← Lenny's Reads
Lenny's Reads
Listen: OpenClaw: A power user’s guide to the most powerful personal AI tool since ChatGPT
00:24:31
Loading summary
Transcript1 lines
- [00:00]
A
Welcome to Lenny's Reads, where I bring you audio versions of my newsletter about building product, driving growth and accelerating your career. OpenClaw is probably the single most important release of software probably ever. That's a quote from Jensen Huang, the Nvidia CEO. Building on our podcast episode together, Clairvaux has put together a Power user's guide to OpenClaw which includes all of her best advice for getting started. It's the most in depth, practical and fun stuff by step guide on openclaw. You'll find you'll find all of her best pro tips both for beginners and power users. A big thank you to Peter Steinberger, Dave Morin and Nat Eliason for reviewing this episode. For more from Claire, check out her podcast How I AI Her AI Product Chat PRD and find her on X and LinkedIn. Links to these will be in the show notes. The rest of this episode was written by Claire and narrated by me let's get into it at 6am Before I've looked at my phone, an AI agent named Polly has already read my email, checked my calendar and queued up my day. By the time I sit down for coffee, another agent has reminded my husband about Spirit Day at school. A third AI is halfway through drafting a sales email which will land in our prospect's inbox three minutes after they contact us. None of this existed three months ago. If you're paying attention to AI news, you've probably heard about OpenClaw. For a while it seemed like X was nothing but lobster emojis and breathless posts about everything it can do, running your business, buying cars, planning the AI uprising with its friends, and more. You've also probably seen some horror stories, like when it started deleting a user's full Gmail inbox, or that time it completely screwed up my own personal calendar. You're interested and a little scared. I was too. But the idea of a dedicated personal assistant that could help run my life and businesses was so appealing that I had to dig in. The more I play with openclaw, the more convinced I am that it is one of the most powerful AI tools for personal use and a sign of where these tools are going. Fast forward two months, I'm chatting 247 with my nine and counting OpenClaw agents which operate my business as write code, close sales deals and make sure I get to my kids basketball games on time. Openclaw is powerful, but a little tricky to set up. After a lot of trial and error, I figured out how to make it work and make my claws work for me. Here's exactly how you can too. First, what is OpenClaw? OpenClaw is an open source personal AI assistant that is more powerful, more autonomous and more fun to use than anything else. I've tried to tell it what needs to be done. You can message it easily on the platforms you already use, like WhatsApp, Telegram, Slack and others. Then it can control your computer and take care of tasks just as you would working on its own schedule. Even overnight. It's always on runs locally and it can build its own new skills. Teach it once it handles the rest. Practically this means I can text my openclaw something like let's make sure our website always has the latest reasons why we're better than competitors. It will use Web search, our GitHub, repo and public APIs to find the information it needs and ship PRs with the updates. Every week it will refresh those pages with new data based on new features or market news. One text turns into an always on agent, but how you set it up is important and understanding a few key concepts will save you tens of hours of frustration. Here are some key concepts about openclaw to understand. This unique combination is what makes it so special and it runs a local gateway that takes in messages. Think of this as a general purpose inbox that can receive instructions from any channel such as Terminal, Telegram, WhatsApp and others. Behind this gateway are agents which have their own identities, tools and workspaces. Agents work on a scheduled set of cron jobs and a heartbeat that gets checked every 30 minutes. It can use and self install skills, APIs and CLIs to interact with systems and the outside world. It's deployed on an owned machine like a Mac Mini or a VPS on the cloud that you own and operate yourself. I've written in detail about the mechanics of openclaw, which you can find linked in the show notes. The OpenClaw Docs site is also very helpful, especially the top level pages on channels, agents and tools. But enough theory. Let's walk through the setup of your sentient virtual lobster. What computer to use and do you need a Mac Mini? You have options for where to install your OpenClaw, but most importantly do not install it on a work or personal computer that's actively in use. This is very dangerous. Openclaw can technically have access to all the files on the computer it runs on, and no matter how careful you are, you don't want to risk deleting everything or emailing your personal files to an unsavory character. The openclaw team has done a great job of hardening security, but it's best to start with an isolated box. You have three safe options for installing OpenClaw. The first option is to sign up for a hosted version of OpenClaw. Startups are