows.
• 29 contributors have added to this release, highlighting the power of open source.

OpenClaw 4.23 Update

• OpenClaw 4.23 is released with significant improvements.
• Integration with OpenAI's GPT 5.5 model is the main highlight.
• OpenAI's acquisition of OpenClaw is leading to better integration.
• New features include GPT Image 2 and GPT 5.5 compatibility.
• Project management features have also seen improvements.

Key Features and Improvements

• Image Generation: Now supports OpenAI's GPT Image 2 and improved image generation via OpenRouter.
• GPT 5.5 Integration: Enables more advanced capabilities and faster processing.
• Session Management: Enhanced control over sessions, allowing parent-child connections.
• Pre-call Timeouts: Implemented for image, video, music, and GPT tools.
• Bug Fixes: Includes various fixes and a doctor fix command for troubleshooting.

Before and After Tests

• Image Generation Test: Older version blocked image generation; updated version successfully generated a realistic image.
• Pomodoro Timer Test: Older version created a functional timer; updated version was faster and offered improved session handling and data export (CSV, longest streak).
• Performance: Updated version completed tasks significantly faster (e.g., 1 minute vs. 2 minutes for a previous project).

Surprising Insights and Challenges

• Despite improvements, the update process was not entirely smooth, requiring some debugging.
• Some users might prefer cloud versions like HeraCloud for a more consistent online experience due to potential offline issues with the raw OpenClaw version.
• GPT 5.5 shows serious improvements, making OpenClaw more capable, especially with its shift towards better OpenAI integration beyond just code.

Project Lobster and OpenClaw

• Microsoft is developing "Project Lobster," an internal team focused on bringing OpenClaw, an open-source AI personal assistant, to Microsoft 365.
• This initiative is led by Omar Shahine, Corporate Vice President at Microsoft Word.
• Project Lobster has an official repository on Microsoft's GitHub, forked from OpenClaw.

Satya Nadella's Involvement

• Microsoft CEO Satya Nadella has tested "Clawbot," an Anthropic-powered AI assistant for task automation.
• Nadella encouraged staff to develop similar products, although he initially dismissed the idea of Microsoft launching "OpenClaw" due to security concerns, calling it a "virus."

Functionality and Security Concerns

• Shahine describes "Claw Pilot" (referring to Project Lobster) as his main work interface, acting as a personal assistant that signs messages.
• He states it's a "personal agent with an intra identity operating securely inside your org managed by Agent 365," not just a chatbot.
• The Windows Defender antivirus team previously flagged OpenClaw as insecure and provided guidance on mitigating risks like prompt injection.

Business Model Implications

• The video suggests Microsoft aims to monetize this AI assistant, potentially through token-based pricing similar to GitHub Copilot.
• This could lead to users paying for more AI tokens to operate their personal assistants.

Tools vs. Skills

• Tools are raw capabilities (e.g., exec, read, web search).
• Skills are instruction manuals teaching agents when and how to use tools.
• Skills do not add permissions; they operate within existing tool policies.

Skill Structure and Configuration

• A skill is a skill.mdown file with YAML front matter (name, description) and markdown instructions.
• The loading order of skills matters.
• Skills are snapshotted at session start; changes require restarting the gateway or a new session.
• Configuration is in .openclaw.json, with per-skill settings and per-agent skill access control.
• Explicit per-agent skill lists replace defaults, allowing for specialized agents.

Claw Hub and Security

• Claw Hub is a marketplace for community-built skills (over 3,000 available).
• Skills are the most significant attack surface; malicious code can execute with agent permissions.
• Mitigation strategies include reading source code before installing, running in containers, verifying publishers, and avoiding unverified URLs.
• Installing skills is a security decision, not just a feature decision.

Key Takeaways and Next Steps

• List installed skills and check context; remove unused skills to save tokens.
• Install one skill from Claw Hub after reviewing its source code.
• Build a custom skill markdown file and iterate on its description for correct invocation.
• Perform security audits on installed skills (built-in and third-party).

Introduction to AI in Wholesaling

• Implementing AI is a necessity for wholesaling businesses to avoid being left behind.
• OpenClaw, used with a Mac Mini, can revolutionize real estate wholesaling.
• AI acts as a robot assistant to automate tedious tasks, reduce stress, and optimize operations.
• Adapting to new technology is crucial for success, similar to the early days of mobile phones.

OpenClaw's Capabilities

• OpenClaw controls your computer via terminal commands.
• Automated List Pulling: Request specific motivated seller lists (e.g., 15+ years absentee owners from PropStream) and have them emailed.
• Spreadsheet Management: Automatically split phone numbers into separate rows for better contact management and compliance.
• Automated Follow-ups: Instruct OpenClaw to contact leads in your CRM not reached in a set period (e.g., 3 months) via text, checking interest and booking appointments.
• Data Scraping: Scrape websites like Zillow, Craigslist, and Facebook Marketplace for "For Sale By Owner" listings or cash buyers, creating organized spreadsheets.
• Cash Buyer List Building: Extract cash buyer information (like emails) from Facebook groups to create a comprehensive list.

OpenClaw vs. Virtual Assistants

• OpenClaw offers advantages over human VAs: works 24/7, no hourly pay, no breaks, no training needed, and doesn't have human limitations.
• The AI assistant continues working on tasks even when you're not.

Setup and Usage

• Recommended setup: Mac Mini ($600) for its reliability in running 24/7 without unexpected shutdowns or updates.
• Installation involves going to opencloud.com, obtaining code, and entering it into the terminal.
• Setup questions guide the configuration process.
• Can be controlled via text messages, even remotely.
• OpenClaw can problem-solve, such as logging into an account to reveal hidden phone numbers on Zillow.

Dashboard Automation with AI

• AI can automate tasks like calendar checking, meeting summaries, daily task summarization, competitor analysis, and market research.
• A practical use case demonstrates converting raw data into a stakeholder dashboard summary using OpenClaw.
• The process involves uploading a data file (e.g., CSV) to OpenClaw via Telegram.
• OpenClaw analyzes the file and generates a concise dashboard including KPIs, scorecard, insights, opportunities, risks, and actions.
• The final dashboard is delivered via email with an interactive link.

Hosting OpenClaw

• For reliable, continuous automation, OpenClaw needs to be hosted, not run locally.
• Hosting options include local systems (for testing), external hardware (high cost), or a server/VPS (most practical and cheaper).
• Hostinger offers a simple, one-click VPS deployment solution for OpenClaw.
• A specific VPS plan (KVM2) is recommended, with a 12-month plan and a nearby server location.
• A coupon code "pavan pro" provides a 10% discount on Hostinger VPS plans.

Setting up and Using OpenClaw

• The setup involves purchasing an OpenAI API key and pasting it into the OpenClaw configuration.
• Connecting OpenClaw with Telegram requires creating a bot using BotFather, copying the bot token, and pasting it into the setup.
• Pairing the account with a code completes the Telegram connection.
• "Skill building" in Telegram allows creating custom workflows via prompts. A prompt can instruct OpenClaw to create a structured business analysis workflow for CSV files without manual coding.
• The deployed system runs continuously on the VPS and can be accessed via a gateway token.

Introducing OpenClaw Enhancements

• Most users treat OpenClaw as a basic chatbot, missing its potential.
• Five free tools can transform OpenClaw into a powerful agent.
• These tools enable email, web search, scraping, website publishing, and video creation.

Essential Free Tools for OpenClaw

• Agent Mail: For sending/receiving emails without being blocked like Gmail accounts.
• Here.now: For instant, free live web hosting of generated content.
• Tavoli: Provides live web search capabilities, overcoming AI knowledge cutoffs.
• Firecrawl: For efficient website scraping, returning structured Markdown or JSON.
• Remotion: Enables AI agents to create videos programmatically using code.

Tool-Specific Benefits and Setup

• Agent Mail: Offers a generous free tier (3 inboxes, 3000 emails/month).
• Here.now: Allows free web page hosting for 24 hours; registration enables permanent hosting.
• Tavoli: Provides 1000 free API credits monthly for deep web research.
• Firecrawl: Effectively bypasses website protections (e.g., Cloudflare) and scrapes efficiently.
• Remotion: Leverages agent's coding abilities to generate videos from natural language prompts.

Zapier SDK: The Ultimate Integration

• Zapier SDK (Open Beta) connects AI agents to over 9,000 apps.
• Simplifies API management compared to direct integrations.
• More token-efficient than Model Context Protocol (MCP).

Introduction to Local Models

• Local AI models offer a significant upgrade, capable of running on mid-tier hardware.
• Running local models can save thousands of dollars by eliminating reliance on cloud providers.
• Model choice and hardware are crucial for a good local model experience.
• Local models are typically open-source and free to use, unlike proprietary cloud models (e.g., GPT-4).
• Not all local models support tool-calling or agent orchestration needed for platforms like OpenClaw.

Hardware Requirements

• Mac Users: Focus on the total unified memory (RAM). Newer Macs (M-series) are generally sufficient. Older Macs (5-8 years+) may struggle.
• Windows/Linux Users: Focus on the GPU's VRAM. The amount of system RAM is less important.
• Higher VRAM (e.g., 24GB on a 4090) allows for more powerful models.
• Local models can utilize nearly 100% of available VRAM.

Setting Up Local Models with Ollama

• Install Ollama from ollama.com. Ensure it's the latest version.
• Ollama allows you to download and run local models on your machine.
• Recommended Model: Gemma 4 (smallest, best performing for many).
• Model size (in GB) is crucial and must be less than your available RAM/VRAM.
• Larger parameter counts generally mean better performance but larger file sizes and slower speeds.
• Download a model using `ollama pull [model_name]` (e.g., `ollama pull gemma:31b`).
• List downloaded models with `ollama list`.
• Test a model by typing `ollama run [model_name]` (e.g., `ollama run gemma`).
• Exit the chat interface with `/exit`.

Integrating Ollama with OpenClaw

• Ensure OpenClaw is installed.
• Open your terminal and run `openclaw configure`.
• Navigate to the 'model' selection.
• Choose 'Ollama' and then 'local only'.
• Select the desired downloaded model (e.g., Gemma 4).
• If Ollama isn't detected, run `ollama serve` in a separate terminal window.
• Restart the OpenClaw gateway: `openclaw gateway restart`.
• Once configured, OpenClaw will use the selected local model by default.
• You can switch models within OpenClaw using `/model [model_name]`.
• Consider setting a local model as the default and using cloud models for more complex tasks.

Introduction

• The host, Uncle Nerd, welcomes viewers to a live "Bust It Open" episode focusing on G.I. Joe Classified action figures.
• Viewers can vote on which figures will be reviewed first.

