turn into a widely known open source system for running Claude Code. OpenClaw is actually a system that runs Claude Code indefinitely, allowing you to set it up as your personal AI assistant.
You may arrange OpenClaw to perform every kind of tasks for you, similar to:
- Reviewing GitHub pull requests
- Analyzing emails
- Browsing the web.
I’ve spent the last week establishing many various OpenClaw systems to act as my bots, specialized in several application areas. In this text, I’ll take you thru how I’ve arrange OpenClaw by myself system, ensuring the implementation is each secure and makes me simpler as an engineer.
I’ll cover the specifics of where it is best to arrange OpenClaw, how you may set it up, the really useful approach to run your OpenClaw instances, and how you can get essentially the most out of OpenClaw.
Why arrange an OpenClaw Assistant?
The primary reason it is best to arrange OpenCline in your computer is solely that it makes you simpler. You may arrange OpenClaw to automate every kind of various tasks and enable you organize your entire work.
For instance, as a substitute of manually scouring through your emails on the lookout for relevant emails and coming up with answers, you may simply tell OpenClo to research your entire emails every day, teach it which emails you think that are relevant and which aren’t, and have it provide you with example responses you may allow it to send.
You too can give OpenClaw access to your GitHub profile, have it notify you every time you’re tagged in a relevant pull request. Analyze the pull request as if it were you and suggest a pull review comment.
One of the best a part of OpenClaw is that it will possibly be personalized:
You may tell it exactly how you can behave and you may teach it over time so it will possibly turn into higher and higher. This is strictly what you wish in a private assistant. It’s very capable right off the bat and becomes much more capable the more you teach it.
How one can implement OpenClaw
On this section, I’ll undergo how I implemented OpenClaw. There are plenty of alternative ways to do it, and plenty of different applications you may arrange. To maintain it easy and specific, I’ll inform you how I did it specifically and which apps I connected it to, to present you inspiration for how you can do it yourself. Nonetheless, there is no such thing as a one true approach to do that, and the optimal way for you relies on your workflow.
Access to Claude Code
The very first thing you wish is access to Claude Code. Claude Code offers three different subscription tiers, which provide you with a set amount of usage per thirty days. That is correspondingly $20, $100, and $200 per thirty days.
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 inside my usage limitations per thirty days.
You too can do Pay Per Use through an API pricing, though I don’t recommend this because it’ll quickly get dearer than the equivalent usage you can get through one in all the subscriptions I described within the section above.
Once you may have access to Claude Code, you may run the command below to establish a token, which you’ll be able to provide to open Claude when setting it up. Remember to not share this token with anyone, because it gives access to your Claude Code subscription.
claude setup-tokenDocker images on a separate computer
You may install OpenClaw through this link. Exactly how you put in it’ll rely upon your operating system so I’ll avoid providing specific commands here. Nonetheless, one other way of setting it up, which may be the easiest method, is to easily tell Claude Code to set it up for you. Provide Claude Code the token we described above, and it will possibly arrange your assistant for you. I actually have done this three different times, and it’s worked each time to completely arrange my assistant exactly how I need it.
Whenever you arrange your assistant, I like to recommend telling Claude Code to set it up as different Docker images in your computer. This has multiple benefits.
- The agent will run in an isolated environment and never get access to things it shouldn’t have access to. That is super necessary for security concerns.
- Running it as a Docker image makes it easy to maneuver and create backups of your agent. You may simply download the Docker image and apply it to a special computer, and store the image on Docker Hub to have a backup of your agent.
You may simply tell Claude Code to set all the pieces up in Docker, and it’ll do it for you. You don’t need to do anything yourself there.
Personalizing OpenClaw
After you’ve arrange OpenClaw with Claude Code, it is best to personalize it. Tell Claude Code to open the OpenClaw dashboard in your required web browser and begin chatting along with your agent. It is going to ask… The agent will ask to your name and what the agent ought to be called, and you may give it a personality.
I’ve arrange multiple different bots. For instance, I arrange a private assistant that has my personality and tries to be super rational and just gives me concise summaries of all the pieces I want to do and things I want to pay attention to.
I’ve also arrange a sales bot which has a little bit of a special attitude, very positively minded, and which I’ve given access to relevant sales material and so forth.
