become a well-known open source system for running Claude Code. OpenClaw is essentially a system that runs Claude Code indefinitely, allowing you to set it up as your personal AI assistant.
You can set up OpenClaw to perform all kinds of tasks for you, such as:
- Reviewing GitHub pull requests
- Analyzing emails
- Browsing the internet.
I’ve spent the last week setting up many different OpenClaw systems to act as my bots, specialized in different application areas. In this article, I’ll take you through how I’ve set up OpenClaw on my own system, ensuring the implementation is both secure and makes me more effective as an engineer.
I’ll cover the specifics of where you should set up OpenClaw, how you can set it up, the recommended way to run your OpenClaw instances, and how to get the most out of OpenClaw.
Why set up an OpenClaw Assistant?
The number one reason you should set up OpenCline on your computer is simply that it makes you more effective. You can set up OpenClaw to automate all kinds of different tasks and help you organize all of your work.
For example, instead of manually scouring through your emails looking for relevant emails and coming up with answers, you can simply tell OpenClo to analyze all of your emails daily, teach it which emails you think are relevant and which are not, and have it come up with example responses you can allow it to send.
You can also give OpenClaw access to your GitHub profile, have it notify you whenever you’re tagged in a relevant pull request. Analyze the pull request as if it were you and suggest a pull review comment.
The best part of OpenClaw is that it can be personalized:
You can tell it exactly how to behave and you can teach it over time so it can become better and better. This is exactly what you need in a personal assistant. It’s very capable right off the bat and becomes even more capable the more you teach it.
How to implement OpenClaw
In this section, I’ll go through how I implemented OpenClaw. There are a lot of different ways to do it, and a lot of different applications you can set up. To keep it simple and specific, I’ll tell you how I did it specifically and which apps I connected it to, to give you inspiration for how to do it yourself. However, there is no one true way to do this, and the optimal way for you depends on your workflow.
Access to Claude Code
The first thing you need is access to Claude Code. Claude Code offers three different subscription tiers, which give you a set amount of usage per month. This is correspondingly $20, $100, and $200 per month.
I utilize Claude Code for my other programming as well and have the max subscription. I can then use this same subscription for my OpenClaude assistance and still be well within my usage limitations per month.
You can also do Pay Per Use through an API pricing, though I do not recommend this because it will quickly get more expensive than the equivalent usage you could get through one of the subscriptions I described in the section above.
Once you have access to Claude Code, you can run the command below to set up a token, which you can provide to open Claude when setting it up. Remember not to share this token with anyone, as it gives access to your Claude Code subscription.
claude setup-tokenDocker images on a separate computer
You can install OpenClaw through this link. Exactly how you install it will depend on your operating system so I’ll avoid providing specific commands here. However, another way of setting it up, which might be the simplest way, is to simply tell Claude Code to set it up for you. Provide Claude Code the token we described above, and it can set up your assistant for you. I have done this three different times, and it’s worked every time to fully set up my assistant exactly how I want it.
When you set up your assistant, I recommend telling Claude Code to set it up as different Docker images on your computer. This has multiple advantages.
- The agent will run in an isolated environment and not get access to things it shouldn’t have access to. This is super important for security concerns.
- Running it as a Docker image makes it easy to move and create backups of your agent. You can simply download the Docker image and use it on a different computer, and store the image on Docker Hub to have a backup of your agent.
You can simply tell Claude Code to set everything up in Docker, and it will do it for you. You don’t have to do anything yourself there.
Personalizing OpenClaw
After you’ve set up OpenClaw with Claude Code, you should personalize it. Tell Claude Code to open the OpenClaw dashboard in your desired web browser and start chatting with your agent. It will ask… The agent will ask for your name and what the agent should be called, and you can give it a personality.
I’ve set up multiple different bots. For example, I set up a personal assistant that has my personality and tries to be super rational and just gives me concise summaries of everything I need to do and things I need to be aware of.
I’ve also set up a sales bot which has a bit of a different attitude, very positively minded, and which I’ve given access to relevant sales material and so on.
In general, you can simply chat with your agent and tell it to remember things. Open Claw will then proceed to store important information in memory and remember it for later.
