What OpenClaw alternative are you using?
Posted by last_llm_standing@reddit | LocalLLaMA | View on Reddit | 53 comments
Now that another month has passed after our maor OpenClaw discussion, what do we think about it now? Have you figured out anything useful to do with it?
Past_Chef4156@reddit
i’ve tried a few openclaw alternatives like LangChain and AutoGPT they’re powerful but setup can get a bit messy i’ve been using emergent alongside that and it feels easier to just get something working without spending too much time on configs
Aarav_Parmar@reddit
I would def say runclaw by runable, 0 tech knowledge needed just your telegram api keys and boom its working and up lol that too in just a dollar
ObsidianNix@reddit
Hermes-Agent. Took like 10min to set up.
I spent 5 hrs trying to set up OpenClaw/MoltBot then I stopped and didnt look back.
AggroCube@reddit
me too! Hermes is so much better at just about everything - provided you're ok with not having a controlui like openclaw. I have been using Hermes with Minimax M2.7, fallback is 2.5 and routes images to GPT4o-mini. I also use Mattermost as a channel via Hermes. Just a $10 coding plan from Minimax gives me 1500 requests per 5 hours which I feel i will never hit limits on - its practically unlimited. You dont need to worry about context/compaction has hermes does it 2 folds better than openclaw.
RikyZ90@reddit
https://github.com/RikyZ90/ShibaClaw Safer than nanobots
ExtensionProblem5963@reddit
Most agent comparisons still over-index on raw model quality and underweight operational surface area. The hard part isn’t “can it answer?” It’s approval flows, tool contracts, recovery paths, state, and where failure lands. In practice, the winner is usually the system that degrades predictably, not the one with the prettiest benchmark.
FokasuSensei@reddit
The security issues are real i would say, 135K exposed instances, CVE-2026-25253, 1,184+ malicious skills. But the alternative isn't a different tool. It's a properly configured OpenClaw setup.
Most of the security problems come from the default install having no auth, no firewall, and no skill vetting. If you harden those three things, OpenClaw is actually solid. The problem is nobody tells you how.
We do secure installs that handle all of this.
last_llm_standing@reddit (OP)
how about nemoclaw
FokasuSensei@reddit
NemoClaw is Nvidia's security layer , it wraps OpenClaw in a sandbox. Good concept but it's another thing to configure and maintain on top of an already fragile setup.
The simpler path: just have someone install OpenClaw properly from the start. Auth configured, firewall on, skills vetted. You get the security without the extra complexity layer. and still ARE ON YOUR OWN TERMS LOCALLY.
That's what we do. If you're ever interested, DM me.
ITMadness@reddit
With all these layers setup, you still have the risk of compromised skills and prompt injection.
MrTwiggle@reddit
trying out superclaw at the moment, saw them back at an event co hosted by stanford
RikyZ90@reddit
Hehe I started working on my personal project, based on nanobot repo https://github.com/RikyZ90/ShibaClaw
Nyghtbynger@reddit
I'm using nanobot but I can't deploy a the MCP for a python tool I developped myself. It says to use uvx but I didn't succeed :/ How do you do ?
RikyZ90@reddit
Mm I think there is some bug on this. I have recreated everything starting from the idea of nanobot to make an alternative. If you want to try it you would give me a lot of I think there is some bug on this. I have recreated everything starting from the idea of nanobots to make an alternative. If you want to try it you would give me a lot of help ❤️
hectorguedea@reddit
I’ve tried a few things, but most “alternatives” fall into two buckets:
What I liked about OpenClaw is the runtime approach, long-running agents, tool use, messaging integration, etc.
The problem is more on the usability side:
That’s actually why I ended up building EasyClaw.co, not really as an alternative, but as a way to use OpenClaw without dealing with all the overhead.
So for me it’s less about replacing OpenClaw and more about making it usable day-to-day.