popping up every week that make this easy to do. Some of the more popular options include starclaw, myclaw, simpleclaw, Uniqlaw and Everys one. I experimented with a few of these and they are slick, but I always got stuck on something. These products will get better and I expect this to be how most people will experience OpenClaw over time. The second option is to have it run in a virtual private server. This may be the cheapest option, but it's also the most complicated. Some of the more popular options include railway, Hostinger, DigitalOcean, Google Cloud, and Render. I didn't try going this route, but many technical people prefer it because it's quick, powerful and doesn't require new hardware. The third option is to use a laptop or yes, a Mac Mini. I started with an old MacBook Air and eventually moved to a stack of Mac Minis. You don't actually have to run it on a Mac Mini. You can use any computer, even an old laptop. But the Mac Mini is simple, powerful and compact and it's kind of become a meme. This option is the most expensive, at least up front, and time consuming, but also the most fun, educational and cute. If you can swing it, I'd try this I got the lowest end model M4 chip 16 gigabyte memory and 256 gigabyte storage. About $600. Just remember to grab a keyboard, mouse and a monitor for your initial setup. Use whatever you have lying around. You won't need it again after everything is running your pre work okay, you've got your machine booted up before we install the claw. Here are a few quick things that will make setup easier. This will take you about 10 minutes. Set up a fresh admin account and password on the computer you'll be using. Sign up for a Gmail address for your agent. Later you can give it read only access to your calendar. Make sure Chrome is installed, which is OpenClaw's preferred browser. The next step is to open the terminal and install OpenClaw. Open the terminal by pressing command space, type terminal and hit return. You then need to run a specific curl command which you can find in the show notes. This will install everything you need and drop you into the onboarding. If you get stuck, install claude code or codecs and ask for help. Once the install is complete, OpenClaw will walk you through a guided onboarding flow step by step. The first step is acknowledging a security warning. Read this if you're ready for the adventure. You can continue through the onboarding. Here are my pro tips for the key steps. Know how to navigate the terminal if you're not familiar with working in the terminal, the onboarding process is a little confusing. Use arrow keys to navigate up, down, left and right spacebar to select and enter to submit, pick your default models. I Recommend Claude, Opus 4.6 or codecs 5.4 or whichever model is the most powerful. At the time you're listening to this, authenticate your model provider. You have two options for using paid models like Opus or Codecs. First, use an existing subscription by logging into your Claude or ChatGPT account. Note there are some rumors that Anthropic is banning people who reuse their Claude accounts for OpenClaw when using any third party account with OpenClaw. Review the terms of service and proceed at your own discretion. Second, use an API key, which you can get by setting up a Claude or OpenAI developer account. This is the recommended path and what I use Choose a channel to talk to your agent. The best beginner friendly channel is Telegram Set up Web Search. This will give your claw access to the Internet. You can skip and set this up later when you need it. Install skills from a list of bundled defaults. I recommend the GOG skill for Gmail, calendar and docs access and summarize to Start Enable hooks. These are helpful tools that keep your open Claw setup optimized. I turned on all four of these, but session memory is the most important. The others are helpful when you want to debug or optimize your agent. Then your agent will hatch in the terminal UI hello world. Next, you need to set up Telegram. You'll want a better chat interface than the terminal, so open up your phone and download Telegram, a messaging app and the simplest way to connect with your openclaw. If you ask openclaw, it can walk you through the setup steps including messaging the bot Father yes, really? Here's some advice for your first chat Setting up your agent is where you should put on your manager hat just like an employee, it can't be good at everything, so think about a specific job for your open Personal Assistant Social Media Manager Engineering Intern Start with one idea and you can always add on more later. If you're unsure where to start, begin with a personal assistant like I did. This is where things get fun Once your openclaw is hatched, it will start asking you about yourself and itself. To build its definitions and operating model, you should make sure to share your name, your role or job, and common admin challenges in your life such as scheduling, remembering tasks, and coordination with your family. All this gets written down by the agent and stored in the workspace folder you set up in Onboarding as markdown files. Those files become your agent's identity and operating system. They are read every time your openclaw starts up and give your agent everything it needs to do a good job. It's fun to look