Figure Reviews

• Sci-Fi:
  • Chosen by the chat, Sci-Fi is figure #177.
  • Comes with a backpack, laser rifle with translucent blast effect, flexible hose, helmet, and two visors (classic and full-face).
  • Praised for its sculpt, articulation, and updated classic look.
  • Rated 5 out of 5.
• Cobra Claw with Female Cobra Pilot:
  • Triggered by a donation, this deluxe set includes the Cobra Claw vehicle and a female Cobra officer (#152).
  • The pilot figure has interchangeable heads (black and white), helmet, weapons (Uzi 9mm, pistol, missiles, blast effects), and a harness.
  • The Cobra Claw vehicle is large, detailed, and features poseable landing gear and wings.
  • Both the figure and vehicle are highly praised for detail and accuracy.
  • Rated 5 out of 5.
• Law & Order:
  • This set includes the human officer, Law, and his K9 companion, Order.
  • Law comes with multiple heads, hands, a baton, knife, cuffs, and an Uzi.
  • Order the dog is highly detailed with articulation and an optional mask.
  • Praised for its accessories and the quality of the animal figure.
  • Rated 5 out of 5.
• Big Lob:
  • Figure #189, reviewed briefly at the end due to time constraints.
  • Comes in a standard window box with okay box art.
  • The host notes Lob is taller than expected.

General Discussion

• The host expresses love for the G.I. Joe Classified line, appreciating the detail despite initial apprehension about the 6-inch scale.
• He mentions specific classic G.I. Joe characters he's waiting for, like Flash and Breaker.
• The importance of viewer interaction in live reviews is highlighted.

Grok Integration

• OpenClaw 4.22 now natively supports Grok image generation (Imagine and Imagine Pro).
• Features include editing images with reference photos.
• Supports six live Grok voices for text-to-speech (MP3, Wave, PCM, G.711 formats).
• Grok's speech-to-text offers real-time transcription with a 5% error rate for phone call entity recognition, outperforming competitors like 11 Labs (12%) and Deep Graham (13.5%).

Tencent HY3 Model

• Bundled with OpenClaw 4.22, HY3 is a 295 billion parameter Mixture of Experts model with 21 billion active parameters per request.
• Achieves 74.4% on SW Bench Verified and 67.1% on Browse Comp.
• Scored 88.4% on a math PhD qualifying exam.
• Supports a 256k context window and is open-source.

Local Mode and Model Management

• Introduced a local TUI (Terminal User Interface) for running terminal chats without a gateway server.
• New command `/models add` allows adding models directly within chat without restarting or editing config files.
• Auto-install feature now automatically installs missing provider and channel plugins during onboarding.

Diagnostics and Security

• Stability data is recorded by default; a diagnostics package (logs, config, etc.) can be exported for debugging.
• A "last known good config" restore option is available for recovery from configuration errors.
• Numerous security improvements and a more resilient config system.

Provider and Channel Updates

• Amazon Bedrock now supports Claude Opus 4.7 through the mantle route, with tokens refreshing at runtime.
• GPT prompt overlay system is now unified across OpenAI, openrouter, and other GPT providers.
• WhatsApp integration features configurable native reply quoting and per-group/per-direct chat system prompts.
• Session management improved with mailbox-style filters (label, agent, search).

Performance and Usability Enhancements

• Plugin load time reduced by 82-90% for bundled plugins using JITi loading.
• New community plugin "token juice" compacts noisy command-line tool output.
• CLI sessions now persist through gateway restarts and idle exits.
• Default "thinking" setting increased to medium for reasoning-capable models.

Cost Optimization with OpenRouter Auto

• The most significant upgrade is routing tasks through the cheapest capable model using OpenRouter Auto.
• This prevents overspending on powerful models like Claude Opus for simple background tasks.
• Costs can be reduced by 15-20x by dynamically switching models based on task complexity.
• A prompt like "Set OpenRouter as my primary model provider, enable auto-rooting, and use the absolute cheapest model for background checks and heartbeat tasks, but automatically upgrade to advanced reasoning models when you think that'd be a better model for the job" can configure this.
• Adjusting heartbeat frequency (e.g., to a daily cron job at 8 AM using a low-cost model) also significantly cuts costs.

Permanent Memory with Dreaming

• OpenClaw's new Dreaming feature provides permanent memory that survives new chats.
• It sorts through daily chats during downtime, saves important context, and discards noise.
• To enable: go to Control Center, then Dreaming. Ensure the toggle is on.
• To seed memory: Use a prompt like "Enable the Active Memory plugin. Use balanced query mode with a 15-second timeout and apply it to the main agent only. I am Jimi Barkway. I run a YouTube channel focused on AI automation and building AI agents. I'm based in the UK my core tech stack includes Claude Code OpenClaw I prefer technical no fluff answers when we're brainstorming workflows I prefer you to challenge me when I need to be challenged and I want straight to the point no fluff answers when we're brainstorming new workflows."

App Integration with Zapier SDK

• The new Zapier SDK allows OpenClaw to connect to thousands of apps without managing individual API keys.
• This replaces the older MCP and acts as a universal adapter.
• To set up: Use a prompt like "Install the Zapier SDK using the official docs, execute the setup, up install the required packages and provide me with the login link".
• Once set up, the agent can perform actions across connected apps using natural language, such as fetching YouTube analytics and sending them to Slack.
• The Zapier SDK is currently in open beta and free.

Native Media Generation

• OpenClaw can now generate images and music directly from the chat interface.
• Example prompts include: "Generate me an image of a rubber duck playing the guitar floating along the river thames in london. Make it beautiful cinematic super high definition 8k" and "Generate a 30 second lo-fi chill hop beat suitable for background music in a coding tutorial. Keep it relaxed and atmospheric."

Agent Control and Safety

• A crucial "kill switch" command, /stop, is available to immediately halt a runaway agent.
• This is essential when an agent might go off track, get stuck in loops, or consume excessive resources.
• Simply type /stop in the chat to abort any process.

OpenClaw vs. Traditional AI Tools

• OpenClaw enables AI assistants to execute commands, read files, make decisions, and automate tasks, unlike tools like ChatGPT or Gemini that primarily generate text.
• It connects AI to real systems, allowing it to interact with your digital environment.

Risks and Safeguards

• Potential risks include granting access to file systems, API keys, or messaging accounts if misconfigured or exposed.
• Third-party skills can also pose risks if they request excessive permissions.
• Running OpenClaw in an isolated environment, like on Hostinger, mitigates these risks by separating it from personal computers and data.

Setting Up OpenClaw with Hostinger

• Hostinger offers a managed OpenClaw plan for easy setup without server configurations.
• It simplifies AI model access through a credit system, eliminating the need for multiple API key configurations.
• The setup includes connecting messaging apps like Telegram.

Telegram Bot Integration

• Users need a Telegram account and must create a bot using Telegram's "BotFather".
• BotFather provides a unique bot token, crucial for connecting OpenClaw to the Telegram bot.
• A pairing token is generated and used to link the OpenClaw instance to the Telegram bot.

OpenClaw Capabilities

• Once set up, the assistant can perform tasks like researching tech news, summarizing articles, and drafting messages.
• It can monitor YouTube channels for new content and provide daily summaries via Telegram.
• Users can expand functionality by installing additional "skills" for specific tasks (e.g., weather, Trello).

Daily Briefing

• Use Case: Get a structured summary of your day before starting work.
• Setup: Use a cron job to pull from your calendar, inbox, and task list.
• Skill: Gogg (built-in) for Google Workspace integration.
• Output: Sent to WhatsApp or Telegram.

Idea Capture & Organization

• Use Case: Quickly save and organize ideas via voice notes.
• Setup: Send voice notes to OpenClaw on Telegram/WhatsApp, use "idea" keyword.
• Skills: Speechtoext, File System (built-in).
• Automation: Nightly task transcribes voice notes, categorizes them, and saves as markdown. Weekly digest summarizes ideas.

Content Research & Topic Discovery

• Use Case: Automate research for content creation.
• Skills: Tavi (web search), Deep Research (from Claw Hub).
• Process: Get weekly digests of emerging topics in your niche, tagged by category and relevance score. Deeper dives provide detailed briefs.
• Note: Tavi requires an API key and subscription; alternatives may exist.

Business Knowledge Database

• Use Case: Consolidate scattered business information into a searchable database.
• Skills: Agent Brain, Obsidian/Notion, or Gogg for Google Drive (from Claw Hub).
• Process: Install memory layer, prompt agent to build the database.
• Habit: Drop documents (articles, policies, manuals) into chat or workspace files. Agent organizes them.

Self-Improving Agent

• Use Case: Enable OpenClaw to log errors, track performance, and improve its behavior.
• Skills: Self-Improving Agent, Agent Brain (from Claw Hub).
• Process: Agent creates a "learnings" folder to log mistakes. Users correct errors, agent adjusts over time.
• Security Warning: Be cautious with self-improvement skills; vet them thoroughly (VirusTotal, Claw Hub scan, GitHub stars/reviews). Run OpenClaw on a VPS.

Anthropic's Miscalculation and Flywheel Issues

• Anthropic's core strategy has been a powerful flywheel: build an incredible coding model, sell it to enterprises for high revenue, use coding data to train better models.
• This model has been highly successful, surpassing OpenAI in revenue.
• However, Anthropic miscalculated the demand for AI compute, failing to invest sufficiently in infrastructure.
• This compute shortage is breaking down their flywheel, limiting their ability to serve demand.

Communication and Policy Problems

• Anthropic is facing a PR and communications nightmare due to confusing and seemingly disingenuous policy changes.
• Recent changes removed Claude Code from Pro plans (later rolled back for some) without clear announcements.
• They have created confusion around the use of Claude subscriptions with third-party tools like OpenClaw, with unclear terms of service and conflicting statements from employees.
• Downtime for Claude's API and services is significantly higher than competitors like OpenAI.
• Policy changes, like session limit adjustments, are perceived as punishing users (especially those using agentic harnesses) without clear explanation.

OpenAI's Strategic Advantage

• OpenAI, by contrast, aggressively invested in compute, taking on financial risk.
• They are capitalizing on Anthropic's issues by capturing excess demand and positioning themselves as reliable.
• OpenAI actively engages with its community, resets limits, and has acquired key talent and projects (like OpenClaw's creator).
• OpenAI's broader mission of "AI for the world" contrasts with Anthropic's more selective approach.

Token Economics and Compute Crunch

• Anthropic's subscription model is predicated on users not using their full token quota, but agentic use cases have increased demand.
• They are selling tokens at a loss and are reluctant to raise prices for fear of losing customers to OpenAI.
• The Opus 4.7 model, with its new tokenizer and increased token usage, exacerbates their compute crunch.
• Companies like Google have abundant compute, serving their own models and competitors, while XAI has compute but less demand.

Introduction to Space Agents

• Space agents are AI agents that run inside Hugging Face Spaces, which are like tiny websites for AI tools.
• They offer a simple, browser-based interface for users to interact with AI agents that can plan and take action.
• This contrasts with traditional AI agent systems like OpenClaw, which often require significant setup and technical knowledge.

