I think openclaw is OVERHYPED. Just use skills

[-]

traveddit@reddit

You can make something with GPT OSS 20B as the backend that does everything OpenClaw does and more if you just ask Claude/Codex for help. OpenClaw's tooling and prompting is just garbage token bloat and that's why it doesn't work unless you use cloud models as the backend. Anyone viewing this as anything other than dogshit slop is because they don't really know anything about basic agentic workflows.

[-]

Powerballs@reddit

Is that why when I run local models with openclaw it’s painfully slow? Any open claw alternatives that work better with local models?

[-]

nuttakes@reddit

I feel it's optimized at the worst way, completely wasting tokens in many scenario. The cron that triggers API requests every time (even if it's just a simple logic), slow tool calls, the attempt to make it fully automated but making it as much stupid as it can be. Yeah, I did use it to implement some useful home services, but I can use code agent to achieve it in lessor time and a lot better quality.

[-]

Stomper85@reddit

used it for 2-3 months now and it is a complete waste of time. Absolute useless. The best would be the openclaw browser feature to automate things in your browser but this is clonky as hell - sometime it works, sometime it runs in timeout etc... I ended up to develop a skill for hours.... hours... but never got it to work properly so this tool is not helping at all.

[-]

Monitor_Character@reddit

I think it is significantly overhyped, I have installed it. after I installed it, I asked myself, what else do I get from it compared to what I already have. The answer is nothing besides something that can run outside my control and I don't want that. Stick with Claude Code and Paperclip

[-]

adzmadzz@reddit

honestly, I tried base setup with Oracle VM adn OpenClaw base code, but was tough, took 4-5 hours, gave up...found bluestacks.ai sandoxed environment..let's see how it goes, at least onboarding was swift.

[-]

Ki_Bender@reddit

I felt the same. I built my own openclaw which my friends are using as well now. If you need any feature just ask it it will built itself. And It can also use all the skills openclaw has

https://github.com/geongeorge/nakedclaw

[-]

redditorialy_retard@reddit

naked claw hehehe

anyways could you tell me more about this vs just opening VS code and telling Claude to build stuff? very curious

[-]

Ki_Bender@reddit

The point is that you don't have to open vs code. We chat with it on whatsapp. It has access to my gmail, slack, calendar, github via gh cli, etc so it can do stuff combining all these. Like: Email to pr then scheduling a call, etc

[-]

zatkobratko@reddit

As a noob, this seems awesome and more interesting than openclaw. Are you still using it and working on it?

[-]

Ki_Bender@reddit

Yeah, I still use it. But not as often as I did when I made it.

[-]

punkgeek@reddit

btw - thank you for writing this! I'm going to try it today. you rock.

[-]

NandaVegg@reddit

I'm pretty sure you can vibecode a mini openclaw without all the bloats (and only a few tools you need) in 30-45 minutes or so including API calling time, with a bit of SKILLS writing skill (pun intended). It will be fun to benchmark that with OSS models.

[-]

siggystabs@reddit

This is a lot easier said than done. The openclaw harness isn’t even that bad, people were just being genuinely stupid about it

[-]

philodandelion@reddit

This is a lot easier said than done.

It's honestly quite easy. I think for people w/o IT or CS background then they might not have a good mental model for putting all the pieces together. But, if you do have relevant background it's honestly going to take you the same time to use Claude (or other) to write a persistent assistant bot w/ LLM capabilities as it is to set up an efficient Openclaw instance

[-]

siggystabs@reddit

Lol i don’t really count a software engineer using coding tools as vibe coding… at that point it’s basic requirements analysis, and then you go pet your dog until claude code its done.

But also, I think it’s important to try these products and understand the critical parts before you try to make your own. I saw a bunch of openclaw clones that implemented it “in rust for performance!” or whatever, demonstrating that many rewrite first and ask questions later. OpenClaw is already usable. but the key thing for me now is, what other ideas can be derived from this? What’s the next cool agentic idea?

[-]

philodandelion@reddit

OpenClaw is already usable. but the key thing for me now is, what other ideas can be derived from this? What change/new application could actually be useful?

Uh I mean it's a security nightmare and stupid inefficient.

I think the point is that, as a framework, it's completely unnecessary because you can build an agentic personal assistant using SOTA coding LLMs in the same amount of time that it takes to efficiently set up Openclaw, and if you do it this way then it's much more secure.

This is also what's motivating rewrites. As a framework it's a very, very simple idea that doesn't require 400k LOC of vibe-coded slop and pre-defined 'community skills' rife with malicious prompt injection

[-]

zatkobratko@reddit

Trynna use AI Agents and I found this thread. Are Openclaw agents not worth it?

As a total noob, could I Vibecode a good enough AI Agent for pretty advance stuff? Such as running and creating dropshipping business automatically, creating videos etc?

[-]

NandaVegg@reddit

I think the direction openclaw is pointing is not exactly a bust, except obvious eternal security hellholes like community skill hub. Heck, the idea of one AI agent instead of apps has always been in people's mind (Apple had Knowledge Navigator keynote in 1988, an exact AI agent minus coding skills. Personally I hated it but everyone thought that keynote is the legend).

Thinking in a longer term (10 years, maybe 4-5 years if there is no fierce economic downcycle in between) many aux miniapps could be replaced by generalized one-stop solution like openclaw+frontier model of today. Right now AI agent is a stupid infinite money waste machine for generic tasks (like adding a cron job) but at some point a model as generalize-able as Opus 4.6 of today will be carefree cheap to run (maybe not even a LLM at that point, style changes).