Normally, you may simply chat along with your agent and tell it to recollect things. Open Claw will then proceed to store necessary information in memory and remember it for later.
Access
After you’ve arrange OpenClaw with its personality, it is best to start by giving it access to stuff. When giving access to your agents, it is best to follow the principle of giving the smallest amount of access vital to perform actions. For my personal bot, I actually have given the next access.
- Slack (where we communicate)
- My emails and calendar so it will possibly read emails and book meetings
- Linear, so I can check the several tasks I actually have to do.
- GitHub, so it will possibly perform actions on my behalf on GitHub.
I’ve also arrange a special agent that acts as a sales bot. This bot has been given access to the CRM system ,where it will possibly 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 will possibly do. If you ought to perform an motion, you must give it access and permissions to accomplish that. Nonetheless, it is best to also watch out with the access you give it, as your agent will act fully autonomously inside these systems.
Skills
One other incredibly necessary a part of establishing OpenClaw is the talents you provided. Should you want OpenClaw to recollect things for later or act in a particular way, you must provide it with skills. To offer the OpenClaw skill, you may simply tell it “store this as a skill” after you provide some information.
I’ll give a number of examples of the talents I’ve created:
- GitHub skill: this tells the agent the way it should act on GitHub on behalf of me. It, for instance, tells him how I do pull request reviews (I made my agent take a look at all my different reviews from previous to research my preferences.)
- Gmail skill: tells the agent which emails it should set to red mechanically, which I don’t care about, and which emails it should inform me about in my every day briefing.
- Slack skill: tells the agent how you can interact on Slack, for instance to at all times respond in threads and never as latest messages.
- Calendar skill: tells the agent exactly how you can read my calendar, informs me of meetings, and how you can book meetings with others, and how you can interact with the Google API for Google Calendar.
Normally, I try to offer the agent with a relevant skill each time I need it to perform an motion.
The skill will then be loaded dynamically every time the agent is asked to do something related to a given skill. For instance, if asked to read Gmails or emails, it’ll read the Gmail skill.
What doesn’t work with OpenClaw
I’ve experimented rather a lot with OpenClaw within the last week. I’ve noticed situations where it really works incredibly well out of the box, and I also noticed scenarios where it doesn’t work as well. There are two most important things try to be aware of that don’t work thoroughly.
- Being vague
- Simply telling the agent to recollect stuff for later.
Being vague doesn’t work because OpenClaw doesn’t plan in the identical way as you do and doesn’t have access to all of the contacts it needs. It’s best to thus make sure that to have a really specific plan and avoid ambiguity every time establishing things for OpenClaw to do.
To attain this, I like to recommend discussing with an LLM beforehand before attempting to implement something, after which being super specific once you are attempting to implement it. It’s not an issue to alter your plan in a while as OpenClaw will adapt, for those who provide a really vague plan, OpenClaw will struggle rather a lot and can likely not have the ability to implement the thing you plan to make.
Moreover, simply telling the agent to recollect stuff for later doesn’t work as well. Normally, it is best to make sure that to store all necessary information in skills. At any time when you teach the agent something specific, tell it to either add it to a previously created relevant skill or make a brand new skill with the data. These skills will then be loaded every time the agent is performing actions.
For instance, for those who provide the agent with an email reading skill, this skill might be loaded every time the agent interacts with emails. So for those who want the agent to perform in a particular way when reading emails or sending emails, it is best to store it in a separate skill.
I thus highly recommend ensuring the agent stores all relevant information in explicit skills and that you simply keep track of those skills and continually update them as you get increasingly information.
Conclusion
In this text, I’ve passed through how you can arrange OpenClaw. You may simply arrange OpenClaw totally free,e on condition that you have already got a Claude Code subscription. When establishing OpenClaw, it is best to set it up on a separate Docker container,s isolating each environment and ensuring different agents don’t have access to one another’s information and keys. Creating an OpenClaw assistance has been incredibly powerful for me, and in lower than per week after setting them up, I’ve already noticed massive efficiency gains. Nonetheless, I’ll also notice scenarios where the agent doesn’t work as well, which you may have to take into consideration when establishing your assistance. The general key, nonetheless, is to be as specific as possible and make the agent store all the pieces relevant as skills that will 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
