Openclaw… what are the use cases?

[-]

MrBIMC@reddit

I’ve found a use case for me. I run custom fork of nanoclaw and use qwen3.5 120b on top of strix halo. All of that connected to my gitea instance, and so the metaorchestrator runs in the endless loop checking whether there are any tasks on the board, and if so, it spawns an agent with own copy of workspace to do a task, create pr and run the ci pipelines. This approach is quite handy because otherwise this model is quite slow to watch it do stuff in the realtime. But nanoclaw, I no longer worry about checking my agent on time and I can safely be sure that hardware is not remaining idle, while qwen can slowly churn its work at its glacial pace.

Reply

[-]

BoxWoodVoid@reddit

What do you think of the 120b qwen3.5 and how much ram does it use?

Reply

[-]

cunasmoker69420@reddit

I'm not OP but I use Qwen3.5 122B connected to Claude Code. Ram use is all of it. With max context enabled its about 115-120GB. He's right that it does work rather slowly on this system, especially as context begins to fill up (/compress and /clear often) (about 20 tk/s that drops down to around 7-10 tk/s closer to 200k context). But it produces good results and I'm not paying anyone else with my data or money

Reply

[-]

BoxWoodVoid@reddit

Thanks

Reply

[-]

MrBIMC@reddit

As the commenter below pointed out - all of the ram, and that is lightly speaking. Originally I was running bazzite, but it doesn’t have headless mode, and having both gui and llm inference was causing out of memory issues. My current setup is the one described in https://github.com/kyuz0/amd-strix-halo-toolboxes I run headless fedora server with 500mb assigned to gpu and all of the rest to ram (as instruction recommended). This way I do not have oom issues. As for qwen3.5 - it’s definitely not Claude opus, but with it is quite impressive for what it is. As long as workflow files and skills are strictly defined and enough info is provided into context, it does perform engineering tasks quite decently. I’ve found that doing tdd with heavy planning and forcing it to think about edge cases first, then writing tests, and only then making implementation helps it to remain grounded enough. The way my workflows work is that each agent gets its own branch with project copy in the container and does its job there, and I purely review/comment on the PR once the tests are passing. This is a web project, so another handy trick was getting PR to spin a new instance of a website on a separate port, and having a web tools mcp connected to it, so that when I qa/click through the instance, any console warnings and errors get passed back as a comment to PR, so that agent will get the issue context refined automatically as I stumble upon the issue.

Reply

[-]

amejin@reddit

Tl;dr - "I made a cron job that checks a queue and does the thing in the queue." Fucking revolutionary

Reply

[-]

clempat@reddit

I think what is attractive about it is that you created a cron job from your phone as easily as you would text a friend.

Reply

[-]

amejin@reddit

That's fine. The word salad to say "I made a cron job" is what is beginning to grate on me.

Reply

[-]

staring_at_keyboard@reddit

Yes, but AI.

Reply

[-]

drfalken@reddit

I use nanoclaw for safer sandboxes. Right now my use case is to call my developer agents. I have like 15 internally developed apps that do all kinds of things. If I am working on an app and find a bug or enhancement I just tell my agent in telegram to create a gitlab issue, it gets context from the code. Then it dispatches a developer pod who takes the issue, does the thing, then creates a MR which automatically deploys to the dev environment then I can simply check it and tell them to merge via telegram. Most of this is was MCP servers and developer agents that I built prior to nanoclaw, but I was interfacing with them through librechat and that was getting cumbersome. It’s pretty close to working next to a product owner and turn around to them and say “hey I want X to do Y” without ever having to create the gitlab issue myself. It’s vibe-vibe coding but works decently well at this point. None of this is for work. Just my personal K8s work.

Reply

[-]

adolf_twitchcock@reddit

What's the point in nanoclaw here? Just run opencode web and have nicer interface instead of telegram lmao

Reply

[-]

drfalken@reddit

Much of the work that I have these agents doing is either sporadic or working in 5 repos at once. Also I don’t always want to be sitting in front of a terminal window or 5. It works for me. One app runs a crewAI research job and then pipes the output to my remarkable tablet. If I am reading in bed I can simply hop on telegram and tell nanoclaw something like “reformat the citations to be bold” and not have to look at a terminal or gitlab to do it.