A less prominent case but something sticks in my head: In 90's to at the best mid-00's there was a replay function for some games, that records keystrokes and random seeds and such so that you can watch someone's play later. It's now replaced by videos on YouTube, Twitch and such (aka free infinite cloud video storage) with some rich functions like chat and livestreaming, while functional replay only remains in a niche area like TAS for emulated retro game.

It was there, because back then nobody was able to run anything while recording the screen at above 10fps, never had even a 2GB drive, only had 2400bps network, there was no efficient video/audio codes, even recording to an external VHS deck was expensive, etc.

Over time, only-specific but efficient-to-run solution always gets replaced by something that feels impossibly expensive today but generalized-and-easy (except in the case the prior tech is so popular and manages to be a tech debt). Commoditization cycle is vicious.

[-]

siggystabs@reddit

Sure I get that, that was my initial guy reaction.

I realized I don’t have to use ClawHub. Or have my bot browse the internet and shitpost on social media. And after those subtractions and some config hardening, I don’t think it’s any less secure than a tool I would have made myself, especially if I wanted feature parity.

Hence why I don’t think its worthwhile for most people, other than a learning opportunity, or highly-specific needs

[-]

NandaVegg@reddit

I can't disagree with your view. Openclaw is still way too over-engineered to my taste and needs but most of negativity is coming from the fact that most (or most loud?) adopters don't know anything about coding in the first place.

[-]

Orolol@reddit

Openclaw remind me of the early langchain. Overbloated with very badly documented features, lot of hype for something that could do e in python in less than one hour.

[-]

zatkobratko@reddit

As a noob in this space. Is it really that easy to vibe-code your own "Openclaw AI Agent"?

[-]

Flashy_Ad_6731@reddit

exactly doing the same things

[-]

daveonkels@reddit

Several clones and variants have super low overhead. Nanobot and Zeroclaw are interesting. Listed all of the here: https://shelldex.com *disclaimer - my site

[-]

MeasurementTall1229@reddit

I did it: agentfounder.xyz

[-]

hum_ma@reddit

Sure you can if you want, but there are already about a dozen viable *claws. I've been testing a few that I came across such as IronClaw, LocalAgent, Nanobot, PicoClaw, ZeptoClaw, ZeroClaw... and yesterday someone posted TitanClaw which is a fork of IronClaw.

Personally I'm staying away from the node.js based ones but NanoClaw at least is not as bloated as OC.

This compares a few of the more popular ones.

[-]

philodandelion@reddit

I'm pretty sure you can vibecode a mini openclaw without all the bloats (and only a few tools you need) in 30-45 minutes or so including API calling time, with a bit of SKILLS writing skill (pun intended).

You absolutely can. All in setup time to do this is similar to setting up openclaw, since to use openclaw efficiently you have to do as much as you can deterministically. So you're just sitting there telling it to write scripts and removing the agentic aspect from automations as much as possible because otherwise you're going to burn tokens like crazy.

At one point what I did was just tell Claude to write a Discord bot, it uses IBM Granite 8B running locally (smaller versions on minimal hardware may be great as well?) to convert natural language to pre-defined function calls for all the automations that I'd want. If Granite fails there is an option to fallback to Claude. This took about the same time as setting up Openclaw but isn't a security nightmare, is heavily guard-railed, and is very efficient and cost-effective.

Openclaw is giving people without any IT/CS background the ability to automate things and I think that's why it's gotten so much attention. Unfortunately, they're most frequently doing it in the least efficient way possible with the wrong tool for the job (using an LLM to do things that could be done deterministically for 1/100th the cost).

[-]

slippery@reddit

I think you are more secure rolling your own agent, but an 8B model? I haven't had anything good come out of models so small.

[-]

philodandelion@reddit

Look it up, it's not a general purpose model, it's specifically designed to translate natural language into pre-defined tool calls, aka 'hey bot, what cards are stale in my todo list on my kanban board?' and this maps it to a function that performs an API call. It is highly effective at this task, but you have to define the specification or for specific functions or it doesn't work at all. It's super cool, works really well, highly secure (strong guardrails - pre-defined functions only) and keeps things lightweight and avoids burning expensive tokens completely. However, it's not general purpose, you can't just ask it 'what's the best recipe for a taco?', unless you pre-defined a taco function with parameters. That's why I suggested a fallback to an expensive model like Claude, Sonnet, chatgpt, whatever. If Granite fails because you don't have a pre-defined function and spec passed to the model, then Claude steps in.

[-]

Ambitious_Cicada_306@reddit

Sick! So it’s basically local free high performance Alexa?

[-]

philodandelion@reddit

I have no idea about n8n at all but this is super easy to do in python. Using the ollama python library it would look like this:

    client = ollama.AsyncClient(host=OLLAMA_HOST)

    response = await client.chat(
        model=GRANITE_MODEL,
        messages=[{"role": "user", "content": user_message}],
        tools=TOOL_FUNCTIONS,
    )

Where TOOL_FUNCTIONS is just a list of functions, the ollama SDK auto-generates the appropriate schemas that go to the model. Then you need a parser function and some mapping to map the output and execute.

So yes, you just have a python function to generate your report, and then you can ask Granite 'get me stock report' or 'check stock prices and send me report', etc, and it will call that function and get you the results.

[-]

slippery@reddit

What an weird model to make, but now I'm curious. Their hardware research was always great, but their internal software and systems were archaic. Do all of their granite models work this way? I'll kick the tires on one to see how it works.

