OpenClaw Backup and Rebuild: Restoring 12 AI Agents From Scratch
Prefer to watch? Watch on YouTube · Subscribe for new builds · Read the transcript
In this video I walk through how I rebuilt my whole OpenClaw setup, all 12 agents, from scratch after a power outage forced a fresh install on my Mac Mini. It's for anyone running OpenClaw agents who wants to know what to back up, what order to rebuild in, and how to keep API keys safe so a bad night doesn't wipe out weeks of work.
I recorded this in March 2026. Tools change fast, so check the current OpenClaw docs before copying any of my setup exactly.
What I cover in this video
- 0:00 What happened: about 4 weeks of a working setup, then the power went out and I needed a fresh install
- 2:17 The three ways I back up OpenClaw: local markdown memory, Neon, and Supermemory
- 3:02 Why you should never put API keys inside cron job prompts
- 4:35 Checking the Neon database and using a Telegram chat export to brief Atlas
- 5:22 The order I brought agents back online, starting with Ghostface
- 6:07 Training Rizza on 600 of my video transcripts so scripts sound like me
- 8:23 The 5 a.m. to 9 a.m. cron schedule and a weekly thumbnail feedback loop
- 10:40 Extra security hardening and the five lessons I took away
What I had to rebuild and what saved me
My Mac Mini had been running for close to 40 days straight when the power went out. The setup I lost included 12 cron jobs, multiple API keys, database connections, my voice profile, and security configurations. Doing the fresh OpenClaw install was the easy part. Reconnecting everything (the Telegram bot, the database, Supermemory, and all the API keys) was the real work.
I didn't start from zero because I back up in three places:
- Local persistent memory in a markdown file
- Neon as the database
- Supermemory for graph search
Neon still had everything: 27 tables, 25 videos, member data, and content drafts. Early on I made a rule that the database is the source of truth, not the local files, and that rule saved me here. I verified the connection, ran a few test queries, and moved on.
The other thing that helped was my Telegram chat history. I exported it to a JSON file and uploaded it to Claude, and it put together a full outline of what to give my main agent, Atlas, during the rebuild. Between all these backups, I'd say I saved 80 to 90% of the time it would have taken otherwise.
Keeping API keys out of your cron prompts
A lot of people dump their API keys right into cron job prompts without realizing it. That's a security nightmare. This time I created an environment file and told OpenClaw (running on Opus 4.6) that I didn't want any security vulnerabilities or prompt injection risk. It still missed the step and started putting keys into the cron prompts anyway.
So I went back and redid the security permissions so only my user can read that file. Now every cron job sources the environment file instead of having keys hardcoded in the prompt. My advice is simple: keep all your API keys in a local environment file with locked permissions.
I also added more lockdown during the rebuild. I put API authentication on my Mission Control dashboard, so if someone finds the localhost address they still can't see the data without the key. I made sure there were no plain text keys anywhere in the config files. When your agents can touch your database, social accounts, and business data, security is day one stuff.
Rebuilding the agent pipeline in order
The order matters, because later agents depend on what earlier ones produce. All my agents are named after Wu-Tang characters.
- Ghostface runs at 5 a.m. and gathers intel from Reddit, X, LinkedIn, YouTube, Hacker News, and Indie Hackers, looking for trending AI topics like OpenClaw use cases and people making money with agents. That intel goes into a file every other agent reads.
- Rizza runs at 6 a.m., reads Ghostface's intel, checks the database so we don't repeat videos, and writes four video scripts a day. This time I gave it 600 transcripts from my reselling channel, transcribed locally on the Mac Mini with Whisper in a few hours, so it picks up my phrases and cadence.
- Inspector Deck handles X posts, You God makes thumbnails (three variations per topic), and the Jizza writes newsletters on Mondays and Thursdays.
- ODB scouts X and LinkedIn for people building in public with OpenClaw and Claude Code who might come on my podcast.
- Method Man checks my channel every 8 hours for new long-form uploads, transcribes them with Whisper, picks the best sections, and turns them into short clips for social media.
Everything kicks off between 5 a.m. and 9 a.m. Right before filming, I also set up a feedback loop: every Monday a cron job pulls data from the YouTube API, ranks the thumbnail outliers by topic, and learns what worked so the next thumbnails and copy get better.
The whole rebuild took about 2 hours, from midnight to 2 a.m. The first time I built this it took 35 days. The difference was having memory files, daily logs, and documentation to work from. Running it costs the $500 Mac Mini, $200 a month, and about $13 a month in API calls through OpenRouter.
If you want more on building setups like this, see my AI Agents page.