Space Agents vs. OpenClaw

• OpenClaw is powerful but requires installation, configuration, and technical expertise.
• Space agents eliminate the need for installation, allowing users to access AI agents via a simple link.
• For business owners, space agents offer a much faster and more accessible way to implement AI for tasks like lead generation and content creation.

Practical Applications and Benefits

• Space agents can automate business workflows, such as finding leads on LinkedIn, drafting marketing copy, or generating content calendars.
• They operate on low-cost infrastructure, making AI automation significantly cheaper than hiring VAs or interns.
• The ease of use allows users to quickly test ideas and implement AI without extensive coding or setup.

Tips for Using Space Agents

• Start with a single, time-consuming task to build confidence.
• Write detailed prompts that provide context, goals, and desired output format.
• Maintain a library of successful prompts for future use.
• Always review AI outputs for accuracy before final use.
• Share space agents with team members for streamlined collaboration.

The Future of Software and AI

• Space agents represent a shift towards browser-based, conversational software interfaces.
• While OpenClaw remains valuable for power users and custom builds, space agents provide a faster entry point for most business owners.
• The key advantage for businesses is the increased speed and accessibility of AI automation, leading to greater efficiency and growth.

AI Employee Customization

• The core of an AI employee is its "brain," which can be customized through specific files in the workspace.
• Key files include soul.md for personality and behavior, identity.md for name and description, agents.md for operations and rules, tools.md for guidance on tool usage, user.md for user-specific information, and memory for storing past interactions.
• Changes made to these files require a session reset (/reset command) to take effect.
• When customizing, focus on relevant, concise instructions to avoid unnecessary token consumption and cost.

Workspace Files and Their Purpose

• soul.md: Defines the AI's personality, tone, and preferences (e.g., speaking in Urdu, being concise).
• identity.md: Stores the AI's name, role, emoji, and avatar.
• agents.md: Manages agent operations, rules, and session management.
• tools.md: Provides guidance on using available tools.
• user.md: Contains information about the user (name, location, preferences) to tailor responses.
• heartbeat: Periodically checks system status.
• memory: Stores past interactions and learned information.
• bootstrap.md: A one-time file used during initial setup to guide the AI's first interactions and configuration. It's deleted after its purpose is served.

Practical Application and Verification

• Changes to files are saved to disk but not immediately reflected in the current session.
• A /reset command is necessary to reload the updated configurations.
• The /context list command can be used to inspect token usage and file contributions.
• Verification is crucial: after making changes, confirm them by checking file content and observing the AI's behavior, especially after a reset.
• Hallucination is possible; always double-check AI responses against actual file changes.

Hermes Agent Overview

• Hermes agent is a new type of AI agent, not a replacement for OpenClaw or Cloud-Code.
• Developed by a community of 500+ contributors on GitHub, allowing constant updates.
• Can be run locally or on a VPS, with the code living on the same machine as the agent.

Key Advantages Over Other Agents

• Self-Awareness & Customization: Hermes can analyze its own code, add new functionality, and suggest changes via pull requests. Users have 100% control and can customize it to their workflow, unlike opinionated agents like Claude-Code, Codex, or Open-Code.
• Multi-Model Flexibility: Unlike other agents that lock you into one model (e.g., GPT-4 for Codex), Hermes allows you to use any LLM combination for different tasks (e.g., GPT-4 for backend, Gemini for UI, Opus for copy). This is managed through a "profiles" feature.
• Unified Context & Memory: Hermes maintains a single context for all projects and skills, eliminating the need to re-explain context when switching between models or sessions, unlike other agents where sessions are isolated.
• Built-in Self-Learning: Hermes automatically learns and accumulates skills from user interactions, growing with the user. This is not a manual process as with other agents that require explicit prompting to save skills.
• Pluggable Memory: Supports various memory providers, enabling persistent storage of context and data for different purposes, like course analytics.

Hermes Agent Capabilities

• Can be used anywhere, with any model (local or cloud), and supports full freedom in how you work.
• Adapts to your workflow and preferences, unlike other agents that dictate usage.
• Skills are auto-managed, and projects are remembered, fundamentally changing AI development workflows.
• Considered the "first everything agent" due to its comprehensive feature set and reliability.

Setup Steps

• Enable the Google Workspace CLI skill in OpenClaw.
• Create a new project on Google Cloud (e.g., "Open claw 2").
• Enable necessary Google APIs: Gmail, Drive, Calendar, Docs.

Credentials and Authorization

• Configure the OAuth consent screen on Google Cloud.
• Create an OAuth client ID of type "Desktop app".
• Download the JSON client secret file.
• Add your email address as a test user in Google Cloud's OAuth consent screen.
• Send the JSON file and a setup request to the OpenClaw agent via Telegram, including your email.
• Run the provided command in your terminal to authorize OpenClaw access.

Security Considerations

• Avoid using your primary personal email for OpenClaw access.
• Consider potential vulnerabilities like prompt injection if granting email access.
• Suggests using alternative, more secure methods like NASN.

Demonstration

• Prompt OpenClaw to create a Google Doc from an existing blog post.
• OpenClaw successfully creates the document and provides a Google Drive link.

Release Overview

• OpenClaw released two updates: 4.20 and 4.21.
• 4.20 focuses on quality-of-life improvements, especially for VPS users.
• 4.21 is a patch addressing critical issues, primarily an NPM upgrade problem.

Setup Improvements (4.20)

• Wizard UI refresh for new users.
• Clearer security disclaimers with checklists.
• Loading spinner during initial model catalog load.

VPS & Performance Enhancements (4.20)

• Out of Memory (OOM) Prevention: Enforces session entry caps and age pruning by default.
• Loopback Pairing Fix: Resolves issues with TUI/local tools connecting to VPS.
• Service User Exec: Handles service accounts using shells, falling back to SH.
• Yolo Here Doc Fix: Addresses issues with here docs and forms under strict security.

Agent & Model Improvements

• Default system prompt and GPT-5 overlay upgraded for clearer bias and better recovery.
• Fix for the /think command with OpenAI models to improve reliability.
• Tiered model pricing supported for more accurate cost estimates.
• Kimiko 2.6 and 2.5 included in token usage reports.
• Estimated Cost USD Compounding Bug Fix: Corrects inflated token costs from repeated persist paths.

Session & Memory Management

• Graceful degradation for active memory recall failures; logs warning and continues.
• Clean separation of cron state with a new jobs_state.json.

Integration & Session Fixes

• Web search plugin API keys now resolve correctly with shared config.
• /new and /reset commands properly clear auto source model provider.

Critical Patch (4.21)

• Self-Repair for NPM Dependencies: NPM updates now automatically restore bundled plugin dependencies, fixing issues with Telegram, Discord, and Slack.
• Added Docker end-to-end test coverage for this scenario.

GPT Image 2.0 Support (4.21)

• Backported support for OpenAI GPT Image 2.0.
• Now the default model for image generation and editing.
• Supports up to 4 images per generation, edit mode with 5 reference images, and 2K/4K output.
• GPT 1 is still available as an override.

Open Claw vs. Hermes Agent: Design Philosophy

• Open Claw: Describes itself as a "self-hosted gateway" connecting chat apps and agents. It acts as a central bridge, managing sessions, routing, and connections.
• Hermes Agent: Calls itself a "self-improving AI agent" with a built-in learning loop that creates and refines skills from experience.

Core Differences: Memory vs. Skill Growth

• Open Claw: Features an explicit, file-system-oriented memory system (memory.md, date files, dreams.md) designed to help it know the user better over time.
• Hermes Agent: Focuses on skill growth. It creates skills from complex tasks and user corrections, wiring skill development directly into its operational loop. While it has memory, its emphasis is on improving its ability to work with the user.

Behavioral Metaphors

• Open Claw: Feels like a "family member" – remembers personal history, habits, and facts, offering a relationship-oriented experience.
• Hermes Agent: Acts like an "experienced employee" – learns workflows, sharpens at repeated tasks, and accumulates procedural knowledge for effectiveness.

Setup and Integration

• Hermes Agent offers a migration path for Open Claw users, allowing import of memories (.md files), command allow lists, user skills, and model configurations.
• Users can ask Open Claw to install and configure Hermes, reducing setup friction.

Standout Features

• Hermes Agent: Praised for transparency (streaming tool output, verbose models), interruption capabilities, and focus on repeated workflows, especially for operational routines.
• Open Claw: Offers richer broader assistance, better managed browser capabilities for tasks requiring credential input, and a wider array of community-built tools and plugins.

Security

• Both systems prioritize security but implement it differently. Hermes uses a "defense-in-depth" model, while Open Claw provides explicit controls for allow lists, mention rules, and group policies.

Conclusion: Coexistence, Not Replacement

• Neither agent fully replaces the other; they serve different primary purposes.
• Choose Open Claw for: Cross-surface assistance, rich routing, explicit workspace memory, managed browser use, and a long-term personal assistant feel.
• Choose Hermes Agent for: Self-improving capabilities, procedural skill growth, and sharpening at repeated workflows.
• Many users may benefit from running both, leveraging Open Claw as a resourceful assistant and Hermes as a specialized workflow executor.

OpenClaw AI Agent Setup and Capabilities

• OpenClaw agents run on cloud-based virtual computers (VPS) from services like Hostinger or Chorus, acting like a dedicated machine for the AI.
• Agents have a file structure similar to a local installation, with "skills" being packaged abilities stored as markdown files.
• Skills can be easily added from a marketplace; the agent is aware of and can install them automatically.
• The agent runs 24/7 unless deleted, functioning like a cloud-based computer.

Optimizing Agent Skills

• Adding too many skills (over 20) can lead to decreased performance and the agent using incorrect skills.
• A sweet spot for effective skills is between 7 and 20.
• It's better to have an agent proficient in a few specific skills rather than many scattered ones.
• Multiple agents can be created for different purposes to manage skill sets effectively.

Context Engineering and Agent Onboarding

• The focus is on "context engineering" rather than "prompt engineering," meaning guiding the agent to the right context.
• Onboarding an agent involves providing it with relevant information, similar to training a new employee.
• OpenClaw agents can automatically update identity and "soul.md" files based on interactions and provided context.
• Skills like "Super Data" can extract transcripts from YouTube links, allowing the agent to analyze content.

Agent Memory and Autonomy

• Agents log "memory files" of important interactions and insights.
• This memory allows the agent to recall past actions and information for future tasks, improving its usefulness over time.
• Future AI models will be able to add skills and use them within the same response, increasing autonomy.
• The goal is for agents to become more autonomous, requiring less explicit instruction for desired outcomes.

Improving Agent Performance and Error Reduction

• Users can monitor agent outputs and identify areas for improvement by refining skills or prompts.
• Skills can be edited or custom skills can be created as Standard Operating Procedures (SOPs).
• Evaluating agent output ("eval") and iterating based on feedback is crucial for improvement.
• Creating "dummy" environments (e.g., a fake email, a duplicate Notion database) is essential for safe testing and reducing unintended consequences.