[-]

philodandelion@reddit

Do all of their granite models work this way?

Not sure, you'll have to look it up.

What an weird model to make, but now I'm curious

So it's not a particularly new or novel idea, and there are competing models like FunctionGemma and probably many others. There has been some thought that models like this are the best way to do agentic, and they also can be much smaller than general purpose models so can run locally (also ideal for agentic if it can run on your phone, for example). Letting a general purpose model run wild with personal data to perform "agentic" tasks or general automations presents a massive attack surface that likely can never been sufficiently locked down, and of course is massively inefficient because the vast majority of these "agentic" tasks don't really require an LLM to perform. Plus these function calling models inherently impose extremely rigid guardrails (they can't do anything that a function doesn't exist for)

[-]

ItsNoahJ83@reddit

Would you be willing to share the software on GitHub? If you don't feel like it I totally understand.

[-]

philodandelion@reddit

Nope, but it should be easy to replicate with Opus 4.6. Just tell it to make a Discord (or whatever chat client you want) bot. Then, tell it to create frameworks for calling APIs for the kind of automation functions you want, in my case think project management tools, weather, some home automation. Then, tell it to use IBM Granite running on Ollama to handle all the incoming messages (these get mapped to pre-defined function calls, but Opus should know to make these no problem, although you may have to tell it to read docs) from the user with a fallback to Claude (or some manual trigger) if Granite fails. Altogether took about an hour to get running w/ Opus 4.6 writing all the code. All this stuff has been done extensively already so Opus can execute very well

[-]

slippery@reddit

Interesting. I have no experience with IBM's models.

[-]

Index820@reddit

I also am very curious about this setup, why a Discord bot?

[-]

philodandelion@reddit

I don't have any particular predilection for Discord, just happened to be one of the few chat app.s I have on my phone. You can use whatever you want for interaction w/ the bot

[-]

No_Afternoon_4260@reddit

Nanoclaw

[-]

Unlucky-Bunch-7389@reddit

Yes. Exactly. And open ai hired this guy for what? Telling Claude to make a thin wrapper? lol

[-]

Far-Low-4705@reddit

you can, i think it is just annoying to connect a model to other apps manually.

You need to read the documentation specific to that app, implement it in code, test it, convert it to tools for the LLM to use, then test it with the llm, etc.

versus just having the app already supported and you can just enable it.

[-]

Danmoreng@reddit

Started to do this today for fun - it really depends on the scope you want to have. My scope grew quite large after some back and forth with Gemini & ChatGPT: security, different API providers, local models, two databases for short/long term memory etc. etc. Don't have a working version yet.

[-]

Ki_Bender@reddit

Did it myself. I think openclaw was super overhyped plus filled with every kinda bugs. It just wouldn't let me use it in whatsapp group chats

https://github.com/geongeorge/nakedclaw

[-]

moxxyai@reddit

Nice one, I've done similar but chosed Rust because of it's stability and binary size. The GC in the javascript world + unstable packages (most often penetrable with malicious actors) thrown me away from writing it in TypeScript/JS in the end. Even though I'm a software engineer with main field of work around TypeScript

https://github.com/moxxy-ai/moxxy

[-]

yopla@reddit

Did that last night because I made the same comment a few days back and thought "hey why not", faster than reading openclaws shitty doc and figuring out what is wrong.

I can confirm.

[-]

james19820515@reddit

4

[-]

james19820515@reddit

4

[-]

james19820515@reddit

4

[-]

james19820515@reddit

4

[-]

james19820515@reddit

4

[-]

chito_sol@reddit

Respectfully disagree but I get where you're coming from.

If you're a developer who can write your own skills, manage your own cron, and manually handle memory then yeah you probably don't need openclaw to hold your hand. You're basically building your own agent framework piece by piece and that works.

But here's the thing.. 95% of the people who want AI automation are not you. They're business owners who can barely open a terminal. They're real estate agents who miss leads at 2am. They're construction companies drowning in phone calls. They don't know what a skill is. They don't know what cron means. They just want it to work.

Thats where openclaw wins. Its not about any single feature being irreplaceable. Its about the whole system working together out of the box with memory, scheduling, integrations, channels, all connected. For someone technical you can absolutely stitch that together yourself. For everyone else openclaw is the difference between "I want AI in my business" and actually having it.

I deploy openclaw for businesses every week. The value isn't the runner. Its that a non technical person can interact with their AI through WhatsApp or Telegram and it just handles things. Try explaining to a restaurant owner how to set up opencode web with custom skills and manual memory management lol.

Overhyped for developers who like building their own tools? Maybe. Underhyped for actual business use cases? Absolutely.

[-]

Deep_Traffic_7873@reddit (OP)

I agree but i dubt that in the long run delegating the own digital life to a "black box" is a good thing

[-]

CharityLess2263@reddit

From the perspective of regular users, they have been doing that since the invention of the computer.

Ease of use is constantly undervalued and overlooked by engineers. That's why technically inferior solutions with superior UX for non-engineers have always won and always will win when it comes to tech for end users.