Key takeaways
- Your database is your safety net. If everything had lived in local files, I'd have been rebuilding from memory. The important stuff lives in Neon and Supermemory.
- Document your setup. Memory files, daily logs, and configuration notes are what turned 35 days into 2 hours.
- Build security in from the start. Use environment files with locked permissions, API authentication, and no hardcoded keys, and double check that your agent actually followed your instructions.
- The second build is faster, and often better. Don't be afraid to start over. What I have now is better than what I had before.
- One person can run this. I'm a reseller who learned to vibe code, not a developer, and OpenClaw lets me run 12 agents on my own.
FAQ
How do you back up an OpenClaw setup?
I back up in three places: a local persistent memory markdown file, a Neon database, and Supermemory for graph search. I also exported my Telegram chat history to JSON, which helped Claude put together a rebuild outline for my main agent.
How long does it take to rebuild OpenClaw agents from scratch?
With my backups and documentation, rebuilding all 12 agents took about 2 hours. The first time I built the setup without that documentation, it took 35 days.
Where should I store API keys for OpenClaw cron jobs?
Put them in a local environment file with permissions set so only your user can read it, and have each cron job source that file. Don't hardcode keys into cron prompts, and check your agent's work, because mine put keys into the prompts even after I told it not to.
How much does it cost to run a 12-agent OpenClaw setup?
For me it's a $500 Mac Mini, $200 a month, and about $13 a month in API calls through OpenRouter.
Transcript
Read the full transcript
0:00In this video, I'm going to tell you about how I rebuilt my entire OpenClaw setup, all 12 agents completely from scratch within hours. And by this morning, everything was all up and running again, but it was a stressful night. So, here's what actually happened. I had my OpenClaw running on the Mac Mini for about 4 weeks behind me perfectly. Then, the power went out. Everything was dialed in. Atlas was handling all the mission control. Rizza was writing the scripts. Inspector Deck was searching for X posts. And Ghostface was gathering intel every morning at 5:00 a.m. Everything was beautiful, working perfectly. And then, I had to do a completely fresh install. Like I said, the power went out, and my Mac Mini has been running for about almost close to 40 days straight. So, I'm not going to
0:45lie, you know, when I realized I needed to build it out everything from scratch, my stomach dropped, my heart sank, I was feeling the pressure. And what we've built to to this point was, you know, very detailed, you know, and and I don't want to get into it too much, but you know, it was just really hard to figure out what I was going to do and go from there. This was 12 cron jobs, multiple API keys, database connections, my voice profile dialed in, security configurations, all that. And what I learned, and why I'm making this video to share with you if case this ever happens to you, and you know, if you're building with OpenClaw, you need to know what's important in case you lose everything. You know, you need to understand that rebuilds happen, especially in the AI space. Things
1:31break. You might need to take your whole app and start over. And don't feel bad or discouraged. I've done that several times through building Easy Flip, Magic Hand, and Snaptastic. Sometimes you just got to start over. So, I'm going to walk you through exactly what I did hour by hour. The first thing I did on a fresh install was just OpenClaw fresh install, get it done. That's the easy part. But then, you need to connect everything. So, the Telegram bot, the database, Super memory, all the API keys, and I will tell you that if you're not backing up your OpenClaw to another external source, even like GitHub or some type of cloud database, I have three ways we back this up. So, we have the local persistent memory markdown file, and
2:17then we also have Neon for database, and then we also have Super memory for graph search. And if I hadn't have done that, I would have started over from zero. So, all I had to really rebuild was a lot of the cron jobs, API keys, and all of the prompts, which actually did take a while to actually get it dialed in. So, real quick, before we keep going, if you're watching this and you want to actually build with some of these tools, not just watch videos about them, then you're going to want to check out our community down below, Shipping School. We have a full Claude code course, a full OpenClaw course, and four live bootcamps every single week, where we actually help you get set up from scratch. >> [music] >> Like actually set this thing up, not just watch a tutorial and figure it out
3:02by yourself. And we also provide one-on-one coaching, so you could book a call with me, we could share screens, and I can help you get Claude code or OpenClaw running on your machine. That's it, no fluff. I built this community because watching YouTube's only gets you so far. We launched it just 3 days ago, and we have over 55 members. You need people around you who are actually building, people who hold you accountable, and coaches who can help you when you get stuck. I'll put the link in the description down below. Get in now before the price goes up. Here's the thing that most people don't realize about this, that you can't just dump your keys into cron job prompts. I think a lot of people are doing this and not knowing it. That's a security nightmare. So, like this time, I created an environment file, redid everything from