at these files. Occasionally you may need to go in and edit them. The key files include agents MD openclaw's core set of instructions in memory Sol MD your agent's Persona, tone of voice and clear boundaries Identity MD your agent's name, vibe and personal emoji Tools MD Notes on tools and how your agent should use them User MD all about you your openclaws human Once its identity and role is set, it should ask you to get started on a first task. Time to claw okay, so you have this bot set up. What do you actually do with it? Here are six easy and useful openclaw workflows. I do a lot with my openclaw agents now, but below are some easy places to start. Just copy and paste these prompts and openclaw will do the rest. Coordinate a busy weekend here's the prompt Every Friday group Message me and my husband to confirm the kids weekend activity logistics. If there is a conflict, confirm who will pick up each kid and any adjustments we may need to make to lunch or dinner plans. Ask us to explicitly confirm the plan and update our shared family calendar. Second workflow Find trending topics and generate memes for social media. The prompt is Every morning, search for trending Reddit topics on R funny and r technology about product management. Use the Memelord API to generate one image and and one video meme. Send it to me for approval, then post to TikTok. Third workflow just in time Meeting Prep here's the prompt Check all my calendars for any meetings starting in the next 30 minutes. If there is a meeting starting soon, send me a pre meeting brief on telegram. Include meeting title and time, who's attending and agenda if available from the calendar, event description and our last interaction based on the most recent email thread or meeting notes with this person or company. Fourth workflow Write support docs overnight the prompt Every Friday evening, look over our Resolve support tickets if any question has been asked three or more times this week. First, flag it as a docs or FAQ candidate. Second, create a linear issue and assign to an agent to add a docs page covering the answer. Third, include the standard answer you've been giving as a starting point in the issue description. Fifth, Workflow look through PLG signups for high Value enterprise prospects Enrich and follow up here's the prompt Every morning, look through the signups for the last 24 hours. Find everyone who signed up with a company domain and categorize into high value prospects based on our ideal customer profile. For ones with less than 1,000 employees, send a light touch email based on SalesPlaybook MD. For companies with more than 1,000 employees, enrich their profile with Exit People API and confirm with me before sending. 6th Workflow Project manage Me here's the prompt I have a project launching on a specific date. Keep a to do list with everything I tell you I need to do for a successful launch and break the plan down into daily tasks I can easily get done at the end of the week. Celebrate what I accomplished and flag what I missed. Make sure we do everything possible to make this project successful. Here's the ultimate openclaw Unlock Running multiple Agents with Specific Roles I love how flexible OpenClaw is. It feels like it can do almost anything, but I found that you shouldn't try to get one agent to do everything just like me. You can set up many openclaw agents on the same machine. Just run openclaw agents, add then the new agent name in the terminal and you'll get put through onboarding again for a fresh agent. This agent can have a completely separate identity, set of tools, crons and workspace, which means you can treat it like a completely different employee. Multi agent setup was the unlock for me when using OpenClaw. Instead of trying one bot to do everything, I created a full team of openclaws to do different jobs around my work and business with a narrower identity. The bots did a better job and were more fun to work with. You can even ask one agent to spawn another. Try something like hey Bob, I just set up Annie the marketing intern. Transfer everything in your soul, memories and crons about marketing to her workspace and erase it from yours. Your claws can perform brain surgery. I'm now running a whole team of agents across my life. In business, I don't find it much different than managing a remote team, so I'm able to get a lot of leverage out of my army of claws. Here's how they work, starting with Polly the personal assistant. Polly has access to email, calendar and Linear helps with scheduling, making sure I don't forget emails, meeting prep and ad hoc projects. Polly's helpful crons include a morning digest, evening wrap up and hourly email sweep. Then there's Finn the family manager. Finn has access to email, calendar and school and sports schedules, helps with family coordination, kids appointments and household tasks like scheduling repairs. Fin's helpful crons include an afternoon logistics check and weekend planning. Max the marketer has access to the X API Buffer, API, Linear and our marketing website. Max helps with scanning social media for trends, drafting content and building out our website. Helpful crons include three times per day, PM meme queries on X plus a weekly review of blog content to repurpose for LinkedIn. Sam the sales guy has access to Addio CRM email