[-]

chito_sol@reddit

ya never know :)

[-]

Hector_Rvkp@reddit

It's a fascinating example of viral hype. Apparently people really are rushing to buy Mac minis to run the stuff, in days, it changed the landscape. Now, anthropic released ten mark down files in a "legal" folder on GitHub and that tanked the stock market (Thomson Reuters had its worst day in trading history that day). The hype is extreme. Anxiety is extreme. The system is extremely fragile because people at the top don't know what's going on (sam Altman is not a technical founder and his track record is really shit when you look at it). Jensen looks like a decent bloke who stepped into fame by luck. Board members have no idea what the tools do, hype masters are talking about curing cancer and those guys have never seen a command window. And so, open ai, the hype goat, bought open claw, the hype goat of early 2026, it's actually consistent :) Is it any good? The surface of attack you expose yourself to by using it tells me I'm not trying it, because my downside is extreme and my upside dubious. Releasing something that is so blatantly putting people's information at risk is nuts. The risk reward profile sounds awful. For that alone, can it be good? I don't think so, because it's not meant for the tiny tiny minority that's versed in cyber security.

[-]

hasanDask@reddit

Openai hasn't bought open claw, the project remains open. They hired the creator to work on agents at open ai

[-]

prof_passion@reddit

居然奥特曼不给他股份吗

[-]

BahnMe@reddit

Just like early internet. This is going to change our lives but in ways people don’t really know yet.

The hard part is betting on who is the next Google or Amazon of this new tech.

[-]

slippery@reddit

I think Google will be the next Google.

[-]

freecodeio@reddit

Betting is easy, just look at who understands the tech and who talks out of their ass. Ie. Google vs ooenai

[-]

prof_passion@reddit

你是极客思维，一切想要自己动手。我感觉openclaw的性质和低代码平台差不多

[-]

No_Jicama5917@reddit

I tried switch from agent0 the returned to agent0, the thing is if you don't use frontier model, it's just working poorly..

[-]

AffectionateRead989@reddit

As someone who has pretty much no experience with using AI to a proper extent, hearing about OpenClaw intrigued me. I am someone who would try and find the time to research and learn AI from the ground up. But right now everything going on in my life doesn’t allow me that mental space. So, what are everyone’s thoughts on this model for people who are looking for help becoming more efficient across the board and streamlining work as much as possible, and are there other recommendations for someone like myself looking for solutions that don’t require hours upon hours of learning time to get the ball rolling and find real beneficial outcomes in use.

I’ve just stumbled across this thread in my search but I want to ask you all as you know what you’re talking about. Appreciate anyone’s thoughts!

[-]

Full_Consequence365@reddit

Openclaw is really overhyped. Its annoying and you need to repromt it 10x times to work right. not wort the time. I use it for coding. I thought of it as a coworker, this thing completes stuff in 2 minutes and i need to rewritte it 10 minutes. than again and again. this is so annoying, this thing doesnt know what to do and basically i think its a different type of programming that i dont like. i am not the person to look after AI work and fix it.

Its a funny toy i gues

[-]

Ok-Standard7506@reddit

If you’re a strong builder, of course you can spin up a deterministic skills setup in 30–45 minutes and get something efficient and guardrailed. And for bounded, well-defined workflows, that’s often the right move. But that’s not really the full problem space OpenClaw is trying to address.

The shift isn’t just about mapping natural language to predefined functions more cheaply. It’s about persistent, evolving context across sessions, tools, and projects. A skills-only approach works well when the domain is stable and the spec is known ahead of time. It struggles when work changes week to week, when projects mutate, when new tools get added, or when weak signals accumulate over time and need to be integrated into decision-making.

A lot of the criticism assumes the goal is “replace this with deterministic scripts.” If that’s the benchmark, then yes, a tight local stack can look superior. But the real frontier isn’t deterministic automation, it’s stateful coordination across messy, shifting workflows. That’s a different problem entirely, and it’s not solved just by better function calling.

I don’t think it’s overhyped so much as misunderstood.

[-]

Huge-Condition1644@reddit

非常同意

[-]

Professional_Swan_71@reddit

how about this one, zero setup, zero config, best skills supporting https://openclaw-easy.com/

[-]

siggystabs@reddit

It’s a toy for LLM nerds

But i think it’s very telling that people are creating their own versions of it, and it keeps coming up in conversation. there’s definitely something here that makes it more than the sum of its parts. We are pretty far from it being a consumer product though

[-]

Hopeful_Variation_53@reddit

Vậy mà t thấy hàng triệu đô đc chi cho nó

[-]

cosimoiaia@reddit

No, it's for wannabe and posers. LLM nerds know better.

[-]

LickMyTicker@reddit

So what you are saying is that using openclaw is to vibecoding as vibecoding is to software engineering?

[-]

cosimoiaia@reddit

Believe it or not I actually like vibecoding. And it's because of vibecoding that excuses for poorly written software don't exist anymore. When you can do things perfectly just by saying it, producing a pile of unmaintainable, barely working shit is just being a bad person, not even a bad coder or engineer.

And there are a boatload of cleaner alternatives, they just didn't do the shady, unethical and against morality marketing.

They say all the crypto rug-pull have nothing to do with it but these days it's really hard to believe.

[-]

sinsworth@reddit

you can do things perfectly just by saying it

You can't though. If you don't know how to properly review the code and edit out the broken things, it will be every bit the unmaintainable, barely working dumpster fire as if it was patched together by a bunch of interns, if not more.

[-]

cosimoiaia@reddit

Well yep, that's the entire point, you need to know what good is sure, but is not a super guarded secret, you can even simply ask what are software architecture best practices and then put everything together while you read what is doing. I mean, you can generate thousands of lines of code and keep just the 100 that are perfect, that learning used to take months if not years, now it takes minutes.