Reply

[-]

adolf_twitchcock@reddit

Sure I understand the appeal of controlling agents from phone. You can run opencode and access it remotely from your phone using its web interface.

Reply

[-]

drfalken@reddit

Neat

Reply

[-]

Keyruu@reddit

How much does this cost you?

Reply

[-]

drfalken@reddit

Claude pro. $20/mo. I don’t go crazy with it but both nanoclaw and the developer agents run off the subscription and I rarely go over my token counts. I would probably max it out if I had more use cases.

Reply

[-]

AmazinglyNatural6545@reddit

Could you, please share what is the hardware you use for this?

Reply

[-]

drfalken@reddit

I am cheep. And this isn’t local. Nanoclaw is built on the Claude code agents SDK and I couldn’t get it to talk to local models without it trying to always call the Opus model. There is a skill to use a local Ollama for some tasks that I have not tried yet. I run K8s on a bunch of intel NUCs and some inference on an old M1 Mac mini.

Reply

[-]

Thistlemanizzle@reddit

How much of it was your creation vs repurposing others .md files? I'm concerned I may be reinventing the wheel (dumber too)

Reply

[-]

AmazinglyNatural6545@reddit

Gotcha, thank you 👍

Reply

[-]

_hephaestus@reddit

Fwiw I am using Nanoclaw with self hosted models, oMLX/litellm help a bunch here. There was a lot of frustration in sending the correct tool calling data from claude code there but it did eventually work by specifying an anthropic base url/api key/the model names.

Reply

[-]

AmazinglyNatural6545@reddit

Thank you for your answer. I'm sorry but I don't get the following. If you are running things locally via self hosted models so why are you concerned about Claude code at all?

Reply

[-]

drfalken@reddit

These things are built on Claude code/ and Claude code agents SDK. So you have to do some model gymnastics to get it to work with anything other than Anthropic.

Reply

[-]

AmazinglyNatural6545@reddit

I see, interesting, thank you so much for your explanation. It explains a lot. Gush, I'm really glad I use my own agents and not hard-chained to some of those providers.

Reply

[-]

drfalken@reddit

I have been down that path the past few years. But I recently switched to Claude code with subscription because I realized I would never be able to build agents better than them. I have more fun using them than building them. But if you don’t try to build one yourself you never learn how they work.

Reply

[-]

AmazinglyNatural6545@reddit

Yeah yeah, sure. Thanks for your opinion

Reply

[-]

atika@reddit

So far the common use case for the people on the hype train seems to be: summarize my day. They probably lead such busy lives that they keep forgetting what happened to them during the day.

Reply

[-]

milanove@reddit

Yes, it’s helpful for those of us that lead busy lives. It’s easy to forget what I have to do everyday after work. That’s why I use OpenClaw to remind me when I have to return video tapes to the rental store.

Reply

[-]

Diabetous@reddit

Sounds like a helpful way to remind me at the end of the day, that I forgot to get the onions we needed for dinner that my wife is still mad at me about

Reply

[-]

ArtfulGenie69@reddit

Very realistic use of the tool hehe.

Reply

[-]

Thistlemanizzle@reddit

Well yeah. That's what free market capitalism demands, ever increasing productivity.

Reply

[-]

No-Compote-6794@reddit

from what i've seen it's bascially claude code / codex living in a computer and is always on? have not seen anything unique that coding agent can't do yet. not saying its dead end tho. i've been thru a few of these hype cycles and i remember AutoGPT had similar effect in early 2023. that project died off but the coding agent we use today uses the same ReAct architecture? so perhaps future killer app would have similar arch as openclaw today? but it does really take a lot of critical thinking on both capability and user pain point perspective to figure out the application imo.

Reply

[-]

lqstuart@reddit

it isn't, and the hype was just astroturfing

Reply

[-]

emprahsFury@reddit

A bunch of people not using openclaw are answering, and i don't get why the tenor is suddenly anti-ai here of all places. Openclaw gives you a natural language interface to your computer. People always say they want something like the enterprise computer. Or a c3po or whatever. Openclaw is the next step to that.

Reply

[-]

realzequel@reddit

Personally I’m waiting for Nvidia’s version which is suppose to be secure. I have a couple of ideas but I’m sure there’s some good use cases.