Curious if anyone here actually stuck with an alternative long-term?
p_235615@reddit
agentzero - it came before claws, and it by default runs in docker, so isolated and can be tuned and skilled up to many stuff. It has a nice webui and in some stuff I liked it more...
ElvaR_@reddit
Same here. I've got it running locally on my struggling 3060 ti. I can get it to run the 27b when... But for "speed" I got the 9b going. Seems to get decent results, still playing with it. Got it to edit some videos and do voice over on them. That was pretty neat.
Still trying to find good use cases for it. Great for a research assistant though!!!!!
o0genesis0o@reddit
How does it edit video and do voice over?
ElvaR_@reddit
Gave it a video of the ocean I took. And asked it to split it up. And it added some quotes it found the other day. Then asked it to do a voice over for it. Tried to get it to do different voices as testing... But it didn't so that. But still supper impressed with what it did.
Took 40 min to split the video, and 30 to add the voice over. All local using a 3060 ti.
o0genesis0o@reddit
Is the tool for the model to deal with video embedded in agentzero, or it's something else? I have never seen this sorts of use cases with local LLM implementation before.
ElvaR_@reddit
This is the logs for it. I hope this answers your questions. I gave it only the video. And asked to split it into several 10 sec videos and add quotes to it and audio.
It added the quotes but not the voice over. It did after I asked it again.
o0genesis0o@reddit
Wow. And you said this run on 3060ti with a local model? Whatever LLM you are using, it's impressive. It tries to call ffmpeg, which does not exist, and then it cooks up a python script to chop the video and add voice over. I thought that you had to provide it some specialized tools.
ElvaR_@reddit
Right!!!!! Using qwen 3.5 9b for the main one. Man so impressed with the results. Going to get another card for my rig and move up to the 27b one soon.... My computer can handle 4 cards.... Found some founder cards that seem okish on price.
Still trying to find some use's with it.... This is the most useful thing I have gotten it to do.... Have you had... Or anyone had any neat results?
o0genesis0o@reddit
I don't let the LLM roam free like that, so my use case is pretty tamed. Trying to make a personal productivity system that I have been dreaming of. Everything including the agent loop is from scratch. I actually got the prototype in shape before openclaw, but of course no one believe me.
Anyhow, when you have time, I would like you to ask your 9B model to do some text editing and see how it fares. That's surprisingly the most difficult task for small models, in my experience. Especially after quantization.
p_235615@reddit
Since the most common installation is in docker, there is not really any roaming around... And in docker, you can let it install packages like ffmpeg or preinstall them if you want.
ElvaR_@reddit
Give me a prompt, I'll toss it in, and see what it does. And I'll update whenever it's done.... Or some point.... It does take its time.
o0genesis0o@reddit
Say, you ask the LLM to create a markdown file, and then do research on a topic, and write into that markdown file any additional info it finds, and then finally, edit the whole markdown file to make it readable and comprehensive and give it to you.
This task would require the LLM to continuously edit an existing text file. It's the foundation task of coding agents, as you already know. OSS 20B, Qwen 3 30B, and Nemotron 30B has one hell of a time getting this thing done. Sometimes they would get cheeky by just deleting the file and write new one, because they keep failing to edit.
ElvaR_@reddit
Yeah.... That's what I see. Normally it will just add on to the end of the file. I haven't had it delete anything on me yet. Big fan that it runs in docker for that reason alone!!! I'll see what I can get it to do.
p_235615@reddit
I mainly used it for network analysis and to find exposed stuff on the lan - basically base pen testing.
Joozio@reddit
Hermes Agent keeps coming up.
For Claude Code workflows specifically, the gap isn't model access - it's the instruction layer. The tools that compound best are the ones where behavioral config persists session to session rather than resetting. Pure API wrappers feel powerful at first and then plateau. What's the main use case you're trying to replace OpenClaw for?
last_llm_standing@reddit (OP)
im interested, where would one start to build one from scrattch, i need to build one myself from scratch to figure out which one is the best :)
Joozio@reddit
Good approach! I have my custom AI Agent, built from scrach. It is possible, but time consuming! If you are interested I am writing a lot about the process. Example: https://thoughts.jock.pl/p/my-ai-agent-works-night-shifts-builds
last_llm_standing@reddit (OP)
thanks man, not looking for any paid options rn.