Access
After you’ve set up OpenClaw with its personality, you should start by giving it access to stuff. When giving access to your agents, you should follow the principle of giving the least amount of access necessary to perform actions. For my personal bot, I have given the following access.
- Slack (where we communicate)
- My emails and calendar so it can read emails and book meetings
- Linear, so I can check the different tasks I have to do.
- GitHub, so it can perform actions on my behalf on GitHub.
I’ve also set up a different agent that acts as a sales bot. This bot has been given access to the CRM system ,where it can get all its relevant material regarding sales. I’ve also given him Slack access where we communicate.
Overall, the access you give your agent is crucial for what it can do. If you want to perform an action, you need to give it access and permissions to do so. However, you should also be careful with the access you give it, as your agent will act fully autonomously within these systems.
Skills
Another incredibly important part of setting up OpenClaw is the skills you provided. If you want OpenClaw to remember things for later or act in a specific way, you need to provide it with skills. To provide the OpenClaw skill, you can simply tell it “store this as a skill” after you provide some information.
I’ll give a few examples of the skills I’ve created:
- GitHub skill: this tells the agent how it should act on GitHub on behalf of me. It, for example, tells him how I do pull request reviews (I made my agent look at all my different reviews from previous to analyze my preferences.)
- Gmail skill: tells the agent which emails it should set to red automatically, which I don’t care about, and which emails it should inform me about in my daily briefing.
- Slack skill: tells the agent how to interact on Slack, for example to always respond in threads and not as new messages.
- Calendar skill: tells the agent exactly how to read my calendar, informs me of meetings, and how to book meetings with others, and how to interact with the Google API for Google Calendar.
In general, I try to provide the agent with a relevant skill every time I want it to perform an action.
The skill will then be loaded dynamically whenever the agent is asked to do something related to a given skill. For example, if asked to read Gmails or emails, it will read the Gmail skill.
What doesn’t work with OpenClaw
I’ve experimented a lot with OpenClaw in the last week. I’ve noticed situations where it works incredibly well out of the box, and I also noticed scenarios where it doesn’t work as well. There are two main things you should be aware of that don’t work very well.
- Being vague
- Simply telling the agent to remember stuff for later.
Being vague doesn’t work because OpenClaw doesn’t plan in the same way as you do and doesn’t have access to all the contacts it needs. You should thus make sure to have a very specific plan and avoid ambiguity whenever setting up things for OpenClaw to do.
To achieve this, I recommend discussing with an LLM beforehand before trying to implement something, and then being super specific once you try to implement it. It’s not a problem to change your plan later on as OpenClaw will adapt, if you provide a very vague plan, OpenClaw will struggle a lot and will likely not be able to implement the thing you intend to make.
Furthermore, simply telling the agent to remember stuff for later doesn’t work as well. In general, you should make sure to store all important information in skills. Whenever you teach the agent something specific, tell it to either add it to a previously created relevant skill or make a new skill with the information. These skills will then be loaded whenever the agent is performing actions.
For example, if you provide the agent with an email reading skill, this skill will be loaded whenever the agent interacts with emails. So if you want the agent to perform in a specific way when reading emails or sending emails, you should store it in a separate skill.
I thus highly recommend making sure the agent stores all relevant information in explicit skills and that you keep track of these skills and continually update them as you get more and more information.
Conclusion
In this article, I’ve gone through how to set up OpenClaw. You can simply set up OpenClaw for free,e given that you already have a Claude Code subscription. When setting up OpenClaw, you should set it up on a separate Docker container,s isolating each environment and making sure different agents don’t have access to each other’s information and keys. Creating an OpenClaw assistance has been incredibly powerful for me, and in less than a week after setting them up, I’ve already noticed massive efficiency gains. However, I’ll also notice scenarios where the agent doesn’t work as well, which you have to take into account when setting up your assistance. The overall key, however, is to be as specific as possible and make the agent store everything relevant as skills that can be loaded dynamically on demand.
👉 My free eBook and Webinar:
🚀 10x Your Engineering with LLMs (Free 3-Day Email Course)
📚 Get my free Vision Language Models ebook
💻 My webinar on Vision Language Models
👉 Find me on socials:
💌 Substack