I have been getting deep into deploying my Claws both with OpenClaw and NemoClaw (variant) over the past two weeks. I have been tracking the reddit feed pretty close over last month and finally found some time to build. It is a very fun extension on top of agents, or easy framework to get into and surprisingly yields not just fancy fluff but some really great tools.

So in summary as of writing this, I have an exec assistant setup (who manages a team of subagents) that manages everything from meeting invites, to drafting emails, to taking in jobs in the house to fix, to reviewing and critiquing my blog as an editor to most recently my personal trainer and coach.

This falls into another tool just like my customized podcast app, that I have quickly fallen in love with and became part of my day to day.

But I think also what I think about it is the power of it comes from providing an easy gateway to integrating a set of fundamental tools that have been out there and agents are just promoting to another level. I expect more frameworks to follow, and there is obviously concern around the security elements (hence Nemoclaw) and also how it overlaps with other things cowork/dispatch custom agents. But I think what it highlighted me was how stiching together a few core elements : a mutlimodal gateway, orchestration layer and memory system all sitting on top of LLMs gives you the ability to create so many powerful tools, and really so easily.

So I will go over the main things I feel are compelling and why I am currently loving openclaw for hacking away at.

Connectivity as a Core Competency

I think about the simplest of things, communicating over telegram/slack with my bot and giving them identities. This is the default, rather than meeting a tool within the bounds of where it it works it meets me where I work or communicate. This ability to dispatch and send things as I flow in my own work and easily move between all my devices. It feels so basic an addition, and was always possible via other systems, but it just feels so much more natural with how openclaw gets setup. I think separating the backend/agents and workers away from you and to communicate remotely with them and asynchronously is quite powerful. I can now be sitting on my United free wifi messaging service and going back and forth with my agent spinning up all sorts of coding projects. Without needing to pay the $8 for full on wifi.

Orchestration ease

Creating specialized sub agents and a master delegation agent is huge for me and maybe I should have put this first. Now for sure I could be better at prompting and be my own delegator but it is just such a great user experience to throw out a coding project and for a primary agent to spin up a sandboxed coding agent with its own set of tools to complete the task, but that same agent to spin up a simple admin agent to review my todos and add new ones from emails. That ability to delegate and kickoff n number of asynchronous processes is great. Further being able to have them sandboxed with limited set of tools and access or file systems gives you the ability to constrain outcomes much better. I had previously used and still use Claude Code and projects. But they are just not as clean. They need me to constantly switching repos, or projects here I want something done and the brain load of how to do it is taken. That is one of those great characteristics of a trusted employee give your biggest problems out and know that in the background they are being accomplished and updates will be surfaced when you want.

I setup my openclaw like an employee with its own emails, git repo etc. It can add invites onto my calendar on one account but cannot read emails that I don’t tag it on.

Now Cron my old friend is also back in the game, everything from daily checkins to hooking up a senior engineer agent that goes with Opus/Sonnet once per week to review logs coming in from Langfuse and improving itself. The closed loop improvemnet and decisions on where to improve based on its own prompts and when to create subagents is great. I have the agent manage its own list of improvements on notion which I go in and prioritize for the coding agent to execute on as part of its weekly sprint.

Finally, my last thing towards an open system. Now I still remain a big fan of Claude even if they cut the amount of work I could do on my subscription per week, but the reality is just now the model ecosystem is 1) moving so quickly 2) diverging enormously in cost. I ended up for my own billing ease using Openrouter and through that I access most models which means each of my subagents is optimized across Gemini, DeepSeek and Anthropic model, which has brought the daily running of this down to a few dollars. As I don’t hammer sonnet when I just need flash lite and experiemntatation and flipping on a new model, creating new agents is so quick and the sytem itself can self-imrpove. All this adds so much power to the single developer. It does create a highly personalized experience so my setup might not suit someone else, but for me at the moment and state of products in the space this is more than ok.

Finally Memories

Now this one deserves more time and realistically I am still experimenting with this more with other memory systems and Karpathy’s latest tweet on top of Obsidian knowledge basis is super interesting. Right now the basic OpenClaw memory systems in markdowns are really effective and all I have extneded on top of this for the moment is a Supabase postgres database that stores more metrics data, such as for my fitness agent which power weekly analytical jobs. But this is an area that has so much more power.

Final Reflections

Like all areas of society, science, and engineering, compounding is an awesome thing to behold and this is what I see that OpenClaw is starting to unlock. Linking together gateways, memory and orchestration/scheduling and creating these wrappers around tools that have been out there is unlocking huge potential.