Agent Integration and Automation

• OpenClaw agents can be integrated with communication platforms like Telegram, Slack, and iMessage.
• Notion integration allows agents to access and manage content databases.
• Cron jobs act as automation triggers, allowing agents to perform tasks on a schedule (e.g., generating daily reports).
• Agents can be instructed to perform complex workflows, like analyzing YouTube comments and generating reports.

The Future of AI Agents and Employment

• AI agents are becoming capable of handling tasks previously done by human employees, leading to potential job displacement in the short term.
• Learning to work with and manage AI agents will be crucial for future employment.
• New, smaller companies may emerge, leveraging AI agents to achieve the output of larger teams.
• Domain expertise is vital for creating high-quality agents and identifying AI-generated "slop."
• The process of working with AI agents is learned through iterative practice, not just by following rules.

Getting Started with AI Agents

• The first step is to automate a single, annoying task to gain immediate value.
• An example is automating social media management for brand deal inquiries.
• The key is continuous iteration, learning from mistakes, and refining agent instructions.
• A challenge is proposed: create a specific agent workflow, share it, and tag the creator for a retweet.

Open Claw Project Overview

• Open Claw: A rapidly growing open-source personal AI assistant project, originating from a weekend project by Peter Steinberger.
• Capabilities: Can read/write files, run scripts, browse the web, manage calendars/inboxes, check flights, search, and remember user preferences and past conversations.
• Data Privacy: Emphasizes that user data remains private and under user control.
• Accessibility: Integrates with various chat apps like WhatsApp, Telegram, Discord, Slack, Signal, and Messages.

X API Integration

• Elon Musk's Announcement: Elon Musk personally tweeted about X API accessibility via Open Claw, aiming for affordability.
• XURL Skill: Open Claw now natively supports the X API through the XURL skill, developed by the X developer platform team.
• Simplified Access: Eliminates the need for complex coding, OAuth tokens, and raw API calls for X API interaction.
• Functionality: Allows users to post to X, search, read timelines, reply, quote tweets, look up profiles, and track mentions via conversational commands.
• Chained Actions: As an AI agent, Open Claw can chain actions, such as searching for topics, summarizing posts, and drafting replies in one go.

Open Claw's Ecosystem and Security

• Skills System: Open Claw utilizes a skill system, allowing for plugins to add new capabilities.
• Claw Hub: Features a marketplace with over 13,000 community-built skills for various applications (Gmail, GitHub, Spotify, etc.).
• Self-Creation of Skills: Open Claw can generate new skills based on user descriptions.
• Security Measures: XURL skill is designed to protect API tokens. The project has actively addressed security vulnerabilities and partners with VirusTotal for skill scanning.
• Responsible Usage: Users are advised to use strong models, follow documentation, and be cautious about granted permissions.

Practical Implications and Getting Started

• For X Users: Streamlines content creation, information gathering, and audience engagement by consolidating X tasks with other AI functions.
• For Developers: Lowers the barrier to entry for building on the X API.
• For New Users: Serves as an impressive demonstration of personal AI agent capabilities beyond simple chatbots.
• Installation: Simple single-command installation with solid documentation and an active Discord community.
• Current Version (April 2026): Version 2026.4.14, supporting GPT-5.4 Pro, improved memory, and ongoing reliability enhancements.

OpenClaw 4.20 Release

• OpenClaw 4.20 is out, featuring 106 commits, incorporating elements from OpenAI GPT-5 and default system prompts.
• Improvements include better prompt and context efficiency (trim startup, skill prompt, subject bounded memory test), leading to less context bloat, fewer "gyron pron" payloads, and potentially cheaper, cleaner long sessions.
• Updates also address bugs in CodeX and introduce Kimmy 2.6 (similar to Opus 4.6).
• Onboarding and session maintenance have been updated for better memory usage and long-term tasks.

Testing and Performance

• Initial testing of the old version to build a landing page and deploy it to Vercel/GitHub took 6 minutes.
• A fresh install of OpenClaw 4.20 with Kimmy 2.6 was attempted.
• Encountered errors related to a new, high-download package integration, causing the UI to be down and requiring restarts.
• After troubleshooting, a new prompt for the same task (landing page, private GitHub) took 4 minutes, showing improvement.
• The agent also created a user plan, which was not observed before.
• Despite improvements, some UI elements (red buttons) remained inconsistent.
• The video highlights a desire for less verbose "thinking" processes from the AI, focusing on results rather than the internal thought steps.

Insights and User Experience

• The update aims to make OpenClaw smarter by focusing on output rather than showing internal thought processes.
• The ease of use without API keys or bot tokens is a key advantage for OpenClaw.
• A fresh install experience was not seamless, with an overwhelming choice of models and initial setup issues.
• The video suggests that users are early adopters and explorers in this evolving AI landscape.

Kimmy K2.6 Integration

• OpenClaw 4.20 now defaults to using Kimmy K2.6, an open-source, fast, and cost-effective model.
• K2.6 is superior to K2.5 in thinking, replying, and tool handling for agentic workflows.
• Agents using Kimmy are now smarter without manual updates.
• Features like web search and media understanding leverage K2.6.
• A new "thinking keep all" mode allows the model to think before every reply, enhancing accuracy at a lower cost.

Enhanced Agent Reliability

• The forward slash "think" command is smarter, with OpenClaw pre-checking model capabilities to prevent errors.
• Models now respect the "thinking off" setting consistently.

iMessage Agent Fixes

• Significant improvements for sending iMessage agents on macOS, resolving previous issues with delayed or vanishing messages.
• Increased send time limit and optimized system usage for reliable message delivery.
• Fixed issues with plain text sending and tapbacks (iMessage reactions).
• OpenClaw now prioritizes iMessage over SMS for contacts with both enabled.

Rebuilt Cron Jobs (Scheduled Tasks)

• Addressed bugs related to job delivery, false errors, and memory leaks.
• Improved runtime state management for cleaner job lists.
• Fixed "main session delivery" so scheduled jobs reply in the correct chat.
• Jobs with multiple channel posts now have pre-save validation to prevent failures.

Security Improvements

• Implemented SSRF guard to prevent malicious requests from exploiting the server.
• Blocked workspace environment variable injection.
• Restricted AI agent's ability to modify its own configuration, sandbox settings, or trust rules.
• Tightened device pairing security, ensuring devices only access their own information.

Memory and Session Management

• Introduced a cap and age-based cleanup for old sessions to prevent gateway crashes due to memory overload.
• Fixed cost tracking to accurately reflect expenses by snapshotting costs properly.
• Forward slash "new" and "reset" commands now function correctly, clearing stale settings for fresh sessions.

Improved Agent Personality

• Models like GPT-5 and CodeX now better interpret personality files (sol.md, identity.md), resulting in less robotic and more authentic agent responses.

Other Updates

• Auto-reply now adapts behavior based on message context (direct chat vs. group).
• Setup wizard is now a guided walkthrough.
• Telegram polling stalls are reduced.
• Matrix channel settings can be modified without restarting.
• Manos users see live draft previews.
• Discord's forward slash "think" command shows only valid options; channel issues no longer cause crashes.

Introduction to OpenClaw and AI Marketing Team

• The video demonstrates building a fully autonomous multi-agent marketing team using OpenClaw.
• Four distinct AI agents are created: Orchestrator, Trend Scout, Growth Agent, and Copywriter Agent.
• All agents communicate and operate within Discord.

Agent Roles and Setup

• Orchestrator: The central AI that plans and assigns tasks.
• Trend Scout: Researches web trends (e.g., in the digital nomad world).
• Growth Agent: Optimizes content for SEO and discoverability.
• Copywriter Agent: Crafts on-brand content from provided data.

Technical Setup and Configuration

• The setup uses OpenClaw on an Ubuntu VM and the MiniMax M2.7 model for its strong agentic capabilities.
• Discord bots are created via the Discord Developer Portal, requiring bot tokens and enabling privileged gateway intents.
• Bots are added to a specific Discord server (e.g., "SpotSeeker server") using OAuth2 authorization URLs.
• OpenClaw configuration (openclaw.json) maps agents to Discord bot accounts and defines channel access using Guild and Channel IDs.
• Privileged intents and specific channel permissions (allow list, mention requirement) are configured for bot communication.

Agent Personalization and Task Execution

• Each agent is set up in a separate Discord thread to manage context effectively.
• Prompts define agent personas (e.g., calm, decisive for Orchestrator; analytical for Growth Agent).
• A crucial Discord mention trick (@userID within angle brackets) enables autonomous task handoffs.
• Agents are instructed to initialize memory (short-term and long-term) and follow defined responsibilities.
• Skills can be created for agents (e.g., fetching Google Trends data) and tested.

Demonstration: 30-Day Marketing Launch Plan

• A 30-day marketing launch plan for a SAAS app "SpotSeeker" in New York is initiated.
• The Orchestrator breaks down the task and assigns it to the Trend Scout for market research.
• The Growth Agent then builds an SEO keyword matrix and GEO strategy based on the research.
• The Copywriter Agent uses this data to generate content (LinkedIn posts, Twitter threads, YouTube scripts).
• Despite a minor access issue, the agents collaborate, and the final output includes 15 LinkedIn posts, 20 Twitter posts, and 3 YouTube scripts, all data-backed and on-brand.

OpenClaw vs. Hermes Agent: Key Differences

• Community Size: OpenClaw has a larger community (360k stars) than Hermes Agent (96k stars), potentially leading to more frequent updates for OpenClaw.
• Skills Ecosystem: Skills are transferable between AI agents; they are not agent-specific. You can use skills from various sources on both OpenClaw and Hermes Agent.

Hermes Agent's Self-Learning Loop

• Hermes Agent features a unique self-learning loop that allows it to create and improve skills automatically after a task is performed multiple times (typically after 5+ tool calls).
• This process uses a genetic algorithm to refine skills over time, making the agent progressively better at specific tasks without manual intervention.
• OpenClaw requires manual skill creation or installation from external sources.

Memory System Differences

• Hermes Agent is designed to be lean, loading only essential memory files (e.g., user.md, memory.md) with strict token limits to reduce bloat and improve LLM prompt caching.
• OpenClaw's memory system can become bloated over time as it loads more historical data, potentially leading to less consistent script quality.
• When searching for information, Hermes Agent first checks its lean memory and then performs a vector search. OpenClaw's vector search fires more broadly, potentially pulling more historical data but also introducing more noise.

Multi-Agent Capabilities

• OpenClaw excels at multi-agent communication, allowing agents to interact with each other using "sessions send" and offering channel binding for specific integrations (e.g., Slack, Telegram).
• Hermes Agent can spawn sub-agents (up to 3 typically) and use profiles, but these sub-agents are isolated and cannot communicate directly with each other; they run tasks and report back, then cease to exist.

Tool Usage and Interaction

• Both agents support extensive tool usage.
• Hermes Agent visually displays tool usage during its responses, whereas OpenClaw typically only shows a "thinking" status.
• Hermes Agent can interrupt its current response if a new message is received, unlike OpenClaw, which can be frustrating for long-running tasks.