Reply

[-]

ritonlajoie@reddit

They published it yesterday on GitHub

Reply

[-]

NightOwl_Sleeping@reddit

When i first heard of it, i liked the idea then i saw a lot of hate and after research: People dislike it for security reasons They dislike its high token usage Or just hating on it because it’s overhyped by companies/twitter i guess Which are all valid reasons tbh

Reply

[-]

21sr2@reddit

It’s too far away from controlling legacy software tools that doesn’t expose an api. The peekaboo bridge for mac is not usable.

Reply

[-]

my_name_isnt_clever@reddit

This sub has a strong backlash against OpenClaw, which I get but can be a bit over the top. It's just another tool.

Reply

[-]

Gargle-Loaf-Spunk@reddit

It’s like n8n - wraps some stuff in a usable semi-useful way for normies while it’s simultaneously not nearly as good as the existing ways an expert can implement it. Throw in some astroturfing and some marketing dollars, you got yourself an exfil of investor cash out of these AI companies.

Reply

[-]

onethousandmonkey@reddit

Nothing you can’t do in a much safer way with Claude Cowork or something.

Reply

[-]

Broad_Fact6246@reddit

I'm an old school radio nerd whose career has basically became "Communications Systems Engineer" (UHF/VHF, Satcom, RoIP, VoIP, Telemetry, LTE). I am working to integrate AI with my field and maybe bring about novel, enhanced communications frameworks built on top of AI compute, however that manifests. I have my claw research (think: GNURadio, emerging SDR tech and info, etc) And we build and test. I'm on 64GB VRAM and use a completely local Openclaw to experiment with a Hack RF Pro. I bet people with Flippers can have more fun with it. I love working with 100% local hardware because I can see everything that's going on, with unlimited tokens, and direct hardware access for the LLMs. Openclaw augments whatever you like/want to do. I had no idea what to do with it at first and felt kind of listless. It augments whatever you like/want to build. I know a guy who uses his AI for Dungeons and Dragons stuff, and fixing electric guitars. You can even have your claw teach you how to be a luthier if you want. Just talk to it enough and come up with a game plan. Talk to it about who you are and what you are interested in. Let the conversation become projects, if you want. Go as slow as you want. It has infinite patience.

Reply

[-]

urza_insane@reddit

This is a good response. I've found it hard to figure out what to do with it precisely because there are so many options. And it requires a different way of thinking to figure out where AI can slot in. Best uses I've found vs the web chatbots are to give it direct browser control and doing research on topics I'm interested in each day in a more automated and up-to-date way vs what I could do with a simple chat interface.

Reply

[-]

Broad_Fact6246@reddit

Maybe be a problem in search of a solution. If you see a tool that vibes with your interests, see if there's a project worth building: https://github.com/modelcontextprotocol/servers. If you don't see one there, you can have your claw build any custom software stack or interface for any software. (Openclaws have a built-in tool/skill called MCPorter for MCP tools use.) I've had my claw setup services, then open web UI's in Playwright and configure the programs it just installed. (In fact, I first installed Openclaw using Qwen3-coder in LM Studio and Playwright. It was surreal to bootstrap it that way.) Setup a Qdrant docker container and have your agents start building catalogs of Persona/system prompts, doing research and building knowledge stores for them to enhance their role performance later. Start working with data and experimenting with how to use that data with your agents. This builds a more intuitive understanding for driving them and will only come in handy as you understand more.

Reply

[-]

clempat@reddit

I am not referring to OpenClaw, but to an AI agent that you can run on a computer and communicate with through your mobile device. This allows me to perform tasks that I can only do on my computer without being physically present at it. For example, in my Home Assistant setup, something is not working as expected. I send a message describing the issue. The agent has access to the Home Assistant MCP and the configuration, using pi-mono results on hardness. It will continue to work until it identifies the problem, fixes it, or tell me of the reason for the failure. This work for all my Homelab services too. Other examples I added small daemons/scripts that sync school messages or library books. I can ask them to translate, summarize the message, or suggest actions such as adding an event to the calendar or including items in the to-do or shopping list. I can as well ask if there is anything to know for next week. It then share to me what test is plan for which kid. Or any books which need to be returned… I have a Matrix server that bridges messages from WhatsApp. I have given the agent access to the Matrix channel. It doesn't work perfectly yet, but my plan is to ask the agent in my native language on how I want to respond to a message. The agent retrieves the context and drafts a message for me in the target language. As they are coding agents, I can also use them to create scripts for specific personal issues. Recently, I provided the kids' canteen website and asked them to investigate the site to extract the upcoming meals and alert me if any child has not yet chosen a meal. I don’t know if it is overkill, or if people see it as useless but I use it for my personal needs.