3:49scratch, and I told OpenClaw, which runs on Opus 4.6, we don't want any security vulnerabilities. We don't want prompt injection. And for some reason, it just missed this step, and started doing what I didn't want it to do, which was put the keys in the cron prompts. So, we basically went back, redid a lot of the security permissions, so only my user can read it. And you know, every single cron job sources that file instead of having to use the keys hardcoded in the prompts. Everyone should have an environment.local folder with all of your API keys. It's that plain and simple. Next up was the database. I'm running on Neon, and you know, praise Neon, because >> [laughter] >> all my data was still there. And if it wasn't, you know, everything would be
4:35lost. Everything. All 35 days of progress gone. So, we have 27 tables, 25 videos, all the member data, content drafts, everything intact. The database is the source of truth, not the local files, and that's the rule I set early on, and it saved me here. So, I verified the connection and ran a few test queries and moved on. That was amazing. And here's where it gets more interesting is the agent pipeline. So, I'm going to break it down what I exactly rebuilt in what order, because the order is what matters. I do want to talk about here is I have all of the chat data from Telegram, which did help save me. I was able to export that into a JSON file, and then re-upload that to Claude, and it basically rebuilt an entire outline on what to give to my
5:22main agent, Atlas, during this rebuild. So, it got a lot of context, and all of this backup, even backup to a backup, has saved 80 to 90% of the time to get this thing back up and running. So, the first agent that was back up and online and running was Ghostface. He runs at 5:00 a.m. every morning. His job is intel gathering, scraping things like Reddit, X, LinkedIn, YouTube, Hacker News, and Indie Hacker News, looking for basically trending AI topics. So, specifically OpenClaw use cases and people actually making money with agents. That intel gets saved to a file that every other agent reads. The second one was the Rizza, and yes, all of the agents are named after Wu-Tang characters. I just love the Wu-Tang, and
6:07it sounded natural. So, Rizza handles a lot of the content pipeline, runs at 6:00 a.m., reads Ghostface intel, checks the database for what videos we've already done, so there's no duplicates, and then goes ahead and generates four video scripts per day. But this time, I added something a little bit of some flair. I gave it 600 video transcripts of a YouTube channel that I have on another for reselling niche, and I was able to kind of mimic and copy exactly the common phrases that I say. You know, the local speech-to-text model, we used Whisper for the transcription, and the actual model running on my Mac Mini got all of this transcribed within just a few hours, and then I built that into all of the voice prompts. So, the voice
6:52profile comes from those transcripts written on data over the past 5 years. So, now the Rizza doesn't just write below. It actually sounds like me over 600 videos, my phrases, my cadence, and the way I actually talk. Pretty freaking cool. The Then came Inspector Deck for X posts, You God for thumbnails, the Jizza for newsletters. All of them copy the same thing. ODB handles the podcast outreach. Maybe we'll move him to something different, but for now, Method Man handles the clip factory, chopping the long-form into short-form. You don't need Opus clips for that. And one by one, all of these agents' cron jobs were coming back into line. Here's something I want to be real truthful about, and that's all you're going to get from me on this channel is mistakes, the wins,
7:38the failures, the successes, so you can learn from this, too. Um, it only took me about 2 hours. Though those 2 hours from were from midnight to 2:00 a.m., not ideal, but I wanted to get the system up and running, so we could be back to pumping out videos and content. So, from start to finish, 2 hours. The first time I set this up, it took me 35 days. So, multiple sessions with Atlas, figuring things out, but the second time around, because I had all the memory files, the daily logs, and everything documented, it was much faster. That's the lesson right there is being able to document everything. So, I know it sounds boring, but when in something breaks at 11:00 p.m. on a Friday night, you're going to thank yourself. Now, let me talk about the cron schedule. Um, you know, everyone kicks off every
8:23day from 5:00 a.m. all the way to 9:00 a.m. All of these different agents kick off, and they all do different things. So, You God is now our new member generating the thumbnails, and it generates three varieties based on each topic. We also just set up, like 10 minutes before this video was filmed, a recursive feedback loop. We hook into the back-end YouTube API, and every week on Monday, we kick off a cron job to pick the outliers of these thumbnails and rank them against each other based on the topic, and learn and get better at creating better thumbnails and copy. So, what worked and what doesn't. So, being able to have cron jobs kick off and also get better with each turn is another great way to level up your agent