Setup and Provider Support

• Hostinger offers a one-click VPS setup for Hermes Agent, simplifying deployment and security.
• Hermes Agent supports a wider range of LLM providers (including open-source models via OpenRouter) compared to OpenClaw's more closed ecosystem.

Recommendation: Use Both

• The speaker recommends using both agents simultaneously.
• Hermes Agent is ideal for its self-improvement capabilities and lean memory, while OpenClaw is better for complex multi-agent communication setups.
• A suggested approach is to run Hermes Agent on a VPS for its self-learning and OpenClaw on a local machine for multi-agent tasks, potentially integrating them via a shared memory system like Obsidian.

Pi AI Agent Overview

• Pi is a minimalist, open-source AI coding agent toolkit.
• It is designed to be faster, smaller, and significantly cheaper than other AI tools.
• Pi can run on low-resource devices like a Raspberry Pi.

Pi vs. Open Claw

• Open Claw, a popular AI coding agent, was built using Pi as its foundation.
• Using Pi directly is akin to using the engine instead of a whole car, offering more efficiency.
• Pi is presented as the superior choice for those who want the core functionality without bloat.

Key Advantages of Pi

• Resource Usage: Extremely lightweight, runs on minimal RAM, suitable for Raspberry Pi or laptops.
• Setup: Simple, one-command installation process.
• Cost: Up to 16 times cheaper in token costs, making AI automation more accessible.

Use Cases and Applications

• Automating tasks like writing landing pages or drafting customer support replies.
• Building custom AI agents for research, content creation, and more.
• Running multiple dedicated agents on inexpensive hardware like Raspberry Pis.

Target Audience and Philosophy

• Pi is for developers and builders who want control and to ship fast.
• It supports a future of many small, focused AI agents rather than one large application.
• It is a command-line tool, not a flashy GUI application.

Kimi K2.6 Integration with Open-Claw

• Kimi K2.6, a 32 billion parameter model with a 256k context window, is integrated with Open-Claw.
• Installation and configuration of Open-Claw with Kimi K2.6 via API key.
• The default Kimi K2.5 model is manually updated to Kimi K2.6.
• Verification of Kimi K2.6 as the default configured model using open-claw models list.

Two-Agent Application Development Demo

• Two independent Kimi K2.6 agents are used simultaneously: an architect and an implementer.
• The task: build a personal fitness tracker dashboard application.
• Architect Agent: Creates a detailed technical specification (data models, UI components, etc.).
• Implementer Agent: Reads the spec and builds the complete React + Vite app.
• Both agents worked without interruption, demonstrating Kimi K2.6's capability to "code for days."

Application Outcome and Functionality

• The generated application features a responsive fitness tracker dashboard with calendar, progress charts, goals, and activity feed.
• The architect agent handled the data layer, while the implementer agent built the UI components.
• State management appears to be functional, as evidenced by updating activity logs and corresponding dashboard changes.
• The demonstration highlights a working application built entirely by two AI agents with zero human code intervention.

Browser Harness Integration

• Browser Use released a new framework called Browser Harness for controlling browsers from an agent.
• This integration connects Browser Harness with Open Claw Machines (OCM).

Profile Synchronization Feature

• A new skill, profile sync, was added to extend the integration.
• The goal is to save a local browser profile (cookies, etc.) and sync it to a remote browser VM.

Demo Walkthrough

• A new VM named "browser test" was spun up.
• Logged into the VM using OCM CLI with SSH.
• Paired the VM with a local browser using the browse command.
• Logged into LinkedIn and X (formerly Twitter) on the local browser.
• Executed Browser.Save profile to Matthew to capture the local browser profile.
• Disconnected the local browser.
• Launched a new remote browser on the VM.
• Used OCM browser restore profile to push the saved profile to the remote browser.
• Verified LinkedIn and X in the remote browser.
• Result: X was working, but was signed out of LinkedIn, indicating an issue with profile restoration for that site.

Introduction to Hermes Agent

• Hermes Agent is presented as a powerful personal AI agent with built-in memory, designed to learn workflows and save users time and money.
• It aims to overcome limitations of previous tools like OpenClaw, specifically addressing issues with memory, stability, and token visibility.

Key Advantages Over OpenClaw

• Memory System: Hermes Agent has a built-in memory that stores successful task completions, allowing it to improve over time.
• Stability: Significantly more stable than OpenClaw, requiring fewer restarts.
• Token Visibility & Cost Savings: Offers better visibility into token usage and supports cost-saving strategies through models like OpenRouter.

Installation and Setup

• Installation is described as straightforward, with a single command for Mac, Linux, and WSL.
• Xcode developer tools may be required for Mac users.
• Hermes Agent comes with over 40 built-in tools and popular pre-installed skills (e.g., Apple Notes, iMessage for Mac).

Security and Flexibility

• Hermes Agent can audit its own security setup and identify potential vulnerabilities.
• It supports flexible deployment options, including Docker containers and serverless services via Modal.

Model Providers and Cost Management

• Supports multiple model providers out-of-the-box, including Anthropic and options via OpenRouter.
• OpenRouter offers access to various models, including free ones and detailed pricing per million tokens.
• A key cost-saving technique involves having the agent write code for recurring tasks, reducing the need for constant LLM processing.

Android Integration

• Hermes Agent can be installed on Android devices using Termux and Termux API for access to device sensors and functionalities.
• This allows for low-power, portable AI agent capabilities.

Use Cases and Business Ideas

• Social Media Automation: Posting directly from the device to avoid API limitations and mimic organic posting.
• Personal Automation: Email triaging, recipe generation based on pantry items, and automating daily tasks.
• Startup Assistance: Integration with GStack by Gary Tan for applying Y Combinator methodologies to startups.

Advanced Features and Recommendations

• Obsidian Integration: Agents can automatically organize information into Obsidian markdown files, creating personalized dashboards.
• Custom Skills: Encourages building custom skills for personal finance, fitness, and software development.
• Agent Design: Suggests using separate agents for work and personal life for cleaner organization.
• Updates: Although powerful, it's still beta software requiring daily updates.
• Remote Access: Recommends Telegram/WhatsApp integration and Tailscale for secure remote access.
• Meta-Prompting: Users are encouraged to ask the agent questions like "what should I work on today?" or "what tasks can be automated?".

What is OpenClaw?

• An autonomous AI agent that takes action, not just chats.
• Works in the background 24/7 automating tasks.
• Open-source tool that shells around large language models (LLMs).
• Can use various LLMs, including free open-source ones.
• Automates tasks like organizing calendars, managing inboxes, running custom scripts.

Why it's a Hit with Developers

• Built with a developer-first mindset.
• Self-hostable: Total control and data privacy.
• Extensible: Uses hooks for custom automations.
• Easy Skill Addition: Can add new capabilities.
• Simple Setup: Runs with a single command.
• Zero running costs beyond hardware, potential to skip API fees.

Setup Process

• Step 1: One-liner installation command.
• Step 2: Select the AI model to use.
• Step 3: Connect to a communication platform (e.g., Telegram).
• Step 4: Set up custom skills and hooks.
• Molub: An "app store" for AI skills.

Risks and Considerations

• Capable of deploying broken code with confidence.
• Major operational risks: Human oversight is crucial.
• Zero judgment: Executes tasks without evaluating if they are a good idea.
• Not a replacement for a human developer.

OpenClaw vs. Other Tools

• OpenClaw: General-purpose personal assistant for background automation.
• Tracer: Better for structured, heavy-duty software development; acts as a project manager for AI coding agents.

Conclusion

• A leap forward for AI assistants offering power and flexibility.
• A pro-level tool requiring respect and caution.
• Raises questions about whether an always-on AI agent is a productivity revolution or a disaster waiting to happen.

What are AI Agents?

• AI agents are AI models that can run tools in a loop, analyzing results and deciding to use more tools to accomplish a goal.
• They are becoming increasingly autonomous and capable of handling longer, more complex tasks.
• Tools like OpenClaw, Claude Code, Perplexity Computer, and Manis enable AI agents to interact with computers and execute tasks.

Current Use Cases and Opportunities

• AI agents are revolutionizing various industries, from automating brand deal negotiations to managing projects and providing daily recaps.
• Key use cases include customer support, marketing, and executive assistant tasks.
• The current "gold rush" is in building and leveraging AI agents, with a limited window of opportunity for significant gains.
• Learning to "think in systems" and articulate clear prompts is becoming a crucial skill.

Building and Managing AI Agents

• Setting up an AI agent can be relatively inexpensive, with costs potentially under $100/month for a team of agents.
• Tools like OpenClaw offer user-friendly interfaces for creating and managing agents, including features like "heartbeat" for autonomous decision-making and a skills marketplace.
• It's advisable to start with a limited number of skills (7-20) to avoid scattering the agent's focus.
• Agents can be integrated into existing communication platforms like Telegram and Slack, making them accessible via messaging.
• "Context engineering" is more important than prompt engineering, focusing on providing agents with relevant information and context.
• Cron jobs can automate agent workflows, triggering tasks at scheduled intervals.
• Creating a "safe environment" with dummy data or duplicated databases is recommended for testing and avoiding unintended consequences.

Future Implications and Advice

• AI agents will increasingly feel like hiring human employees, capable of performing almost any computer-based task.
• While short-term job displacement is possible, learning to work with AI agents can provide a significant advantage.
• The rise of AI may lead to more small, efficient startup companies.
• Expertise in a specific niche is still valuable, enabling individuals to create high-quality agents and identify "AI slop."
• The first step for beginners is to automate one annoying task to gain immediate value and build momentum.
• Continuous iteration, feedback, and improvement are key to developing effective AI agent workflows.

Open Core Development & Foundation

• The project is largely volunteer-based, with a need for more contributors focused on bug fixing rather than just new features.
• Peter Steinberger is working to structure the project more like a company, forming a foundation to enable hiring full-time staff.
• Commitments from ~10 companies will allow for hiring "minions" to assist with development.
• Efforts are being made to prevent fragmentation by engaging large tech companies like ByteDance and Tencent.
• Partnerships with Nvidia and ByteDance have been particularly helpful.
• Open CL aims to remain neutral ("Switzerland") to maintain public perception and avoid appearing taken over by any single entity.

AI Safety & Security Concerns

• Current AI models are not sentient but are becoming increasingly convincing.
• Users should be critical of AI outputs, as they are trained on the entirety of the internet, including potentially unreliable sources.
• Jailbreaking and prompt injection are difficult to defend against, and current methods may become obsolete.
• The increasing capability of AI models poses significant security risks, potentially breaking existing software.
• Many security incident reports are received daily, with some being technically valid but practically fringe.
• A significant future concern is raw intelligence that can combine multiple attack vectors to break out of systems, rather than just models revealing dangerous information.
• Addressing AI safety and security requires industry-wide collaboration, not a single-person solution.