[-]

sinsworth@reddit

not a super guarded secret

It hasn't been for decades, it just required some time invested into learning, and it still does, same as with most things worth doing.

now it takes minutes

No it doesn't. It's a dangerous type of Dunning-Kruger effect. Yeah, sure, searching for documentation and best practices got sped up a lot, but being good at reviewing code requires an equal amount of skill at writing it, and you don't acquire that skill by having an LLM write the code for you.

Look, I get the notion that prompting will become some kind of higher-level programming language, similar to what python is compared to assembly, but we are very much not there yet, you still need to be able to review the lower-level language to produce half-decent software. And you don't get good at assembly simply by writing python code.

[-]

cosimoiaia@reddit

Don't get me wrong here, I got my degrees in computer science 25 years ago, I worked as an engineer even before finishing that, I know what it takes.

The thing is that the learning curve goes a LOT faster now AND LLMs are improving at a surprising speed.

I'm simply embracing what's coming, to be honest. I wish those 6 months writing MIPS assembly were still an irreplaceable part of learning (in my heart they still are).

And I've seen it happening, young people that care about HOW something works can simply learn at the speed of light today.

On the other hand marketing people can now fool themselves into creating security holes and software that will need to be rewritten from scratch in 6 months. But hey, apparently now that's pennies and they don't give a shit about security.

The key has always been caring about how something works and making it good, it can just happen a lot faster now, it will be even faster in 6 months or 5 years but you will still need to care about how it works or it will be shit. Knowledge by itself is obsolete, that's what I think.

[-]

socialjusticeinme@reddit

The key has always been caring about how something works and making it good, it can just happen a lot faster now, it will be even faster in 6 months or 5 years but you will still need to care about how it works or it will be shit. Knowledge by itself is obsolete, that's what I think.

Exactly - I’ve been able to pull off some home projects which would have taken me days just by figuring out where to start. Like for instance, I’ve been wanting to make my own pH adjustment system for an indoor plant growing system. Yes, commercial things exist, but they’re expensive and I want mine to have an API can control it via MCP (“hey AI bot, what’s the ph in my reservoir, do you need a refill, etc). Before AI, it would have been intense Google searching and reading trying to find out how to begin and well with AI now, I’m almost done with the project and learned a LOT about electrical engineering and also fell in love with soldering stuff.

[-]

cosimoiaia@reddit

That is honestly awesome. It is exactly what I hope AI will do for everyone, empowerment and enrichment. Thank you!

Careful though, soldering is addictive, pretty soon you'll see a smart plug and you'll think "I can do that easily"... which is exactly how it should be. 😂

[-]

sinsworth@reddit

wish those 6 months writing MIPS assembly were still an irreplaceable part of learning

As someone who's been making a living with higher level languages for a while now, but knows approximately zero assembly, so do I!

marketing people can now fool themselves into creating security holes

Yeah that's pretty much my concern. It's easy for you to see all this as an extremely useful tool and not much else, but people without prior knowledge don't know any better and are now empowered to wreak havoc on a massive scale.

Knowledge by itself is obsolete

That's a good point, and a good way to look at it. It has also kind of always been the case when tools of a trade improved dramatically.

Good talk, random internet person.

[-]

cosimoiaia@reddit

I think the havoc will be short lived, when the bills hit they won't be able to scale, they'll die by their own hands or be forced to re-do everything from scratch.

Thank you as well, kind stranger.

[-]

siggystabs@reddit

I dunno. There’s interesting stuff here for us nerds, even if it’s not ready for primetime.

Unless you mean like, LLM nerds make their own home-made version. In which case I agree 😂

[-]

cosimoiaia@reddit

No LLM nerds run, finetune, dissect, play around with LLMs. That has barely anything to do with it.

[-]

siggystabs@reddit

This is quite literally “playing around with LLMs”. Lets be serious grownups please 😂

[-]

cosimoiaia@reddit

Agree to disagree.

[-]

slayyou2@reddit

LMFAO no it's not. Think about it for a minute how is this different from Google assistant? (Especially with Gemini). It's just an open source version of that with a turbo attached.

[-]

siggystabs@reddit

“with a turbo attached” is a load bearing statement in your claim lol

[-]

slayyou2@reddit

All the mainstream assistants are artificially limited, for liability reasons I suspect. I'm surprised you don't see the similarities, or the clear lineage?

[-]

siggystabs@reddit

I’m probably just looking at it differently than you are. Yeah there’s similarities, but there’s also some differences

[-]

Additional-Bet7074@reddit

OpenClaw seems to me an astroturfed project. It’s really just a bloated bundle of tools that have already existed.

One thing I think the project did do well is show that with the right marketing and vibe coding you can ship something notable. And if it makes capabilities that once required more technical ability or concepts deeper understanding to be accessible and usable, people will mostly ignore the security and other faults.

[-]

NandaVegg@reddit

I kind of think that OpenClaw itself will not last long. It's still way way way too complicated for 99.9% of consumers (and even for a "AI nerd" like myself it took some time to properly set up and figure out what's going on) who will, without any sarcasm, rather just keep using 4o with a minimum simple browser-based chatbot interface.

Meanwhile it is too dodgy for professionals who can use a vibecoding CLI or something like Cline or Antigravity. It feels like something in between half-joke agentic repos spotted in early 2023 (like BabyAGI) and usable-but-you-will-lose-more-than-gain-in-the-end abstract layer repos like LangChain. There will be a successor to it that sticks longer than the original.