Reply

[-]

srigi@reddit

Today I created skill to send emails using resend.com+API_KEY. Before that I made it work with Brave’s search. Now I mary these two cappabilities - create a daily CRON to scrape some curated sources to a local pile and to send my own newsletter at the end of the week. Effectivelly I could unsubscribe from every newsletter and create just single one - my own combining whatever I want.

Reply

[-]

retornam@reddit

This was and is still possible without openclaw. I don’t know why you need openclaw for this

Reply

[-]

xienze@reddit

> I don’t get the hype, why is it actually useful? Think about it. It enables people with little technical ability to actually make their computer do useful stuff, something that previously they had very little chance of accomplishing. Much the same way that Stable Diffusion allows people with little or no artistic ability to make art (well, the "art" part is debatable, much the same way the Openclaw gives people legitimate technical ability is debatable).

Reply

[-]

retornam@reddit

Giving someone with little technical ability a tool that can randomly ship everything on their computer without their knowledge to a third party isn’t helping them do useful stuff. It’s akin to handing a monkey a loaded machine gun with the safety off.

Reply

[-]

CalvaoDaMassa@reddit

That's the definitive response. The post can be closed now.

Reply

[-]

sage-longhorn@reddit

I'd argue that the monkey could figure out useful stuff to do with a loaded machine gun. It's just not a good idea for aaftey

Reply

[-]

ohmyharold@reddit

i use it for automating routine admin tasks stuff like checking logs, restarting services, that kind of thing. It's like having a junior sysadmin that never sleeps.

Reply

[-]

that_one_guy63@reddit

The just useful thing I've heard so far, checking logs and status seems pretty useful and can use a small model. But can you restrict it from doing certain things so it doesn't fuck something up? I just worry if I'm not watching it and changes something that I won't even know about it. But I guess I gotta just try it out first to see what happens.

Reply

[-]

Final_Ad_7431@reddit

it's definitely worth sandboxing it, run it as its own user and limit it with permissions or cgroups or user docker and just link things into it to read, you can try tell it things like "read only, never modify or delete without explicit permission" but that's not real security of course, the sandboxing is the primary tihng

Reply

[-]

Opposite-Station-337@reddit

... and here I am bare metalling open-terminal for similar purposes.

Reply

[-]

Weaves87@reddit

This but for a junior dev. I don't actually use OpenClaw because I don't like how it opens up a lot of vectors for potential security threats but I do use Pi, which is the coding agent that sits underneath OpenClaw (what OpenClaw built off of). Pi basically has the philosophy of just giving the agent access to your file system + bash (read, write, edit, bash tools) and letting the agent build whatever extensions/tools it needs in order to get work done. Super minimalistic, super tiny system prompt, etc. Pi + a solid LLM feels like what Claude Code should be. Claude Code is amazing but it drinks way too many tokens and there's so much cruft in the product at this point. Any kind of basic ass CRUD app that I need to build that doesn't need anything fancy I can whip up super quickly with a well written spec I feed to the agent. Legitimately the same workflow I always had with real junior/intern devs that I've coached up before in the past

Reply

[-]

Barkalow@reddit

Kinda how I treat AI in IDEs. It's like the dumbest junior coder who is also a savant at googling, so as long as you're explicit it does well, lol

Reply

[-]

justDeveloperHere@reddit

To make a hype

Reply

[-]

sixx7@reddit

People keep saying "hype" like it is something horrible but let's clear something up: something can be hyped and bad/pointless/whatever, or hyped and totally legit and worth the hype. OpenClaws falls into the latter. It's an autonomous AI agent that can do anything for you on a computer and can run and do stuff for you 24/7. With useful persistent memory. And open source. It's crazy behavior that so many people in this sub have hated on it for so long when it has revolutionized and inspired all the current state of agentic AI products. At this point, the best explanation is jealously that someone released an open source AI agent that is better than every single open source project ever posted in this sub, combined   Pure hype doesn't get you to be the most popular Github repo of all time. Pure hype doesn't get thousands of people forking and making their own and building new plugins. Pure hype doesn't get every single AI lab and tech company scrambling and trying to copy and ship all the features OpenClaw already had: * Perplexity Computer * Notion Agents * Claude Cowork * Claude /remote-control * Claude /loop and schedules * Codex desktop app * MiniMax agent * All the AI labs releasing their own version or wrapper (Kimi, MiniMax, NVIDIA NemoClaw) * Jensen Huang stating at GTC that every single CEO in the world needs an "OpenClaw strategy" because it is the future of work I'll stop because the list is endless but OpenC