I have been getting deep into deploying my Claws — both with OpenClaw and NemoClaw — over the past two weeks. I have been tracking the Reddit feed closely over the last month and finally found time to build. It is a very fun extension on top of agents, and an easy framework to get into that surprisingly yields not just fancy fluff but some genuinely great tools.

In summary, as of writing this, I have an exec assistant setup (which manages a team of subagents) that handles everything from meeting invites to drafting emails, taking in household jobs to fix, reviewing and critiquing my blog as an editor, and most recently acting as my personal trainer and coach.

This falls into another tool — just like my customised podcast app — that I have quickly fallen in love with and made part of my day-to-day life.

What I think about most is the power that comes from providing an easy gateway to integrating a set of fundamental tools that agents are now elevating to another level. I expect more frameworks to follow, and there are obvious concerns around security (hence NemoClaw) and also how it overlaps with other things like Cowork or Dispatch's custom agents. But what it has highlighted for me is how stitching together a few core elements — a multimodal gateway, an orchestration layer, and a memory system, all sitting on top of LLMs — gives you the ability to create so many powerful tools, and really quite easily.

So here are the main things I find compelling and why I am currently loving OpenClaw as a hacking environment.

Connectivity as a Core Competency

Consider something as simple as communicating with your bot over Telegram or Slack and giving it an identity. This is the default: rather than meeting a tool within the bounds of where it works, it meets me where I work or communicate. This ability to dispatch tasks as I flow through my own work — and to move easily between all my devices — feels like a basic addition, and was always possible via other systems, but it feels so much more natural with how OpenClaw gets set up.

Separating the backend agents and workers from your direct interface, and communicating with them remotely and asynchronously, is quite powerful. I can now be sitting on my United free WiFi messaging service going back and forth with my agent spinning up coding projects — without needing to pay the $8 for full WiFi.

Orchestration Ease

Creating specialised sub-agents and a master delegation agent is huge for me — perhaps I should have put this first. For sure I could get better at prompting and serve as my own delegator, but the user experience of throwing out a coding project and having a primary agent spin up a sandboxed coding agent with its own set of tools is great. That same orchestrator can simultaneously spin up a simple admin agent to review my todos and add new ones from emails. The ability to delegate and kick off n number of asynchronous processes is powerful. Further, being able to sandbox each with a limited set of tools and file system access gives you much better constrained outcomes.

I previously used — and still use — Claude Code and projects, but they are just not as clean. They require constant switching between repos or project contexts, whereas here I want something done and the cognitive load of how to do it is removed. That is one of the great characteristics of a trusted employee: hand over your biggest problems and know that in the background they are being resolved, with updates surfaced when you want them.

I set up my OpenClaw like an employee, with its own email address, its own git repo, and so on. It can add invites to my calendar on one account but cannot read emails I haven't tagged it on.

Cron, my old friend, is also back in play — everything from daily check-ins to a senior engineer agent that runs once per week with Opus/Sonnet to review logs coming in from Langfuse and improve itself. The closed-loop improvement and decisions on where to improve, based on its own prompts and when to spin up sub-agents, are great. I have the agent manage its own list of improvements in Notion, which I go in and prioritise for the coding agent to execute on as part of its weekly sprint.

Finally, on openness: I remain a big fan of Claude even if they have reduced the throughput available on my subscription. But the reality is that the model ecosystem is 1) moving very quickly and 2) diverging enormously in cost. I ended up using OpenRouter for billing simplicity, through which I access most models. This means each of my sub-agents is optimised across Gemini, DeepSeek, and Anthropic models, which has brought the daily running cost down to a few dollars — I'm not hammering Sonnet when I just need Flash Lite. Experimentation and flipping to a new model or creating new agents is so quick, and the system can self-improve. All of this adds enormous power for the solo developer. It does create a highly personalised setup, so my configuration might not suit everyone, but for me at the current state of the product landscape this is more than fine.

Finally, Memories

This one deserves more time, and realistically I am still experimenting further with other memory systems. Karpathy's recent post on LLM knowledge bases built on top of Obsidian is super interesting and very much on my roadmap. Right now the basic OpenClaw memory system in markdown is really effective, and the only thing I have extended on top of it for the moment is a Supabase Postgres database that stores metrics data — such as for my fitness agent, which powers weekly analytical jobs. But this is an area with so much more potential.

Final Reflections

Like all areas of society, science, and engineering, compounding is an awesome thing to behold — and this is what I see OpenClaw starting to unlock. Linking together gateways, memory, and orchestration/scheduling, and creating these wrappers around tools that have been out there for some time, is unlocking huge potential. As Jensen Huang put it at GTC 2026, OpenClaw is the operating system for personal AI — and I'm inclined to agree.