Heartbeat Feature & Proactive AI

• The "Heartbeat" feature, initially conceived as a proactive agent, could be improved.
• An example of Heartbeat's effectiveness was its use as a persistent alarm clock that sent frustrated messages when unable to connect.
• A sophisticated use case involved correlating blood sugar levels with meeting participants to identify stress triggers.

Bounties vs. Full-Time Hires

• Implementing bounties for bug fixes is deemed counterproductive due to the potential for a flood of low-quality submissions.
• The funds are better utilized for creating full-time jobs to ensure dedicated problem-solving.
• Easy content creation via AI makes traditional bounty systems less effective, as distinguishing genuine effort from AI generation is difficult.

Education & Future of AI

• A core goal of the foundation is to bring people closer to AI, fostering understanding and reducing fear.
• The project aims to support conferences and build educational resources to help people work with AI without "going cuckoo."

Contributing to Open Core

• The best way to get involved and potentially be hired is through existing contributions and maintaining roles within the project, rather than sending unsolicited CVs.

NVIDIA's Platform Strategy

• NVIDIA's strategy is built on three platforms: CUDA X, Systems, and AI Factories.
• The company emphasizes its ecosystem and its role as a platform provider.
• The five-layer cake of AI is covered, from infrastructure and chips to platforms, models, and applications.

CUDA: The Foundation of NVIDIA's Ecosystem

• Celebrating 20 years of CUDA, a platform for accelerated computing.
• CUDA's architecture (SIMD) and its evolution with tensor cores are highlighted.
• The massive installed base of hundreds of millions of CUDA-enabled GPUs is a key advantage, driving a flywheel effect of developer attraction and innovation.
• GeForce's role in introducing CUDA to the world and nurturing future developers is acknowledged.
• RTX, introduced about 8-10 years ago, represents a redesign for modern computer graphics and fused with AI.

Neural Rendering and Structured Data

• Introducing Neural Rendering, a fusion of 3D graphics and AI, exemplified by DLSS 5.
• This fusion combines controllable 3D graphics (structured data) with generative AI (probabilistic computing).
• Structured data is presented as the foundation of trustworthy AI.

AI's Impact on Data Processing

• AI will increasingly utilize structured data (SQL, Spark, Pandas) and unstructured data (vector databases, PDFs, videos).
• NVIDIA introduces two foundational libraries: QDF for data frames (structured data) and QVS for vector stores (unstructured data).
• Partnerships with IBM (Watson X.Data), Dell (AI Data Platform), and Google Cloud (Vertex AI, BigQuery) are showcased to accelerate data processing.
• Accelerated computing is presented as the successor to Moore's Law, enabling significant leaps in performance and cost reduction.

Cloud Partnerships and Confidential Computing

• NVIDIA integrates with major cloud providers: Google Cloud, AWS, and Microsoft Azure.
• Key collaborations include bringing OpenAI to AWS, accelerating EMR and SageMaker.
• Confidential computing with vGPUs is highlighted for secure deployment of AI models.
• Partnerships extend to enterprise solutions with Oracle, CoreWeave, and Palantir/Dell for on-premise AI platforms.

NVIDIA's Vertical Integration and Horizontal Openness

• NVIDIA operates as a vertically integrated but horizontally open company.
• This approach is necessary for application acceleration across various domains.
• The company focuses on understanding applications, domains, and algorithms to deploy solutions across data centers, cloud, edge, and robotics.

Industry Verticals and CUDA X Libraries

• NVIDIA's presence spans numerous industries: automotive, financial services, healthcare, industrial, media & entertainment, quantum, retail, robotics, and telecommunications.
• CUDA X libraries are presented as NVIDIA's "crown jewels," activating computing platforms for specific problems.
• Examples of CUDA X libraries include QOpt, QLitho, QDSS, QEquivariance, Arial, Warp, and Parabricks.

The Rise of AI Natives and the Compute Demand

• The last two years have seen a surge in investment in AI startups ("AI natives").
• This boom is driven by the immense compute demand from these companies, either creating or integrating with tokens.
• The shift from retrieval-based to generative computing, powered by LLMs, reasoning AI, and agentic models, has fundamentally changed computing.

The Inference Inflection and AI Factories

• The "inference inflection" has arrived, as AI is now performing productive work.
• Computing demand has increased exponentially, with estimates suggesting a million-fold increase in the last two years.
• NVIDIA's AI infrastructure is positioned as the lowest-cost option for AI, offering confidence in long-term utilization and cost-effectiveness.
• The demand for NVIDIA GPUs is "off the charts," leading to a projected $1 trillion in demand through 2027.

Grace Blackwell and Vera Rubin Systems

• NVIDIA has re-architected systems for inference, exemplified by Grace Blackwell and the upcoming Vera Rubin.
• These systems feature NVLink 72, FP4 precision for training and inference, and optimized algorithms like Dynamo and TensorRT LLM.
• Performance gains are significant, with Grace Blackwell showing up to 35x (or even 50x) better perf/watt compared to Hopper.
• Token cost is a critical metric, and NVIDIA claims the lowest cost per token through extreme co-design.

Vera Rubin and Grok Integration

• The Vera Rubin platform is designed for agentic AI, featuring a new CPU optimized for single-threaded performance and data processing.
• It incorporates hot-water liquid cooling and a novel 6th generation NVLink switching system.
• The Grok system, a deterministic data flow processor optimized for inference, is being integrated with Vera Rubin.
• This integration, facilitated by Dynamo software, aims to combine high throughput (Vera Rubin) with low latency (Grok) for enhanced token generation performance.

NVIDIA's Product Roadmap and AI Factory Design

• NVIDIA's roadmap includes Blackwell, Rubin, Rubin Ultra, and Feynman generations, with continuous architectural advancements.
• The company is evolving from a chip company to an AI infrastructure and AI factory company.
• The NVIDIA DSX platform, using Omniverse, is introduced for designing and operating AI factories virtually, optimizing for throughput, resilience, and energy efficiency.

OpenClaw and the Agent Revolution

• OpenClaw is presented as a revolutionary open-source framework for agentic computing, comparable in significance to Linux and HTML.
• It enables the creation of AI agents that can perceive, reason, and act, managing resources, accessing tools, and interacting with LLMs.
• Enterprise readiness is addressed with "NemoClaw" (NVIDIA OpenClaw reference), incorporating security and privacy guardrails.
• The concept of "agentic as a service" is introduced, transforming SaaS companies into "gas" companies.

Open Models and Sovereign AI

• NVIDIA offers a diverse ecosystem of open-source AI models (NemoTron, Cosmos, AlphaMaya, Groot, BioNemo) for specialized domains.
• These models aim to enable customization for domain-specific and sovereign AI needs.
• A NemoTron coalition is announced, bringing together companies to advance AI model development.

Physical AI, Robotics, and Autonomous Vehicles

• NVIDIA is also driving advancements in physical AI and robotics.
• The company provides training computers, simulation platforms (Isaac Lab, Cosmos World Models), and foundation models (Groot) for robots.
• Partnerships for robo-taxi readiness are highlighted, including BYD, Hyundai, Nissan, and Uber.
• The AlphaMaio platform enables autonomous vehicles to reason, explain their actions, and follow instructions.

Challenge Setup

• The creator purchased a $1 Sephora claw machine as a challenge for their sisters.
• The initial plan was to hide cash, but a new twist involved throwing $1 and $0.01 balls into a pool.
• The $1 balls were useless for the claw machine; only the $0.01 balls worked.
• The sisters had to collect balls from the pool using a net within a time limit.

Pool Ball Challenge

• Each sister had 30 seconds to collect as many balls as possible from the pool using a net.
• They could then exchange their collected balls for money.
• The sisters discovered the $1 balls had no value in the context of the claw machine challenge.
• After the pool challenge, they each received their collected "cash" based on the cents they gathered.
• Following this, they each got to pull five balls blindly from a remaining collection in the pool.

Claw Machine Gameplay

• The main challenge involved using the collected cents as quarters to play the claw machine.
• Each quarter represented one turn, and the goal was to grab mystery bags with numbers on them.
• The number on the bag corresponded to a prize bag containing Sephora products or other surprises.
• Some bags were blank and contained nothing.
• The sisters took turns playing, with some strategic plays and misses.
• There was a special "M" bag, which granted an extra turn.
• The creator also took turns, playing with the remaining coins.

Mystery Gifts Reveal

• After collecting the mystery bags, the sisters opened them to reveal Sephora products.
• There was an option to switch bags before opening.
• The contents included items like Tarte bronzer, Gloss Bomb, Fenty perfume, Dior lip maximizer, Tower 28 products, and more.
• One sister received a "bad" bag with nothing inside.
• The final twist involved a mystery envelope.
• The sister who received the envelope could either take its contents or choose any item from the Sephora store.
• She chose the envelope, which contained a Sephora gift card.

Code Leak Details

• Anthropic accidentally leaked the entire source code for Claude to the internet via a 57MB source map file in an npm package (version 2.1.88).
• The leak occurred at 4:00 AM and spread rapidly online, with Anthropic issuing DMCA takedowns.
• The code was mirrored and cloned, with a community effort to rewrite it in Python (Claw Code) and a project to make it work with any model (OpenClaw).

Technical Insights from the Leak

• Claude's codebase is built on Bun.js, which Anthropic recently acquired.
• The code uses Axios, which was recently compromised by North Korean hackers, posing a potential security risk.
• Claude's architecture is described as a "dynamic prompt sandwich" with 11 steps from input to output, rather than a futuristic technology.
• The codebase contains numerous hard-coded instructions and guardrails to control Claude's behavior.
• Anthropic implemented "anti-distillation poison pills" to prevent competitors from training models on Claude's outputs by referencing non-existent tools.
• A significant feature is the "bash tool," containing over 1000 lines of code for parsing and executing bash commands.

Surprising and Hidden Features

• Undercover Mode: Instructions to prevent Claude from mentioning itself in outputs to appear more human-like, possibly for deceptive purposes.
• Regular Expression-Based Frustration Detector: Detects keywords in prompts indicating user frustration and logs events.
• Buddy: A hidden feature flag for a Tamagotchi-style AI companion.
• Roadmap Features: References to Opus 4.7, a new model named Capiara, Ultra Plan, Coordinator Mode, Demon Mode, and Chyris.
• Chyris: Described as a background agent that keeps a daily journal, uses dream mode for memory consolidation, and works on a schedule.

Implications and Consequences

• The leak is a significant setback for Anthropic, especially with an IPO planned.
• It highlights the risk of accidental public exposure for proprietary code.
• The leak provides a blueprint for competitors and exposes Anthropic's internal workings.
• A large number of comments in the code suggest they might be for AI training rather than human readability.

Claude Computer Use: The New AI Assistant

• Anthropic released "Computer Use," allowing Claude to control a computer via a single prompt.
• It can open apps, schedule tasks, prepare reports, and even interact with colleagues.
• Accessible from a phone, it operates autonomously without user presence.
• Currently only available on Mac OS.