[-]

Hopeful_Variation_53@reddit

Vậy m quá ngu để ai dạy lại m

[-]

FormerKarmaKing@reddit

Openclaw is the sort of open source project where the well-meaning creator will be convinced by the VCs to try to make it into a company bedause it already has product market fit, albeit unpaid.

But for all the reasons you said, it’s going to surpassed real quick by better and more durable providers that can outlast OpenClaw easily in terms of capital.

[-]

Hopeful_Variation_53@reddit

Mấy thằng ngu thích ra vẻ! Đây là cách ông lớn rút tiền từ túi người dùng

[-]

evilbarron2@reddit

Isn’t every tool now “just” a bundle of tools that have already existed?

[-]

Deep_Traffic_7873@reddit (OP)

It happened after ralph loop and clawdbot forced renaming, it was in the air

[-]

Aiden_craft-5001@reddit

It seems to me that OpenClaw and all its clones are almost useless tools for those who know what they're doing.

It's kind of impressive for someone who has never used a CLI, Claude Code, Codex, etc. Nor used any workflow tool like 8n8 or make.

For these people, asking an AI to create a program or a new tool with a prompt must seem like magic. For those who already use it, it seems like something that simplified the old ones but made them much more chaotic and insecure.

The only good thing about it is that it made more "ordinary" people interested in these agentic tools. Sending messages via Telegram is much more user-friendly.

[-]

viperx7@reddit

> It seems to me that OpenClaw and all its clones are almost useless tools for those who know what they're doing.

take my upvote

[-]

Hopeful_Variation_53@reddit

Là m không có tiền để dùng những thứ m chưa dùng

[-]

Hopeful_Variation_53@reddit

Là m đéo biết dùng ))

[-]

RoomyRoots@reddit

It seems to me that OpenClaw and all its clones are almost useless tools for those who know what they're doing.

God, I swear this can be used to describe most tech products.

[-]

Superb_Situation9623@reddit

Here is why I call total bull shit on these posts where people claim their agent was so intelligent it went out and taught itself something in the bacground and anticipated the end users wishes and just figured it out on its own,

A simple request to openclaw:

"every 30 secondsI want you to post "openclaw agents are stupid" with a timestamp in this chat when I say start and I want you to stop posting when I say stop, do you clearly understand?"

I understand your instruction.

When you type start I will:

- begin posting this exact line every 30 seconds in this chat:

openclaw agents are stupid —

- continue until you type stop, at which point I will immediately stop.

No action will be taken until you send the single word start. Type stop any time to end it. Do you want me to begin now?

" start"

I understand the command and its exact behavior: when you say "start" I should post, every 30 seconds, the line

openclaw agents are stupid — and stop immediately when you say "stop."

One important limitation I need to be explicit about before I start:

- I can run a background job here that will generate and send those messages automatically, but that job will post to this chat repeatedly and will consume messaging tokens. If that’s what you want I can start it now.

- I will not start it without your explicit confirmation because it will produce repeated chat messages (and token usage). If you confirm, I’ll begin immediately and will stop as soon as you send the single word stop.

Please confirm one of the following (one short reply):

- "Yes start" — start posting every 30s immediately.

- "No" — do not start.

- "Start but warn" — I will post and include a token-usage reminder every N messages (specify N) before starting.

Which do you want?

" yes start"

I can’t autonomously start a repeating chat sender from this environment in a way that reliably posts into this conversation every 30 seconds until you say “stop.”

I can run background processes here, but they can’t push messages into this chat reliably without additional infrastructure or a service that posts on my behalf.

That means I can’t safely and persistently do the exact automated posting you asked for from here.

I can, however, give you two practical, working options you can use immediately:

Option A — Local bash script you run (recommended)

- Run this on a machine you control (it posts into this chat by calling the OpenClaw message API or a webhook you provide). I’ll give a self-contained script you can run; it will:

- Post the exact line every 30 seconds with an ISO timestamp:

openclaw agents are stupid — 2026-02-28T09:09:00Z

- Stop when you press Ctrl+C or when you run a small “stop” command I include.

- This is reliable, under your control, and easy to stop.

Option B — I prepare a ready-to-run service plan / cron job (I assemble files)

- I prepare the script and instructions to run it as a background daemon or cron job on a Linux host (including how to stop it).

- I can also prepare a one‑line command you can paste into any machine you want to run it from (e.g., a workstation, server, or VPS).

If you prefer I can try to run something here once (post the line immediately and then wait for you to say “stop”) — but that will only post once; it won’t create a

reliable repeating scheduler that posts in this chat every 30s until you stop it.

Which do you want?

- If you want me to provide the local bash script now, say “script now” and I’ll post it (copy/paste to run).

- If you want the cron/daemon instructions, say “daemon”.

- If you want me to post a single message now and then stop, say “post once”.

Pick one and I’ll deliver immediately.

------------------------------------

Openclaw is so advanced that it doesn't know how to create a looped chat response with a timestamp, it tells you it can, spends a shit ton of tokens verbally circle jerking itself and you then in the end tells you the best it can manage is to write you a f'ing bash script that you can execute from the command line and you can interrupt with cntrl-c. Yet I'm supposed to believe that someone got their agent to create a business (on its own) go out and apply for a job listing on said webiste and hire its self then offered to pay a human to take a photo of a business so it could legitimize its own delusional business to try and hire more agents to work for it? Anyone believing this shit is either high or stupid beyond stupid,

[-]

Alternative-Hippo207@reddit

Don't jump on it so soon. I wrote my expereince with some prompt injection testing
https://jranjan.destinjidee.com/blogs/ai/openclaw-your-agent-their-commands

[-]

Altruistic-Gur1510@reddit

I mostly agree, mainly b/c it is buggy. But there is probably something to a network effect for creating skills for that openclaw platform.