I have been getting deep into deploying my Claws — both with OpenClaw and NemoClaw — over the past two weeks. I have been tracking the Reddit feed closely over the last month and finally found time to build. It is a very fun extension on top of agents, and an easy framework to get into that surprisingly yields not just fancy fluff but some genuinely great tools.

In summary, as of writing this, I have an exec assistant setup (which manages a team of subagents) that handles everything from meeting invites to drafting emails, taking in household jobs to fix, reviewing and critiquing my blog as an editor, and most recently acting as my personal trainer and coach.

This falls into another tool — just like my customised podcast app — that I have quickly fallen in love with and made part of my day-to-day life.

What I think about most is the power that comes from providing an easy gateway to integrating a set of fundamental tools that agents are now elevating to another level. I expect more frameworks to follow, and there are obvious concerns around security (hence NemoClaw) and also how it overlaps with other things like Cowork or Dispatch's custom agents. But what it has highlighted for me is how stitching together a few core elements — a multimodal gateway, an orchestration layer, and a memory system, all sitting on top of LLMs — gives you the ability to create so many powerful tools, and really quite easily.

So here are the main things I find compelling and why I am currently loving OpenClaw as a hacking environment.

Connectivity as a Core Competency

Consider something as simple as communicating with your bot over Telegram or Slack and giving it an identity. This is the default: rather than meeting a tool within the bounds of where it works, it meets me where I work or communicate. This ability to dispatch tasks as I flow through my own work — and to move easily between all my devices — feels like a basic addition, and was always possible via other systems, but it feels so much more natural with how OpenClaw gets set up.

Separating the backend agents and workers from your direct interface, and communicating with them remotely and asynchronously, is quite powerful. I can now be sitting on my United free WiFi messaging service going back and forth with my agent spinning up coding projects — without needing to pay the $8 for full WiFi.

Orchestration Ease

Creating specialised sub-agents and a master delegation agent is huge for me — perhaps I should have put this first. For sure I could get better at prompting and serve as my own delegator, but the user experience of throwing out a coding project and having a primary agent spin up a sandboxed coding agent with its own set of tools is great. That same orchestrator can simultaneously spin up a simple admin agent to review my todos and add new ones from emails. The ability to delegate and kick off n number of asynchronous processes is powerful. Further, being able to sandbox each with a limited set of tools and file system access gives you much better constrained outcomes.

I previously used — and still use — Claude Code and projects, but they are just not as clean. They require constant switching between repos or project contexts, whereas here I want something done and the cognitive load of how to do it is removed. That is one of the great characteristics of a trusted employee: hand over your biggest problems and know that in the background they are being resolved, with updates surfaced when you want them.

I set up my OpenClaw like an employee, with its own email address, its own git repo, and so on. It can add invites to my calendar on one account but cannot read emails I haven't tagged it on.

Cron, my old friend, is also back in play — everything from daily check-ins to a senior engineer agent that runs once per week with Opus/Sonnet to review logs coming in from Langfuse and improve itself. The closed-loop improvement and decisions on where to improve, based on its own prompts and when to spin up sub-agents, are great. I have the agent manage its own list of improvements in Notion, which I go in and prioritise for the coding agent to execute on as part of its weekly sprint.

Finally, on openness: I remain a big fan of Claude even if they have reduced the throughput available on my subscription. But the reality is that the model ecosystem is 1) moving very quickly and 2) diverging enormously in cost. I ended up using OpenRouter for billing simplicity, through which I access most models. This means each of my sub-agents is optimised across Gemini, DeepSeek, and Anthropic models, which has brought the daily running cost down to a few dollars — I'm not hammering Sonnet when I just need Flash Lite. Experimentation and flipping to a new model or creating new agents is so quick, and the system can self-improve. All of this adds enormous power for the solo developer. It does create a highly personalised setup, so my configuration might not suit everyone, but for me at the current state of the product landscape this is more than fine.

Finally, Memories

This one deserves more time, and realistically I am still experimenting further with other memory systems. Karpathy's recent post on LLM knowledge bases built on top of Obsidian is super interesting and very much on my roadmap. Right now the basic OpenClaw memory system in markdown is really effective, and the only thing I have extended on top of it for the moment is a Supabase Postgres database that stores metrics data — such as for my fitness agent, which powers weekly analytical jobs. But this is an area with so much more potential.

Final Reflections

Like all areas of society, science, and engineering, compounding is an awesome thing to behold — and this is what I see OpenClaw starting to unlock. Linking together gateways, memory, and orchestration/scheduling, and creating these wrappers around tools that have been out there for some time, is unlocking huge potential. As Jensen Huang put it at GTC 2026, OpenClaw is the operating system for personal AI — and I'm inclined to agree.