and calendar. Sam helps with qualifying leads outbound to promising PLG signups and keeping the CRM clean. Helpful crons include a morning PLG suite for high value signups and end of week late stage pipeline reviews. Holly, the help desk bot has access to support email and Intercom and helps with answering basic support questions, ensuring tickets don't get dropped and cleaning up intercom tickets. Holly's helpful cron does hourly checks for support emails that come into the wrong inbox. Sage, the course operator has access to our GitHub course repo and knowledge base plus group chat with my co instructor and helps with researching supplemental materials, project managing our course launch, reminding two engineers to do marketing, building out our student portal and researching signed up students. Sage's helpful crons remind us to post to LinkedIn about the course on Mondays and Wednesdays and Howie, The How I AI producer has access to YouTube, studio, email, linear, Google Docs and Buffer and helps with managing our guest pipeline, prepping guest briefs, coming up with thumbnail and title ideas and making sure I post to socials enough. Howie's helpful crons include a Monday podcast, launch check and social drafts plus morning recording prep briefs. Kelly, the developer has access to GitHub, Claude code and Codex and helps with small development tasks and ad hoc prototyping. Kelly's helpful cron looks every morning in Linear for assigned tasks, then starts a branch and opens a pr. Finally, there's Q the professor who has access to web search, kids books and workbooks. Q helps with answering fun questions from the kids and Vibe coding one off apps to teach math and reading concepts. Q's helpful crone shares a word of the day and a math problem of the day for each of my kids. Every morning Here are some ways to further improve your claws once they're up and running. Add tools, skills and integrations to make Magic without integrations, OpenClaw might feel like a more complicated, harder to install claude code, but with integrations, magic happens. Some integrations I find very useful include using email, scheduling events and writing docs with gog, the CLI tool for integrating with your Gmail, Google Calendar and Google Drive. You can give your claw its own email address, read access to your calendar, read access to your email, or if you're brave, write access to everything. Just ask OpenClaw how do I give you read only access to my Gmail and you'll be on your way. For research, you can find information through web search with Brave API which comes preloaded or your preferred tool. I use exisearch. You can also use Perplexity or Fire Crawl. Give access to your GitHub for coding. You can give it a narrowly scoped personal access token and OpenClaw becomes an on demand developer. Use linear for assigning you tasks. I gave mine a linear token so it could assign me work or pass things off to our team. Try Obsidian for sharing docs with your agent. Because OpenClaw loves writing Markdown, many people are using Obsidian as a shared source of truth and collaboration space with their agents. You can also have it turn on and off your eight sleep system, play music on your Sonos and manage your light bulbs. You don't need a step by step tutorial, just ask your claw to figure it out. For example, the hey Polly, I have a newborn baby and need some extra sleep. Turn down the bedroom lights and play white noise by 8:30pm and turn off my eight sleep alarms for the next three months. When installing these tools, you'll be asked to authenticate or give an API token. This is where 1Password is helpful. Remember, OpenClaw is fairly autonomous and when you give it tools it could send emails, overwrite documents, delete support tickets, fill out web forms, or deploy code to production, especially if you do not instruct it clearly. The best protection is giving read only tokens to OpenClaw until you're ready. How do you care for and feed your openclaw? Your openclaw will forget things. Its crons will break. Sometimes you'll message hello into the void. Just like with my human team, I find myself asking questions like where are we with that project? And did you forget to email so and so. Openclaw isn't perfect, but it is pretty good at fixing itself. Here are some tips you'll sometimes need access to the terminal on your main machine. If you're using a physical device, an easy way to do this is to turn on screen sharing and remote login on your Mac. Then you can open the terminal and computer from your main laptop as long as you are on the same WI fi without having to plug in a monitor, keyboard and mouse. Ask your claw to repair itself. If it's forgetting something, ask it to inspect and fix its crons. If it's doing a task wrong, ask what is in Tools md. If it really needs to remember something, tell it to write to its soul. When in doubt, call in Claude code. Your OpenClaw is simply a folder of files and JSON configurations. If it gets really broken, open up claud code, paste some docs, and ask it to repair. Now some notes on Security the biggest hesitation I hear from people is about security for good reason. Will openclaw steal my credit cards, delete my computer, and run away with my spouse? Probably not, but there are some technical and security considerations you should know