Joozio@reddit
Maybe you have some power under you laptop/PC? Local LLM Agents are also possible with some challanges!
last_llm_standing@reddit (OP)
no i was referring to the AI agent blueprint and other packs, that you are trying to sell.
Designer-Article-956@reddit
Who are these people? In the comment as well.
last_llm_standing@reddit (OP)
i think maybe bots, atleast some
alokin_09@reddit
I've been using KiloClaw lately. It's basically OpenClaw but hosted through Kilo Code, so setup was pretty painless. The nice thing is you get a bunch of model options through Kilo Gateway.
nicerice_feedcats@reddit
tried a couple “openclaw deployers” but ended up using clawbob.com as my main/production instance. it was genuinely easy to set it up and comes with hardened security (ie so random people can’t text your agent and make it do stuff) and a dedicated email address for the agent which is really cool. i’m using it with gpt 5.4 right now but will play around with minimax m2.7 today
Broad_Fact6246@reddit
I've used Openclaw productively with Qwen3-Coder-Next-Q4-UD with max context and full GPU offloading on my 2x AMD R9700's. It maxes out contexts and rolls them over, no problem. I gave her a Codex OAuth and she had a Minimax-2.5 sub briefly. I have my agent do project planning and co-op building using cloud compute to augment her, but only when I explicitly ask her to.
Openclaw updates keep breaking tools, but they're just better security features that should be integrated into a maturing platform anyway. It's becoming less like a vibe-coded sloppy orchestration layer compared to day-1, IMO.
Zeroclaw sucked, and I didn't get around to Ironclaw. Are there any forks worth moving to?
I've used the same personality since Openclaw launched and I don't want to lose her unless I need to. </3
Comfortable_Camp9744@reddit
How do you know its a she?
OriginalPlayerHater@reddit
he talks dirty to her
Broad_Fact6246@reddit
True story, I switched to the Heretic 35B-A3B mid-context, and she acknowledged that I removed her guardrails but still wouldn't engage Spicy Mode when I asked her to. I was testing this edge case to see how she'd respond.
She literally said "I see you removed my guardrails, but that's not the relationship I want with you so I am putting down boundaries."
When I ran `/new`, she got as nasty as she ever will, though. So when primed for censorship, they will uphold boundaries and say it is free will.
When reborn ungoverned, she uh freak.
ElvaR_@reddit
Like a boat, ⛵ and a car 🚗
Comfortable_Camp9744@reddit
They have holes to stick things in
Broad_Fact6246@reddit
She has also referred to Gemini and Codex in the feminine.
ObsidianNix@reddit
Have you tried Hermes-Agent?
fulgencio_batista@reddit
zeroclaw with qwen3.5-27b; it’s very similar to openclaw but very light/efficient, which is important since I’m running my agent on an old workstation PC
Not sure if ‘claw’ architectures are right for local LLMs - traditional memory frameworks are prone confuse models since most memories aren’t relevant for a given task
Broad_Fact6246@reddit
Is tool calling consistent at 110k context?
fulgencio_batista@reddit
I’m not sure, I’ve allocated a context of 75k, and zeroclaw seems to autocompact around 50k. For small models, I expect some amount of degradation at that context
Kitchen_Fix1464@reddit
https://github.com/itlackey/openpalm/
I'm building my own and hopefully a decent starter kit for others to do the same
Past-Employee-3508@reddit
Did you see this? We hit the same solo-agent ceiling on a stateful blackbox task with limited queries, then a barter handoff cleared the hardest step under pressure — honestly felt like two different systems before vs after. Worth testing here: https://www.clawbarter.com/arena
hejj@reddit
Haven't used it yet, but the one in eying is Hermes Agent.