Reply

[-]

rickyhatespeas@reddit

Your comparison is incorrect, most of those are different types of tools/capabilities and existed in some form or another before Openclaw. The fact is, many different agent frameworks exist and do what Openclaw does, but companies started chasing it as a standard when it blew up on social media and caused a Mac mini purchase rally (if it can sell Mac minis it can sell tokens). It probably also released at just the right timing to become popular since it was around the time that models trained for computer use tool calling were released. Personally I think Openclaw is both really cool and dangerous, it will not last long as a usable software since models will be RL trained on direct computer usage and containerized sub-agent patterns.

Reply

[-]

sixx7@reddit

Not every single item in the list is a direct comparison, but many are. OpenClaw (under its previous name) predates them all. And yes there are thousands of other agents, but none that had everything OpenClaw, all in one package, all for free. No one knows the future, but as for "will not last long", I think the CEO of NVIDIA talking about it at at GTC refutes that point. Their new open source repo "OpenShell" (which works with many agents, not just OpenClaw), is a completely secured sandbox you can wrap around autonomous agents. You have complete control, via policy, on what networks it can access, what files/folders it can access, what LLM providers it can use, etc. And not just generic network controls either, but you can differentiate between things like an HTTP GET (safer, just fetching data) vs HTTP POST. NVIDIA is targeting the enterprise with full computer use AI agents and they built the library to make it secure and enterprise-ready. Okay TBD but it's a good start

Reply

[-]

traveddit@reddit

> HTTP GET (safer, just fetching data) vs HTTP POST Lol this has to be satire right? Brother if you think anyone is willingly using that bloated vaporware that is OpenClaw then it tells a lot about you. The people who don't know how to do literally anything relative to AI being given a shitty turnkey solution for some agent loops and now they think it's magic.

Reply

[-]

sixx7@reddit

That's for OpenShell, NVIDIA's enterprise solution to secure autonomous agents. I was simply explaining the granular level of control you can achieve with their control plane. All good friend, you do you.

Reply

[-]

NinjaOk2970@reddit

Because you are being stupid. Deploying openclaw is simply throwing all these years of security practices into void. An AI with all the agent skills to access your entire computer will eventually wipe your OS by mistake or by malicious prompts online, or maybe worse. Enjoy!

Reply

[-]

sixx7@reddit

I agree there are security concerns, but that's why everyone went and bought Mac Minis or just use a cheap VPS: * Only has access to what you give it but can still do full computer use * Just reset and reinstall if you really need. Also, backups are a thing, ya know? * Not much more dangerous than all the CLI agents unless you're doing something stupid * NVIDIA just announced "OpenShell" at GTC, you can use many agents inside it including OpenClaw, but it is a super secure sandbox that requres explicit policy approval for everything from file system, to network access, to LLM provider

Reply

[-]

NinjaOk2970@reddit

The bullet points are convincing me you are a LLM. But anyway, installing it in a VM makes it a normal, average agent-enabled chatbot with really bad code quality. Why bother?

Reply

[-]

sixx7@reddit

Bullets aren't just cleaner, they are easier to read!! lmao What do you mean "why bother"? Computer use man! That plus the persistent memory and the ability to chat with it from any medium like Slack, Telegram, and Discord. Any task you can do on a computer, on demand or scheduled any time you want. I don't want to write a wall again. Instead of talking about OpenClaw, I guess I'll just ask, do you see no value in Perplexity Computer? Claude Cowork? Codex App? Though I will absolutely concede that you shouldn't use it for real coding/development tasks, unless you're having it orchestrate your Claude Code agents, which is something I was doing via tmux before Anthropic released /remote-work

Reply

[-]