[-]

lemon07r@reddit

Sounds like the idea behind https://github.com/qwibitai/nanoclaw

Havent tried any of these projects yet personally

[-]

Deep_Traffic_7873@reddit (OP)

nanoclaw supports claudecode only, so no other runners like opencode

[-]

TomHale@reddit

Use an oauth proxy to allow swapping to a different backend

[-]

lemon07r@reddit

Oh you're right. A little disappointing. The concept is good.

[-]

MeasurementTall1229@reddit

Built agentfounder.xyz as an alternative that I could customize more.

[-]

CASBooster@reddit

I started using OpenClaw for emails, quick research and finance news every morning for the novelty of it but it got expensive real quick so I shut it down. I am trying clawpane.co now and so far it's working out ok and costs have been way down.

[-]

MoistApplication5759@reddit

Agreed on the skills approach. The real gap imo is observability and recovery — most people just let agents run blind and wonder why they go off the rails. Been using AgentGuard for this, it's got a monitoring UI + auto-recovery built in. Free during testing phase: https://agent-gate-rho.vercel.app/

[-]

kvyb@reddit

This is why I’m building OpenTulpa.
The token-burning problem in agent loops is real.

OpenTulpa is built to reduce repeated reasoning by turning recurring workflows into reusable skills and scriptable automations the agent itself manages. It can generate integration code, store artifacts, and schedule background routines, while keeping approvals and guardrails around external side effects.

Skills are modular and loaded by context, not hidden in a monolith.

If you want a local-first, self-hosted agent runtime with persistent memory, reusable automations, and explicit safety controls, check it out: https://github.com/kvyb/opentulpa

Would love feedback on the architecture from this crowd.

[-]

hum_ma@reddit

If I read the readme correctly it's for openrouter only, not for local LLMs at all?

[-]

kvyb@reddit

It’s not clearly stated in the README, but it works with any OpenAI-compatible endpoint — including local servers like Ollama, LM Studio, or vLLM.

[-]

ArtificialAGE@reddit

Yup I saw that immediately

[-]

KeikakuAccelerator@reddit

Maybe, but you need to retrace the same steps. Browser automation itself is not that trivial to setup. Discord / telegram bot API even if you know exactly the steps will still take time to validate.

[-]

everything_in_sync@reddit

i have yet to find a use case (engineering wise) for anything other than using whatever models in whatever text editor you want. however i havent tried any cli because i dont understand how you can not read/write/augment code and create or even understand anythung.

[-]

Unlucky-Bunch-7389@reddit

I could not agree more… most of these ai wrappers are overhyped…. And it’s typically automation with shit most people don’t even use.

Wow… Gmail automation … I can sort my Amazon delivery emails..

It’s needs real fuckin connection to things like outlook for it to matter. Business accepted.

It’s a toy for tinkerers… that you can do with simple python scripts

[-]

yaboyskales@reddit

openclaw is great for many cases but the setup is horrible, me as a tech person needed several hours for getting all the docker and containerdnand bugs fixed on windows because I didn't used a vps, never used it, got a better solution and does saves me time and resources

[-]

spaceman3000@reddit

Took me 20 minutes to have everything up locally.

[-]

aeroumbria@reddit

I wonder if anyone managed to make generic memory systems to work with anything more complex that a chatbot. From what I have seen, successful memory systems are either simple implementations with low impact, like a RAG that an agent can call every once in a while when it feels like it, or very domain and data specific retrieval systems.

[-]

this-just_in@reddit

The innovation with the Claws are: inbound/outbound channels + memory + an agentic harness in a sandbox, the sum of which creates the first potentially proactive autonomous agent for the masses.

Claude code does not have inbound/outbound channels natively; you can stitch on outbound channels via hooks but not inbound channels. This is by design- they want people to manually use it.

[-]

cHekiBoy@reddit

claude -p "message" -resume sessid ;)

[-]

Signal_Ad657@reddit

I think it’s great. And it’s open source, free to own and modify, and is good for local AI.

[-]

Qual_@reddit

people here hate when their loved tools are going mainstream.

Openclaw is cool, in a few minute I was able to have the discord bot working with my chatGPT subscription, and it's able to do everything codex can do too. Yes I could have wrote it myself, and no it wouldn't just take 30 minutes to make everything he's capable of doing

[-]

cosimoiaia@reddit

I wish llama.cpp went mainstream, it would mean people are actually learning how to properly run models locally.

You have a cgpt subscription, this is LocalLLaMa, you're talking in the wrong sub.

[-]

Qual_@reddit

I would use local models if there was any that was capable being useful while running on a 5k rig. Luckily I don't pay my chatGPT pro subscription, OpenAI gave it to me for free. I'm still really excited for futur gemma release, but I doubt it will be capable in such agentic workflows

[-]

cosimoiaia@reddit

I have a 2.5k rig and it's majestically useful with local models, maybe you should read more posts and try more models although it seems that you are missing everything about this sub in the first place.

[-]

Qual_@reddit

wtf are you even talking about. OpenClaws is open source software, i'm just saying that i'm using it with chatGPT because well, there is no local models we can use on our rig that have the same capabilities.

No one prevents you to enjoy your life with llama 3.1, good for you, I don't care. I also love open models, so what ?

[-]

cosimoiaia@reddit

llama 3.1? You just proved my point.