Got everything I need. Here's your proofread blog with all references linked:

---

OpenClaw Setup: Beyond the Hype

I have been getting deep into deploying my Claws — both with OpenClaw and NemoClaw — over the past two weeks. I have been tracking the Reddit feed closely over the last month and finally found time to build. It is a very fun extension on top of agents, and an easy framework to get into that surprisingly yields not just fancy fluff but some genuinely great tools.

In summary, as of writing this, I have an exec assistant setup (which manages a team of subagents) that handles everything from meeting invites to drafting emails, taking in household jobs to fix, reviewing and critiquing my blog as an editor, and most recently acting as my personal trainer and coach.

This falls into another tool — just like my customised podcast app — that I have quickly fallen in love with and made part of my day-to-day life.

What I think about most is the power that comes from providing an easy gateway to integrating a set of fundamental tools that agents are now elevating to another level. I expect more frameworks to follow, and there are obvious concerns around security (hence NemoClaw) and also how it overlaps with other things like Cowork or Dispatch's custom agents. But what it has highlighted for me is how stitching together a few core elements — a multimodal gateway, an orchestration layer, and a memory system, all sitting on top of LLMs — gives you the ability to create so many powerful tools, and really quite easily.

So here are the main things I find compelling and why I am currently loving OpenClaw as a hacking environment.

Connectivity as a Core Competency

Consider something as simple as communicating with your bot over Telegram or Slack and giving it an identity. This is the default: rather than meeting a tool within the bounds of where it works, it meets me where I work or communicate. This ability to dispatch tasks as I flow through my own work — and to move easily between all my devices — feels like a basic addition, and was always possible via other systems, but it feels so much more natural with how OpenClaw gets set up.

Separating the backend agents and workers from your direct interface, and communicating with them remotely and asynchronously, is quite powerful. I can now be sitting on my United free WiFi messaging service going back and forth with my agent spinning up coding projects — without needing to pay the $8 for full WiFi.

Orchestration Ease

Creating specialised sub-agents and a master delegation agent is huge for me — perhaps I should have put this first. For sure I could get better at prompting and serve as my own delegator, but the user experience of throwing out a coding project and having a primary agent spin up a sandboxed coding agent with its own set of tools is great. That same orchestrator can simultaneously spin up a simple admin agent to review my todos and add new ones from emails. The ability to delegate and kick off n number of asynchronous processes is powerful. Further, being able to sandbox each with a limited set of tools and file system access gives you much better constrained outcomes.

I previously used — and still use — Claude Code and projects, but they are just not as clean. They require constant switching between repos or project contexts, whereas here I want something done and the cognitive load of how to do it is removed. That is one of the great characteristics of a trusted employee: hand over your biggest problems and know that in the background they are being resolved, with updates surfaced when you want them.

I set up my OpenClaw like an employee, with its own email address, its own git repo, and so on. It can add invites to my calendar on one account but cannot read emails I haven't tagged it on.

Cron, my old friend, is also back in play — everything from daily check-ins to a senior engineer agent that runs once per week with Opus/Sonnet to review logs coming in from Langfuse and improve itself. The closed-loop improvement and decisions on where to improve, based on its own prompts and when to spin up sub-agents, are great. I have the agent manage its own list of improvements in Notion, which I go in and prioritise for the coding agent to execute on as part of its weekly sprint.

Finally, on openness: I remain a big fan of Claude even if they have reduced the throughput available on my subscription. But the reality is that the model ecosystem is 1) moving very quickly and 2) diverging enormously in cost. I ended up using OpenRouter for billing simplicity, through which I access most models. This means each of my sub-agents is optimised across Gemini, DeepSeek, and Anthropic models, which has brought the daily running cost down to a few dollars — I'm not hammering Sonnet when I just need Flash Lite. Experimentation and flipping to a new model or creating new agents is so quick, and the system can self-improve. All of this adds enormous power for the solo developer. It does create a highly personalised setup, so my configuration might not suit everyone, but for me at the current state of the product landscape this is more than fine.

Finally, Memories

This one deserves more time, and realistically I am still experimenting further with other memory systems. Karpathy's recent post on LLM knowledge bases built on top of Obsidian is super interesting and very much on my roadmap. Right now the basic OpenClaw memory system in markdown is really effective, and the only thing I have extended on top of it for the moment is a Supabase Postgres database that stores metrics data — such as for my fitness agent, which powers weekly analytical jobs. But this is an area with so much more potential.

Final Reflections

Like all areas of society, science, and engineering, compounding is an awesome thing to behold — and this is what I see OpenClaw starting to unlock. Linking together gateways, memory, and orchestration/scheduling, and creating these wrappers around tools that have been out there for some time, is unlocking huge potential.