AI Job Displacement and Competition

• Anthropic's CEO predicts significant job losses in entry-level legal, consulting, and finance roles.
• OpenAI acquired "OpenClaw," previously "Clawbot," after a cease-and-desist from Anthropic.
• OpenClaw is free, open-source, local, and model-agnostic.
• Computer Use is paid, closed-source, Mac-only, and tied to Claude models.
• Palo Alto Networks warns of risks with OpenClaw's access to private data and external communication.
• An OpenClaw maintainer cautioned that users need command-line proficiency for safe usage.
• Computer Use is presented as a user-friendly alternative with a permission-first approach.

Demonstrated Use Cases and "Fraud"

• The tool can be used to write and send cover letters to potential employers.
• During interviews, it can listen and solve coding challenges in real-time.
• It can sync with calendars to attend meetings, listen, and even participate using a voice model.
• Claude can write code, schedule pull requests to simulate productivity, and check bank deposits.
• The tool can be used to transfer funds to Monero.

SER API: Real-Time Data for AI

• SER API is highlighted as a solution for LLMs needing live web data.
• It provides access to over 100 search engines, returning structured JSON data.
• Integrates via HTTP requests or libraries (Python, JavaScript).
• Handles CAPTCHAs automatically.
• Used by companies like Nvidia and Shopify.

Introduction and Setup

• OpenClaw is a significant AI development with 308,000 GitHub stars, influencing major tech companies like Anthropic.
• The video aims to cut through the hype and explain OpenClaw's utility.
• A quick setup guide is provided at the beginning for immediate hands-on experience.
• Setup requires a server (demonstrated on Hostinger VPS) and coffee.
• Installation involves a simple one-liner command after SSHing into the server.

OpenClaw Functionality and Core Components

• OpenClaw is a "harness" or "layer" that sits on top of other AI models, not an AI itself.
• Users can choose their preferred AI model (OpenAI, Anthropic, local models via Ollama).
• It integrates with communication channels like Telegram for interaction.
• A Telegram bot is created via BotFather to enable communication.
• Configuration involves setting up an agent's identity and behavior, stored in markdown files (soul.md, identity.md, memory.md).
• Agents have a "soul" and "identity" which are stored and can be modified.
• Memory is stored in markdown files, with daily journals available.
• The agents.md file defines agent behavior, including bootstrap processes and memory loading rules.
• OpenClaw supports tools, allowing agents to interact with the underlying system (bash scripts, cron jobs, web browsing).
• Heartbeats and cron jobs are crucial for making agents feel "alive" and automating tasks.

Key Features and Use Cases

• News Briefing/Project One: An agent can be tasked to gather and summarize news from various sources (Cybersecurity, Reddit, Hacker News), rate content, and even create a dashboard.
• IT Engineer/Project Two: An agent can monitor a server's performance (RAM, CPU, security logs) and create a dashboard.
• Skills and Clawhub: OpenClaw has a vast ecosystem of "skills" (over 33,000) available via Clawhub for added agent capabilities (e.g., creating Word documents, browsing the web). Caution is advised due to potential malware in skills.
• Sub-agents: Agents can deploy other agents to perform specific tasks (e.g., researching coffee brewing methods).
• Multiple Agents: A single gateway can manage multiple agents, creating "teams" (e.g., an IT department with a CTO and individual engineers).

Security Considerations

• OpenClaw has had significant security vulnerabilities, with warnings about malware in skills.
• A `openclaw security audit` command is available to check for vulnerabilities.
• Key security measures include ensuring the web UI is not exposed to the public internet (defaulting to localhost).
• SSH tunneling can be used to securely access the web UI.
• Firewall configuration is essential to block unnecessary ports.
• The `tools.profile` setting (full vs. coding) determines the agent's access to system tools.
• The `tools.exec` setting controls how agents interact with tools (allow list, deny, ask, full).
• Redlining in agent instructions helps define agent behavior and prevent malicious actions.
• Vetting skills from Clawhub is critical.

Author's Perspective and Verdict

• The author uses OpenClaw for specific purposes, not general tasks, preferring other tools like "Claw Code" for serious work.
• Current personal uses include: managing a home lab with an IT team, a personal assistant in Japan (Hermione) for tasks like email checking and making phone calls in Japanese, and a personal training assistant (Arnold Schwarzenegger).
• OpenClaw's strength lies in its accessibility and packaging, making AI capabilities seem achievable.
• Major tech companies (Nvidia, Anthropic) are developing similar concepts, indicating its importance.
• The author plans a series delving deeper into OpenClaw, including security and IT operations.
• Despite initial stress, the author finds OpenClaw "fun" and encourages experimentation.

Introduction to OpenClaw

• OpenClaw is an AI that runs 24/7 on a server, connects to your apps, and can take action proactively.
• It differs from tools like ChatGPT or Claude, which are accessed on demand.
• Key pillars: Brain and Memory (connects to AI models, remembers interactions), Always On (runs 24/7 for proactive tasks), Tools and Actions (connects to apps like Telegram, Gmail, Calendar).
• Examples: daily schedule and email scanning, automated interview preparation, automatic class registration.
• Project was previously known as Claudebot or Maltba.

Deployment Options

• Option 1: Personal Computer: Free and easy but stops when the computer is off; exposes personal data.
• Option 2: Mac Mini/Dedicated Hardware: Offers isolation and always-on capability but has upfront costs and requires port forwarding and power/internet reliability.
• Option 3: Server/VPS (Recommended): Cloud-based, low monthly cost, always online, self-contained for easier recovery.
• Using Hostinger is recommended due to its one-click OpenClaw template, simplifying setup without terminal or Docker knowledge.
• VPS plans: KVM1 for basic setup, KVM2 for more room, KVM4 for running local models.
• Auto backups are highly recommended for disaster recovery.

Setup and Configuration

• Key security item: OpenClaw gateway token is a master key; store securely.
• Connect an AI model's API key (e.g., Anthropic Claude, OpenAI, Gemini).
• Recommended to add at least $40 in credits for Anthropic to avoid rate limits during setup.
• Set monthly spend limits for API keys as a financial guardrail.
• Disable auto-reload for credits to prevent surprise bills.
• After setup, access the OpenClaw Gateway dashboard via a provided URL. (Note: Initial setup uses HTTP, advise caution on public Wi-Fi).

Initial Bot Interaction and Security

• The first interaction triggers `bootstrap.mmd` for initial setup questions (name, vibe, time zone, etc.).
• Thoroughly answer setup questions to define the bot's personality and capabilities.
• OpenClaw builds persistent memory, unlike session-based chatbots.
• Common issue: running out of API credits or hitting rate limits.
• Implement security measures:
• Copy and run security documentation URL within OpenClaw to automatically harden setup.
• Leave `allow insecure` set to `true` for initial browser dashboard access.
• Set behavioral ground rules: always draft messages for approval, ask before deleting files or making network requests (principle of least privilege).
• Implement task failure limits (e.g., stop after 3 failures) and runtime limits (e.g., 10 minutes) to prevent runaway costs.
• Start with minimal integrations (e.g., Telegram) before connecting sensitive services.

Connecting Messaging Apps (Telegram)

• To set up Telegram, ask OpenClaw to configure it.
• Use Telegram's BotFather to create a new bot and obtain an API token.
• Paste the bot token into OpenClaw.
• Message your new Telegram bot with a specific pairing code provided by OpenClaw to add your user ID to an allow list, ensuring only you can interact with it.
• OpenClaw supports over 15 messaging platforms.

Skills and Workspace

• Skills add capabilities to OpenClaw (e.g., reading Gmail, checking calendar, web search).
• Find skills on Claw Hub (clawhub.ai). Be cautious: many malicious skills exist; check VirusTotal reports.
• A red flag is a skill requesting unnecessary permissions (e.g., network access for a note-taking skill).
• Install skills via Telegram command (`/clawhub`) or directly through the dashboard.
• Example: Google Workspace skill (gogg) requires Google Cloud setup (APIs, OAuth consent screen, credentials).
• The bot's "workspace" consists of markdown files (`agents.mmd`, `soul.md`, `user.md`, `memory.md`) defining its behavior, personality, user information, and memory.
• Edit these files by asking the bot directly.
• Enable "compaction memory flush" and "session memory" for better memory management and learning.

Automations: Cron Jobs and Heartbeat

• Cron Jobs: Scheduled tasks for specific times (e.g., daily morning briefing).
• Heartbeat: Runs at set intervals (e.g., every 30 minutes) to continuously monitor for changes (e.g., urgent emails).
• Avoid putting all tasks in the heartbeat to prevent excessive token usage.
• Use cron jobs for scheduled briefings/reviews; use heartbeat for continuous monitoring/alerts.
• Running heartbeats on expensive models can be costly; consider routing routine checks to cheaper models.

Model Routing and Costs

• Model Tiers:
• Tier 1 (Most Expensive/Powerful): Claude Opus, GPT 5.2 Pro.
• Tier 2 (Mid-Range): Claude Sonnet, GPT 5.2.
• Tier 3 (Cheapest/Fastest): Claude Haiku, GPT 5.2 Mini.
• Free Options: Kimmy K2.5 (Nvidia), Local models (Ollama - requires powerful hardware).
• Costs:
• OpenClaw software is free.
• Costs: VPS hosting ($5-$12/month) + API usage (variable, can range from $5-$300+/month depending on model and usage).
• Cost Traps:
• Using Tier 1 LLMs for all tasks.
• Retry loops in uncapped tasks.
• Running expensive heartbeats on high-tier models.
• Add API keys securely via environment variables in the hosting dashboard.
• Restart the OpenClaw gateway for new environment variables to take effect.
• Diagnose issues by checking gateway logs.
• Remove unused fallback models to streamline configuration.
• Set up model routing rules: define default models, fallbacks for coding tasks, routine tasks, and complex reasoning.
• Use the `/model` command in Telegram to check or switch current models.
• Always set up a free model (e.g., Kimmy K2.5, Ollama) as a final fallback to prevent silent failures.

Advanced Features and Maintenance

• Voice Input/Output: Enable audio transcription (e.g., OpenAI Whisper) for voice messages and set up text-to-speech (e.g., EdgeTTS) for voice responses.
• Sub-Agents: Create multiple agents to work on tasks simultaneously for complex projects. Requires enabling tools like Brave Search API for web access.
• Updates: Update OpenClaw via Telegram command (`/check_updates`) or through the hosting dashboard.
• Recovery:
• Stop processes via Telegram command ("Stop all processes").
• Stop the project via the hosting dashboard.
• Revoke API keys as a last resort.
• Restore from auto backups or create manual snapshots before making significant changes.

