Interesting video on this -https://youtu.be/6OXE65fjjsU
1 Like
new “ultra lightweight” alternative to openclaw called nanobot
1 Like
I don’t see the point of giving an agent root level access to my entire life (by letting it into my iCloud/google accounts etc), what sort of things are people planning to do which they currently can’t using this tool?
What openclaw is able to do is to make anyone run an autonomous agent. Be it for work or spam ( https://www.reddit.com/r/aiagents/comments/1qywy30/in_china_this_is_already_how_some_people_are/)
It is able to do many rights - you can spin up agents , get tasks done, get response now or later - all under one config umbrella.
Others are catching up but are not there yet. Nanobot can talk or run jobs. But it cannot converse > do async actions and revert. That is one big missing point.
Openclaw is like working in a firecracker factory. One wrong move and you are blasted. Unfortunately setting up guardrails is not out of box or a piece of cake .
1 Like
This is pretty good advice and secure way to setup openclaw
Openclaw founder has joined OpenAI.
1 Like
now it’ll be nerfed
Just Set Up OpenClaw and really Loving It !
Mine’s installed in an unprivileged LXC container, and talking to me over Telegram
Used Opus for initial bootstrap (Helps with setting up really good identity and soul for the agent)
using Cheaper Qwen and Deepseek v3.5 (All through openrouter) for regular tasks (Depending on what i need it to do).
Ultimately plan to use MiniMax but a bug prevents minimax usage through openrouter currently
Really impressed by how extensible it is with skills, and how light weight it is in terms of resource usage ( I have put it on a Ryzen 7730u Mini )
Once its identity, soul, and user’s identity are properly setup, it kinda feels like a real assistant…. but i have set it up to feel like “one of us”
Planning to setup few more agents other thn the main one, especially an ncert agent (for my daughter). I already have all NCERTs converted to markdown, cleaned and neatly processed and ingested into llamaindex, will teach my ncert agent to do RAG over it.
Would love to hear if anyone else is tinkering with it, and got ideas for helpful automations
6 Likes
I’m planning same on one of my Mini PC setup, However, I’m still not able to decide which model to go as i wanted to integrate with my claude code setup but now Anthropics banned it so only able to use API which is costly so thinking about either to go with GLM 5 which is offering 3 month at 80 USD or Kimi K2 which i think its still little context aware compare to GLM 5 or should go with OpenAI Codex, Does anyone have any suggestion on this?
That’s a really nice use case, Can you talk more on the hardware that you are using
1 Like
I am actually “Eating my own dog food”
I am running openclaw on this Ryzen 7730u mini pc
It also runs the n8n and some other playground containers for random stuff.
consumes around 5.5w Idle, When its running in monk mode (Headless, no io devices connected except LAN). This workload utilizes only fraction of its capacity though.
Directly or via docker?
Inside unprivileged LXC container - no docker.
So it kind of has access to every thing inside that lxc
I had tried with docker before, the experience was not tht great
So created a seperate lxc just for openclaw.
1 Like
One thing lurking in my mind is, feed and build a RAG on all the books written by Gandhi. and build a bot with his persona, thoughts, beliefs and values. Even though opinions about him vary But i think it can unlock his ideas.
Just don’t let it preach and don’t get Token anxiety while doing so.
1 Like
This video gives an excellent short explanation of the complete architecture of OpenClaw and how it works. He has explained it very nicely.
And more importantly, he has demonstrated the RISKS which are involved in using it and how to mitigate them. Please watch from timestamp 5:44 to become aware of these risks -
Thanks to everyone for sharing helpful insights on this thread.
PS: This video is not saying that you should not use OpenClaw, instead it just tells about the vulnerabilities involved and how to use it wisely.
1 Like
Word of advice to anyone running local LLMs. Don’t run them continuously at full load if your motherboard, CPU, GPU or SMPS is close to its limits. Continuous peak draw is different from gaming bursts.
You need proper headroom on the SMPS so it handles sustained load without degrading or failing.
My SMPS handled heavy gaming fine, so I ignored the fact that my desktop was already near its power ceiling. After running a local model on the GPU continuously for around 4 days, the SMPS failed. It started pushing unstable voltage and caused super bright electrical arcing near the GPU power connectors.
If you’re building or upgrading for local models, overspec your PSU. Aim for at least 30 to 40 percent headroom over real peak draw. Don’t trust cheap high-wattage ratings. Sustained compute load stresses hardware differently than gaming.
I’m replacing the SMPS and using this as an excuse to build a higher spec desktop. Hoping the GPU survived.
2 Likes