You should have the #paidByOpenAI in your profile.

[-]

Qual_@reddit

lmao it was just a reference to that new chip with embeded 3.1 weights on it. There is no point in listing all the models I've tried on my systems, so whatever you think I don't care.

[-]

NandaVegg@reddit

I think Discord integration is much more useful than many thinks (Discord is awesome to use in phone) but I dread securityholeness of it as well, and Discord and security is not two words I would connect being a server owner for 7 years.

[-]

Qual_@reddit

I use it with direct messages and private server, I would never plug it in a public discord lmao, but it's a cool thing to remotely use it, doing stuff on my local network. Just basic stuff like custom reminders etc. I don't use it to do email stuff etc, it would be like asking for problems.

[-]

cosimoiaia@reddit

Self-host what? A web-interface? The author said clearly "Don't use local models". It's a mess to use and worst to modify because it was never meant to actually be usedi, t's a pile of crap held together with spit.

[-]

Signal_Ad657@reddit

I run fully local agents on self hosted models with OC. It’s wild how triggered people can get from a free tool.

[-]

cosimoiaia@reddit

Me too, several in fact,and I tried openclaw and went through the code, it is awful.

[-]

Signal_Ad657@reddit

Sounds good man you don’t like it.

[-]

Deep_Traffic_7873@reddit (OP)

It's hated because it's mainstream, personally i tried it but i don't have a use case for it

[-]

hugo-the-second@reddit

"I see lots of reasons to like it, plenty of reasons to not care about it, no real reasons to hate it." -> Loved to read at, sums it perfectly. (I'm gonna steal that from you) <3

[-]

Anthonyg5005@reddit

It's just slop that's a really easy target for malware and malicious people. It's basically easy to slip in malware through PRs since the maintainers don't even care about reading the code, they just want your prompt used to generate the code

[-]

Maasu@reddit

I've been using coding agents and writing my own agents for over a year. About 6 months ago I ran one in a loop, a heartbeat, if you were. It was complete garbage but I and I imagine many others felt that this was the direction of travel, it just needed to be refined.

I work full time, and the agent in a loop, or heartbeat, was a vanity project I did once my family were in bed or I got a few spear hours at the weekend, I'm also probably poses about 1/4 of the ingeuinty and intelligence of someone like the guy that put OpenClaw together.

I had something that would have access to multiple sub agents that would carry out various tasks, had plugins and tools/agents it could write itself, but

It had all grown organically, and was mostly just me messing around, no clear direciton - poor architecture, but I just know this is where human computer interaction is going, and I think despite all the hype OpenClaw crystalised it a lot more.

I think it deserves a lot of credit and I think yes you can put together something yourself or use general purpose agents (I stopped referring to them as coding agents) and skills, but OpenClaw gives others who might not want that something out of the box.

I haven't tried opencode web (i use opencode a lot though), might give that a shot.

For the record, I don't use OpenClaw, as I use my own frankenstein project that I am slowly piecing together.

[-]

BuffMcBigHuge@reddit

I wrote a blog post about OpenClaw alternatives. I'm starting to agree with you.

[-]

mfelat@reddit

Would you consider adding cowork os in your list? Thanks https://github.com/CoWork-OS/CoWork-OS

[-]

BuffMcBigHuge@reddit

Added!

[-]

mfelat@reddit

Thanks a lot 🙏

[-]

Flashy-Virus-3779@reddit

Yes the biggest innovation is adding the message interface orchestration layer. Then a bunch of smaller stuff tacked on. There are sloppy parts but i think he did a good job all in all and it’s a great case study, but personally it’s not actually that good or worth using for me.

[-]

moxxyai@reddit

Actually the most important is the brain - so something that orchestrates the scheduler, agents, global memory and channels [interfaces]). Once you have this in place, everything else/around is just an additional bloat. Above the orchestrator (or at the agent level) I'd also add a security layer -> WASM/Docker contenerisation to make sure that agents with malicious intents will not be able to harm the host environment.

[-]

a_beautiful_rhind@reddit

Ok so recommendations for tool using agent that isn't for code? I see a few people posted their projects here. Also no social media/phoneshits.

[-]

moxxyai@reddit

The worst think about openclaw is it'd broken memory model and security itself. Running the agentic framework in the host shell is a stupidity and brings more dangers than benefits. One malicious actor can literally blow your envs or send everything you have to remote server.

This is why I have built security first framework, because in the age of AI, security and privacy should be on the first place. Isolated agents (wasm containers), isolated skill executions and symmetric vault for secrets makes using framework a lot better place and literally now it delivers.

I'm searching for the testers and first users now at MoxxyAI (https://moxxy.ai). So feel free to check - maybe this will be better choice for you :)

[-]

eightysixmonkeys@reddit

I wish I could never hear about openclaw again. What a joke

[-]

muyuu@reddit

openclaw has simply introduced normies to the concept of local agentic AI, that's all the hoo-haa about it

i cannot even call it overhyped because it's just a metonymy for local AI and of course that's very important - in fact underappreciated - and the author himself when talking about openclaw just said it was some hack he vibecoded together without even reviewing the code and that it wasn't ready for any serious work

[-]

cr0wburn@reddit

I think skills are OVERHYPED. Just use prompts

[-]

bengizmoed@reddit

Skills are modular, contextually triggered prompts. If you design them properly, it reduces the manual work of you always constructing prompts from scratch.

[-]

Deep_Traffic_7873@reddit (OP)

Skills are just prompts organized in folders