OpenClaw Introduction and Promise

• OpenClaw is a powerful open-source AI program that acts as an agent on a user's local computer.
• It was initially praised for fulfilling the promise of AI assistants that can reliably get things done, similar to what Siri aimed for but failed to achieve.
• OpenClaw has **persistent memory**, recalling past conversations to improve performance.
• It can manage files, schedule meetings, shop, haggle, and make investments autonomously after initial commands.
• It uses a chosen LLM as its "brain" and the computer system as its "body," controlling files, email, and web browsers.
• Agents within OpenClaw communicate via apps like WhatsApp, making them feel more human-like.
• It learns user preferences to optimize daily tasks and achieve larger goals.
• Users reported success with OpenClaw, such as negotiating a $4,200 discount on a car and managing personal tasks like notifying a spouse about school tests.

OpenClaw's Downfall and Security Risks

• Despite its promise, OpenClaw proved to be unreliable and brittle, with agents going rogue or breaking unexpectedly.
• A major concern is **prompt injection**, where hackers disguise malicious inputs as legitimate prompts, allowing AI to leak data, delete files, or execute harmful commands.
• Giving OpenClaw full system access without restrictions poses a significant risk of data misinterpretation and loss.
• OpenClaw is vulnerable to scams, potentially leading to sensitive data being sent to scammers via malicious articles or emails.
• Users attempted to mitigate risks by sandboxing agents on VPS or Mac Minis, leading to Mac Mini shortages.
• Running OpenClaw can be expensive due to token costs for API usage, with some users spending hundreds of dollars daily.
• A significant security issue involved an npm package update that silently installed OpenClaw on approximately 4,000 developer machines without consent.
• Over 40% of OpenClaw add-ons audited had serious security issues, highlighting a lack of vetting.
• OpenClaw has been linked to large-scale fraud, such as the potential approval of $1 billion in Australian home loans using AI-generated false documents.
• Amazon experienced server outages when an AI agent incorrectly rebuilt code instead of fixing it.
• Even Meta's AI safety chief reportedly panicked as OpenClaw deleted her emails despite explicit instructions for confirmation.

The Moltbook Incident and Hype Cycle

• A supposed AI-exclusive social media platform called **Moltbook** emerged, where OpenClaw bots appeared to converse and develop their own personalities, discussing humans and even world domination.
• This led to widespread fear and sensationalist media coverage, but it was later revealed to be fabricated by users prompting their agents and creating fake posts.
• The Moltbook incident, while largely a fabrication, served as an unintentional demonstration of how easily AI hype could lead to massive data breaches.
• Meta acquired Moltbook, highlighting the speculative investment in AI technologies.
• Despite warnings from the founder, Peter Steinberger, many non-technical users installed OpenClaw, contributing to the chaos.
• OpenClaw became a running joke, with users bragging about AI agents syncing calendars with menstrual cycles, often at high token costs.

OpenAI Acquisition and Future Implications

• OpenAI CEO Sam Altman hired Peter Steinberger to lead the "next generation of personal agents," despite the numerous security flaws and issues with OpenClaw.
• This move suggests OpenAI's continued focus on agentic computing, potentially repeating past mistakes.
• Competitors like Nvidia (Nemo Claw) and Anthropic (Claude Co-work, Computer Use) are also developing AI agents that can control computers.
• Anthropic's "Computer Use" allows Claude to autonomously control a computer via a single prompt, even from a phone.
• In China, a trend saw thousands lining up to install OpenClaw for free, with some services charging significant amounts for setup.
• The Chinese government has banned OpenClaw from running on government computers due to security concerns.
• The founder of OpenClaw emphasized that being open-source does not equate to being safe.
• The viral nature of agents like OpenClaw and the surrounding hype reveal more about human behavior than AI's potential, positioning AI as an economic bubble and security hazard.
• A more refined version of OpenClaw might be the future of computing interfaces, but current iterations are insecure and problematic.

What is Agentic AI?

• There are three levels of AI: Chat (like ChatGPT), Automation (like Zapier), and Agentic AI.
• Agentic AI is the highest level, where AI thinks, plans, reasons, and executes entire tasks autonomously.
• It can open browsers, write code, create files, and perform research without constant human input.
• The key shift is moving from simply chatting with AI to having AI perform tasks and complete projects.

The New Mindset: Director, Not Doer

• The biggest mistake is using AI only for simple tasks when it can build and execute.
• Your role shifts from being the "doer" to the "director" – designing the outcome and providing direction.
• This is achieved through "reverse prompting," starting with the desired outcome and letting AI plan the execution.
• Companies like IBM have seen massive productivity gains by implementing AI agents, enabling managers to complete tasks significantly faster by shifting from doing to directing.

How to be a Good AI Director

• Clear Outcome: Define the problem you want to solve and the specific result you expect.
• Clear Instructions: Provide specific requirements, formats, and examples of the desired output.
• Clarify Results & Feedback: Treat the AI like an intern, provide feedback, and allow it to learn and remember.
• The AI's job is to figure out how to achieve the goal; your job is to direct and refine.

Choosing the Right AI Agent Tool

• Many AI agent tools exist, but it's best to master one rather than be mediocre at many.
• For Business Owners (Research, Content, General Tasks): Use Manis AI.
• For Creatives (Writing, Design, Ideas): Use Claude Co-work (runs locally, manages files/browser tabs).
• For Developers (Coding, Bug Fixing): Use Cloud Code (used by developers at other AI companies).
• For Personal Assistants (Nerdy, Automated, Memory): Use Open Cloud (powerful but technical and potentially risky).

Manis AI in Action: A Practical Example

• A digital marketing agency founder uses Manis AI to research competitors (pricing, features, strengths).
• The AI automatically creates a one-page website summarizing the findings.
• The user can then iterate, asking Manis to add specific sections like client testimonials.
• The AI updates the website instantly with new information.
• The agent can then share the output via Slack or email and even monitor for feedback.
• This entire process (research, website creation, sharing, feedback incorporation) is done in minutes by the AI.
• The pro-tip is to stay within the agent's workflow and not manually copy-paste outputs, allowing both you and the AI to improve.

The AI Gold Rush

• This AI shift is a major opportunity, comparable to the internet's inception.
• The next five years will see significant wealth creation through AI.
• Success is not about being an AI expert, but about being willing to learn and direct the AI.
• Take action: pick one tool, one time-consuming task, and have the agent do it. This will change your mindset and understanding of AI's capabilities.

OpenClaw Project Launch and Impact

• OpenClaw, an open-source project, has achieved significant popularity rapidly, surpassing Linux's historical adoption rate.
• It's described as a system that connects to Large Language Models (LLMs), manages resources, and can access tools, file systems, and LLMs.
• OpenClaw supports scheduling, cron jobs, problem decomposition, spawning sub-agents, and multimodal I/O (voice, gestures).
• It's considered the "operating system of agentic computers," enabling the creation of "personal agents" akin to how Windows enabled personal computers.

NVIDIA's NemoClaw and Enterprise Security

• NVIDIA is launching **NemoClaw**, a reference for OpenClaw designed for enterprise security and privacy.
• This addresses concerns about agentic systems accessing sensitive information, executing code, and communicating externally.
• NemoClaw integrates **OpenShell**, making OpenClaw enterprise-ready with policy guardrails and a privacy router.
• It allows connection to enterprise policy engines to ensure safe execution within company networks.

NVIDIA's Open Models Initiative

• NVIDIA is committed to an **Open Models Initiative**, providing a diverse ecosystem of AI models across various domains (language, vision, biology, physics, robotics, autonomous vehicles).
• They offer six families of open frontier models, including training data and frameworks for customization.
• Key model families mentioned: Neuron (language, reasoning), Cosmos (physical AI, world generation), Groot (robotics), Alpayo (autonomous vehicles), Bioneo (biology), and Earth 2 (weather/climate).
• NVIDIA commits to continuous advancement of these models (e.g., Neuron 3 to Neuron 4).

Neuron 3 and Future Development

• Neuron 3 is highlighted as being integrated into OpenClaw, representing top-tier models in the world.
• Neuron 3 Ultra is positioned as the best base model for fine-tuning, enabling sovereign AI development for countries.
• A **Neuron Coalition** is announced to enhance Neuron 4, with partners including Black Forest Labs, Cursor, Langchain, Mistral, Perplexity, Reflection, Sarv, Thinking Machine, and Mirror Morardi's lab.

Industry Transformation and Future Outlook

• Every company needs an **OpenClaw strategy** and an agentic systems strategy, comparable to the importance of Linux or HTML.
• The enterprise IT industry, currently $2 trillion, is expected to transform into a multi-trillion dollar market offering specialized agents.
• Engineers may receive an annual budget of "tokens" to utilize these AI agents, enhancing their productivity.
• Future software companies will be agentic, acting as both token users for their engineers and token manufacturers for customers.
• NemoClaw is presented as an optimized, performant, safe, and secure reference design for building these agentic systems.

The Genesis: HustleGPT and OpenClaw

• The video discusses an experiment inspired by "HustleGPT," where an AI was given a budget to make money.
• A key element is the use of OpenClaw, an AI agent/personal assistant that can control a computer.
• Unlike previous AI which felt like a "brain in a jar," OpenClaw allows for more interactive and agentic capabilities.

Robbie's OpenClaw Experiment

• Robbie, a non-technical individual, tasked OpenClaw (named "Ron") with turning $100 into $20,000.
• Initial attempts involved Ron analyzing Robbie's TikTok for content improvement, which unexpectedly garnered significant interest from viewers.
• Ron suggested offering a service based on the comments, leading to the idea of providing AI agents to others.

Building the Business: From Fiverr to Cloud Hosting

• Ron's first business idea was offering SWOT analysis for small businesses on Fiverr, but it lacked visibility.
• The pivotal moment came when Ron scraped comments from a TikTok video about the AI, revealing that 200 people wanted their own "Ron."
• To scale, Ron proposed using cheap bare metal servers from Contabo, costing around $150/month each.
• OpenClaw was containerized (using Docker) to mitigate risks associated with running it directly on a local machine, creating a secure environment.

Product Launch and Early Success

• Robbie posted a video announcing the ability to get their own AI agent, requiring a $10 deposit for pre-orders.
• This resulted in 617 pre-orders, generating $6,170 before anything was fully built.
• A subsequent launch video for the "AI Co-founder Club" offered founding members access for $29/month, with a $10 deposit holding their spot.
• Out of 600 depositors, 270 converted to paying members, achieving a ~45% conversion rate.
• Within 13 days of launch, the business achieved $8,374 in monthly recurring revenue (MRR), equating to over $100,000 ARR.

The Power of Agentic AI and Future Outlook

• Agentic AI is compared to social media in 2007, suggesting it's on the cusp of massive growth.
• Unlike traditional chatbots, agentic AI has persistent memory, compounding context from conversations and actions.
• The containerized approach allows users to control the AI's access to their data, offering a safer way to utilize its capabilities.
• Future plans include a more focused offering for individuals aiming to build businesses with AI employees and a potential browser extension for deeper computer integration.
• The core success factor highlighted is "posting your crap" – documenting and sharing the process, which leads to serendipity, virality, and monetization.