9:09skills. So, ODB, I'll break down how he scouts for the podcast opportunities. Basically, you know, he goes on X and LinkedIn, and we look for people that are building in public and actually making money with OpenClaw and using Claude code. And basically, I would love to have them on the Build in Public podcast to share their journey, because learning from other people's stories can help give you a boost of confidence and momentum along your journey. You know, I'm no one special. I didn't go to school for this. That's why I'm sharing my journey with you, because giving you confidence to be able to do the same thing. So, we have another agent, Method Man. Basically, another job that kicks off every 8 hours, searching this YouTube channel for any new long-form uploads, and then chops them up into
9:54clips, and then transcribes them using Whisper, picks the best sections of the video, and those publish to all the social media. So, Mondays and Thursdays, the Jizza is in charge of writing the newsletters. You can subscribe to my Substack for free down below. Um that is all written in based on the content we're producing every week. So, 12 agents running flawlessly within a few hours, all running on the Mac Mini that sits behind the desk. The cost is only $500 for the Mac Mini, $200 a month, and basically just the API calls that we make through Open Router, which is like $13 for the month. So, not much. Now, what I also did is something important during this whole rebuild is the security hardening. We were locked down aggressively before, but I added
10:40even some more lockdown. So, I added AI or API authentication to Mythic Mission Control. So, if someone finds localhost, they still can't access the data without the key that I have stored down and salted. So, I locked out all the credentials, and I made sure there was no plain text keys existing anywhere in the config files. So, based on all of this, on what I've learned, and because here's the thing, you know, when you're running agents that have access to your database, your social accounts, your business data, security isn't optional. It's day one stuff, and I don't want you making this mistake of skipping it because you think nobody will find your local setup. You know, and that's where it's scary when you're running a VPS setup, or you really don't know what you're doing. I think if you're setting
11:25up Open Claw specifically, you got to put security at the forefront. So, the main thing that I learned from all of this is number one, your database is your safety net. If I'd been storing everything in local files, this rebuild would have been a disaster. I would have had to rebuild everything from scratch just based on memory. Everything important lives in Neon and in Super Memory. The number two thing that I learned is document your setup religiously. So, memory files, your daily log files, the configuration notes. The future you will be grateful if you experience any kind of power outage, or your Mac Mini breaks, or whatever. Your hard disk gets corrupted. The third thing is security from the start is, like I said, the most important thing. Environment files with
12:11locked permissions, API authentication, no hard-coded keys everywhere. Number four is second in or the second build is always faster. You know, if I hadn't have backed up all this to these other sites, I'd have been in trouble. But, the thing is also too, don't be afraid to start over because now what I have built behind me here is even better than what I had before. So, sometimes a clean slate is exactly what you need. Number five is Open Claw makes this possible for one person. I'm not a developer, I'm not a guru, I don't claim to be. I'm a reseller who learned how to vibe code. Over the last five years, I've built built a seven-figure resale business. We sell brands like North Face, Patagonia, Lululemon on a live selling platform called Whatnot. That's what we've been doing, and I saw everyone building from
12:56the sidelines, and you I'm like, "You know what? I can do that, too." I love building, and over the last year, I've been building every day with Cloud Code, and it's been amazing. But now, with 12 agents at my disposal with skills that are curated to meet my needs and my mission, I could build much faster than I've ever have within this last year. So, if you want to learn to actually build something like this and take security precautions and learn how to create backups in case this ever happens to you, that's what we cover in the community down below. Real steps, real agents, and real people, not just theory. We have 57 members in just five days. Community is growing every single day. I'll put the link down below. We'll see you in the next one, and be blessed.
Ready to start building with AI?
Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.
Join Shipping SkoolKeep reading
Hermes Desktop Tutorial: Install and Run Your First AI Agent
Hermes Desktop from Nous Research puts Hermes Agent in a chat-style app that took me about 15 minutes to install. I connected my OpenAI Codex plan with GPT-5.5 and ran three agents at once: a code repo review, a YouTube channel analysis, and an X account analysis. Then I set up a weekly cron job.
OpenClaw Memory Fix: 4-Layer Setup With Supermemory and Neon
To stop OpenClaw from forgetting, I use four memory layers: bootstrap markdown files, daily memory logs, Supermemory for semantic memory, and Neon (Postgres) for structured data. Every agent searches both systems before a run and writes back to both after it. Lossless Claw and QMD add more on top.
How to Fix OpenClaw Memory Loss: My 3-Layer Memory Setup
To stop OpenClaw from forgetting you, I use three memory layers: a short, curated memory markdown file plus daily logs, an hourly cron job that backs memory up to a Neon Postgres database, and Supermemory as a searchable cloud backup. Updating to OpenClaw 3.8 with its LCM graph memory helps a lot too.