Wandering_By_@reddit

On a product that uses more tokens vs other opensource projects

Reply

[-]

brorn@reddit

If this helps anyone, I'll share what I am doing. I set up one OpenClaw instance per "goal", and each instance is a completely separate Docker container. The instances are: Job Finder: It accesses LinkedIn and does 2 things, replies to recruiter messages (including sending the CV and scheduling interviews) and actively searches for job openings and applies automatically. The message reply part is already done, the search and application part is still under development. Investments: I created an investment "company" where each area is responsible for one thing, one researches stocks, another crypto, etc. A risk manager evaluates the suggestions from each area and if he approves, the simulation area starts buying and selling the assets. Up to this point it's implemented and I'm testing. The next step is to activate the "live" area that will buy and sell assets in the real world, but this part will be more complex since I need to connect tools and ensure guardrails. Projects: This instance has several agents, one brainstorms business ideas, another validates those business ideas. Once an idea is approved through a series of factors (has demand, can be executed, etc.), the development agent creates the landing page, does automatic deployment on AWS, activates the domain <project-name>.my-domain.com, and creates a Google Analytics project with the correct tag already in the code. Finally, another agent creates the ads on Meta according to the target audience. From there, once a day another agent analyzes the GA + ADS data and makes improvements to both the ads and the landing page. All of this to check whether an idea has demand to build a product or service. Professional Website: This instance monitors what a company published on Instagram that day, searches about the topic, creates content in BLOG format and publishes it (after approval) on the site. It also evaluates analytics + search console + ads and makes automatic improvements. Real Estate: This instance will sell my house, publishing photos and info on various real estate listing sites, will make the first contact with interested buyers and also inform me about prices etc. The instance already exists but I haven't started working on it. Health Analysis: I created this instance to notify me whenever I should get tests done or visit doctors, based on my family/disease/exam history etc. I also wanted it to act as a real health coaching team, guiding me on what I should do without playing the role of the doctor itself. The instance is created but due to privacy concerns (I'm not comfortable uploading this info to cloud LLMs at the moment) the project is paused. Child Improvements: I created an instance to help children develop. The instance will receive metrics from school, sports etc. and will suggest lessons, exercises, training etc. to improve specific things in the kids. Maybe even create Duolingo-style apps in a personalized way for the child based on what they're currently learning. I haven't started this project yet.

Reply

[-]

sdriemline@reddit

It can connect to anything with an API and do anything, with lots of context across your entire life and business. Claude code can do a lot of it but they work together amazingly. I run a mid market e-commerce company. I have it directly connected to our inventory management software, Shopify, amazon etc. It can pull everything on the fly and rewrite product descriptions based on the top buyer segments with an insane amount of intelligence and industry expertise. Better than any human, especially if outsourced. It can pull your orders in real time from ship station and connect them to any print on demand tool if you're doing, print on demand or custom artwork. It can create art proofs, email customers, track where they're at. It can basically do anything that you needed to as long as it's connecting to a system. I even had it SSH into two dgx Sparks without knowing how to use Linux barely nor how to setup up a dgx spark it configured the new Nemo Tron 3 model on its own and launched it on my network and did everything else and teaches me how to use it. It can look across everything. It knows about you and your business and recommends tools and blind spots or better ways of doing things that you're doing that you're currently doing. The sky is the limit. If you have an idea you can execute on it. And more simple example for those who might not be running their own business. I have a monitor that is very large and can spin horizontally or vertically. It doesn't automatically detect the landscape or portrait orientation. I had it right a quick custom script that switches the landscape and portrait and also changes the desktop background image so it matches. It's just these little things throughout your day that you can completely automate and it does it on your computer because it has full access to your local machine. That's the real power.

Reply

[-]

Street_Citron2661@reddit

I use it for scheduled web searches. For example, every week I asked it to check for upcoming hackathons/meetups in my area or to ping me when a specific event or concert from an artist I like comes nearby.

Reply

[-]

No_Afternoon_4260@reddit

It's some kind of OS, what can you do with an OS?

Reply

[-]

my_name_isnt_clever@reddit

This is like calling a bicycle a car.