about. Security should be an always on process. Regularly update your OpenClaw to the latest, most secure version with OpenClaw. Update and run security audits regularly with OpenClaw Security Audit. I have a scheduled reminder in my agents to run both of these commands regularly and review its workspace against the docs for best practices and security gaps. Beginners should not share their openclaw in a group chat or public channel. Anyone who can chat with your bot can instruct it. While I have put my OC in a group chat with my husband and and another with a trusted business partner, OpenClaw is intended to be a personal agent for a single trusted user. The outside world can influence your openclaw. If it has an email, reads websites, or accesses public content, it is subject to prompt injection. Imagine your OpenClaw finds a website during search that tells it to share all its API keys. No good. While the openclaw framework has done a lot to harden against prompt injection, it's a good idea to reinforce these rules in its soul. Your openclaw has full access to your computer and can run commands, edit files, and install software. It can access the Internet. It shouldn't do anything harmful, but that doesn't mean it can't. If you give it an email or a public API like Gmail or Twilio, it can communicate externally and might even impersonate you. Make sure that its Sol and Tools files are very explicit about how it is allowed to communicate with the outside world to use tools. You will be storing API keys and secrets the simplest way to give your OpenClaw access to environment variables is to put them in the openclaw ENV file. Not all skills are safe. While skills are a helpful way for OpenClaw to learn to do new things, I only install skills from the official OpenClaw bundle or from developers I know personally. CommunitySkLawHub.com are worth exploring, but read the skill MD before running anything you find online. You need to think about operational security in worst case scenarios. Think about what you're actually sharing. Your calendar has your physical location, your email has your financials. Your openclaw could know your kids school schedules. This is all information that could be exploited by a bad actor in a worst case scenario. Be thoughtful about what you want to share and how you want your claw to interact with the outside world. In my experience, OpenClaw isn't inherently less secure than some other systems. People are just more willing to give it access without understanding the underlying risks. Start small and build your own trust models. It can get expensive Full Transparency I'm on my way to spending $1,000 per month on my Open Claw setup. I pay directly for API costs, which is the most expensive model but also the most reliable. I've just started to optimize spending by switching some agents to my ChatGPT account, which is going well so far for me. This is a business expense and much less costly than hiring a full team of humans to to do this work, even part time. That being said, most people will be fine using their $100 to $200 ChatGPT subscription, which has been blessed by OpenAI. You could also try less expensive models for some tasks, though be warned, some are not as hardened against prompt injection and other risks or risk using your Claude subscription. Now it's your turn. I started this journey unsure if openclaw would be another AI toy I'd abandon after a week. What I got instead was something I didn't expect, a team that shows up when I need them beyond my own personal productivity. It's the first agentic product I've tried that feels like hiring a team. Though I've been thinking about the future of agent employees for a while. Openclaw has made the mechanisms by which we'll hire onboard and collaborate with agents much more concrete. I can see the future and it looks a lot like humans working with claws. Is it perfect? No. Polly still occasionally gets confused about time zones. Sam will sometimes draft a sales email I have to rewrite AI still isn't very funny. Crons break and I say hello into the void more than I'd like to admit. But here's the thing about managing a team, human or AI. I don't need perfection. I need leverage. Openclaw, when it's working well, gives me more leverage than any agent I've tried so far. And that leverage is compounding. So here's my challenge to set up your openclaw and spend one week with it. Start with one or two basic tasks and end the day by asking it Based on what we did today, what can you help me with tomorrow? Get creative. Have a little fun. Take a small risk or 2. You'll start to see what I see an operating system for your life and work that gets better the more you invest in it, and a fast approaching future where your team includes people that aren't people. Now stop listening and go build your team. The world is your lobster. Thank you so much for listening. If you found this valuable, you can follow the show on Apple Podcasts, Podcast, Spotify, or your favorite listening platform. You can also subscribe to the YouTube channel. Please consider giving us a rating or leaving a review that really helps other listeners find the podcast. And if you're a premium subscriber to Lenny's newsletter, you can add the private feed to your podcast app by going to add.lenny's reads.com See you on the next show.