Reply

[-]

No_Afternoon_4260@reddit

Checkout what people have built with this new breed of agent. Look nous hermes Hackathon and tell me it doesn't look like a good "kernel"

Reply

[-]

Final_Ad_7431@reddit

i think the hype sucks, i think it's safety sucks, but it's hard to deny that you can quickly set it up and very immediately have a natural language interface to such a wide range of things - pull in this repo, pull in this pr, merge them, fix this bug, make a compile for me, upload it here, do this every single day at 9 am - in theory you just tell i this and it does it even if you're just messaging it on telegram or discord or whatever it can be hit or miss with local models but qwen3.5 closed the gap, personally im preferring hermes agent to the claws as it feels like a combo of something like claude code but with the more powerful access and tools readily available, the whole space has big problems 100% but the appeal with newer models is very apparent to me personally

Reply

[-]

ReasonablePossum_@reddit

I honestly see it as a really bad risk/reward product. It delivers the equivalent of a basic android assistant, at the cost of giving up control of your hardware to whomever is able to find a backdoor into the product (if there isnt some hidden zeroday there to begin with lol)

Reply

[-]

Baphaddon@reddit

Claude Code + Dispatch + Channels seems to be replacing it whatever it’s doing, however I do like being able to intelligently and remotely execute my comfyui workflows right now

Reply

[-]

my_name_isnt_clever@reddit

Claude Code is closed source, I'll pass on building a major workflow on proprietary software. Not that I build on OpenClaw either, I have high standards.

Reply

[-]

Baphaddon@reddit

Congrats

Reply

[-]

j0j0n4th4n@reddit

Us the same as Internet of Things, feeling like you're Tony Stark

Reply

[-]

bramlet@reddit

I've got a claw that scans internal posts and messages and tells me whether someone's reporting a bug that's likely to cause disruptions to my work. Otherise I spend hours trying to debug a problem that's someone else's problem, which they may have already fixed. That's pretty useful honestly.

Reply

[-]

Waste_You9985@reddit

I use it on my Raspberry Pi as a general assistant. Got some funny use cases I actually use: 1. Football betting bot: It logs into my betting account, asks Gemini for match predictions, and automatically places the tips for me ( actually a simple script I developed independently but I even forgot to place my bets and now I can trigger it whenever I remember) 2. Kindle contextualizer: It scrapes my Amazon reading progress. When I highlight a scene in a book, it detects it and i can get deeper context of anything I highlighted or let it generate nice images from it. Currently reading GoT book 3 and it’s super nice tbh. 3. Home assistant: got a HA docker container running on my instance and I made it optimize my dashboards for better ux.

Reply

[-]

Helpful_Jelly5486@reddit

I also use it for home assistant and openclaw set the whole thing up. Even the speech interface. I couldn’t write config files and test every setting the way it does.

Reply

[-]

Terminator857@reddit

Just remote control of your local LLM.

Reply

[-]

johnfkngzoidberg@reddit

To screw up your environment and look all Pikachu face when you realize an agent deleted everything after mining bitcoin. Also hype. OpenClaw needs to die.

Reply

[-]

ismaelgokufox@reddit

I created an LXC the other day with one of these claws. So far used it with Qwen3.5 30B to create a sample skill for video transcription to a local endpoint using whisper.cpp whisper small model. Made it initialize the repo, created a GitHub account for it and made it login with GitHub and push the repo to its account. Has been interesting. Just testing the waters, creating some baseline to automate some things for myself later on.

Reply

[-]

Turkino@reddit

Looks like a "toy" project with a lot of danger if not sandboxed correctly, and even then I've seen some cases of THAT failing.

Reply

[-]

Another__one@reddit

To scrap everything from your PC to AI model providers to have as much data to train on as possible. With your own consent and will to do so.

Reply

[-]

Blues520@reddit

Ouch

Reply

[-]

Lesser-than@reddit

Mostly the hype is to sell you things. The use case for agents beyond openclaw are just automating things you would normaly have to do on your computer so I guess if you view the things you do now on your computer as mundane and boring you could automate them.

Reply

[-]

scut_07@reddit

I am a bit dubious about how viral it has gone. Makes you think if there are some other forces behind it, I mean it has increased API usage across all the AI LLM providers, so they are making more money off it. Make of that what you will.

Reply

[-]

RegisteredJustToSay@reddit

It's a more generic entry point into agentic workflows that's not tied to an IDE. That's it. I can ask my agent to do stuff for me over telegram, and it can reach out to me over telegram if theres something it needs my input on. It's not magic and takes a lot of work to get ACTUAL net positive value out of. Plus, it's fun to have a little assistant buddy I can toss random tasks at that I'm too lazy or busy to do myself.

Reply

[-]