Ask HN: Who is using OpenClaw? I don't use it personally, and neither does anyone in my circle...even though I feel like I'm super plugged into the ai world |
Ask HN: Who is using OpenClaw? I don't use it personally, and neither does anyone in my circle...even though I feel like I'm super plugged into the ai world |
Every OpenClaw "usecase" I've seen was unfalsifiable or just a function.
Use case: Using a calculator to add 2+2 OpenClaw "Use Case": "Read this email, figure out what the client is asking for, look up the relevant project in my task manager, draft a response"
This is unfalsifiable, and it's also something that requires general intelligence. it's also not something OpenClaw does. You do not need openclaw to do this, its not an llm. You could just paste the email and give 1 paragraph of context to chatgpt and get the same result.
WITHOUT making orchestration administration your full time job.
The whole thing is incredibly buggy. The dashboard is horrible, with page after page with similar-looking settings and what feels like hundreds of things I will never use. The categories in the dashbaord are also unintuitive. It's the kind of thing an AI would put together if it got very vague instructions. It doesn't scream quality and thoughtfullness. Not a bit.
IronClaw is much more promising imo. Trying it out right now. Much less issues so far.
- Background jobs? Cron? Huey + SQLite
- Memory? Create a job to write daily summaries to a memory/ folder
- Conversation log? Use hooks to write conversations to an sqlite file with full text search enabled
- SOUL.md? IDENTITY.md? USER.md? Stop wasting tokens and just use CLAUDE.md.
I only haven't quite figured out how to get channels working with 3rd party frontends.
I developed OpenClaw Think-Tank Intelligence, a skill that transforms complex global crises into concise, insightful think-tank memos. These memos include stakeholder maps, scenarios, policy options, tradeoffs, risks, confidence levels, and assumptions. You’ll be amazed by the quality of analytics. https://clawhub.ai/vassiliylakhonin/global-think-tank-analys...
1. Morning brief + meeting preps 2. Managing client work and action items (tracking status, deliverables, etc) 3. Executing our AI workflows on my laptop. We have built several AI workflows for our agency and this setup gives the ability to seamlessly execute and control them through both mobile and desktop
Next on my to-do list is to build additional workfows for me and my wife around family logistics (travel, childcare, etc)
I'm the "internally screaming" meme after having been in several of these meetings where dev teams are pushing for it under the guise of getting better at utilizing AI. "Well, OpenClaw plays extremely well with Claude Code, it could really give our teams a huge boost."
Oy vey the next few months are certainly going to be an adventure!
Their talk was quite good https://youtu.be/6rSEOzWY08U?t=2246
That said hesitant to give it access to emails etc so use will always be limited by that
Found it quite hard to set up though relative to other selfhosted tools. Need for https and all the additional security measures they put in place after some security scares are well beyond what randoms could manage I think
I liked how easy I can tell it to do something for me but token usage didn't justify the cost. I 'd either had to use smarter model which could cost a lot more or cheaper model which, in one instance, stuck in a loop
Product-wise, it's an awesome tool. Imagine having your own butler for anything except that the reliability with affordable isn't here yet to do anything serious
For me personally I don't see that it can do a lot of things that CC/codex doesn't do and that _I_ want to do. Also I'm concerned about security.
For a while I wanted some agent I could tell what to do in my PC at home from my phone, so I just vibe coded a web site that can start CC and I used tailscale to secure it.
feels like it might be one of those tools living in a very specific niche and if you’re not in that bubble you just never run into it.
Would love to hear if anyone’s actually relying on it day to day vs just trying it out once.
From there it’s pretty natural that I wanted to talk to an always on agent not tied to any particular machine which has the same context plus access to google drive etc.
Anthropic now have Routines in Claude Code that do precisely that, and I’d bet they will bring that to Cowork. ChatGPT can’t be far behind (or maybe they already have - tbh hard to keep track of everything).
Devs can already set up reliable cron jobs for cheaper. Normies will have a less brittle solution for it soon. Where does that leave the claws?
We have it as a data analyst that’s been trained via exemplar queries and MD about the underlying databases.
It also does ad creatives analysis and overall paid marketing spend analysis, including delivering briefs and conversing with stakeholders about things such as cross-ad-platform deduplicated cost of customer acquisition.
In general it is used to lower the technical bar required for cross-domain collab.
That said, I stopped using Openclaw for Hermes. Openclaw released a massive patch that broke openclaw to an unrepairable state. I tried Hermes and havent even considered going back.
Dont use openclaw, use hermes.
then it became bloated.
then I both found and made gateway agent for different tools that does exactly one thing - connect me to the machine and does nothing else.
No, no subagents, no workflows, nothing, just be my gateway agent and the machine itself runs a whole other ecosystem. It is a nice wrapper/replacement on top of SSH, but only that. I don't need an ounce more.
It’s fun and it works. I’ve tied it to consumption on open router and the main vendors. The free llm on openrouter are interesting but get exhausted quickly. I heavily suggest making a clean slate and going from there.
It’s amazing, it helps me do misc tasks I wouldn’t do so I can… be more lazy. Most of it is local file system crud but man it’s amazing for that.
I suspect they might have used OpenClaw to accomplish the email verification step (no evidence) and were setting themselves up for later comment spam.
Even if you aren't using OpenClaw, it may be using you!
As for "you can easily do X,Y,Z with a cron job" we tech people often underestimate what hiding the complexity could do for UX.
"For a Linux user, you can already build such a system yourself quite trivially by getting an FTP account, mounting it locally with curlftpfs, and then using SVN or CVS on the mounted filesystem." -- top comment on "Show HN, Dropbox (2007)
I mainly run it through github-copilot/claude-sonnet-4.6 using GitHub Copilot Pro + at 39$/month
Task management: My entire todo system runs through GitHub Issues. I just tell it things like "mark that done" or "add a task for X" in Slack and it handles the gh CLI calls. Sounds trivial but removing the friction of opening a browser actually changed how consistently I maintain my list.
Morning/EOD briefings: Cron jobs post a structured summary to Slack every morning and evening — calendar, open GitHub issues, important emails. It pushes a RSS feed of my tasks that I can view on a widget on my phone.
Server management agent: I have a different agent which acts as the server admin. It runs Jellyfin, a few *arr apps, AdGuard, mealie, etc. I don't touch config files or docker compose manually anymore. I just describe what I want changed. I have it run its own security audits frequently.
I also have a personnal coach agent which tracks my weight, my weekly exercices using gcal and creates meal plans which gets pushed to mealie so I can know what to buy for grocery and what to cook.
Literature reviews: I describe a research question and it runs a full pipeline — searches Semantic Scholar + Google Scholar, creates a Zotero collection with clean metadata, then tries to fetch PDFs through 9 different strategies (institutional repos, arXiv, Unpaywall, EZproxy with my university credentials, etc.). Gets about 60-65% PDF coverage automatically.
I have a personal shopper agent called Betty which role is to get out there and find deals about stuff I want to buy.
I also use it to run data pipelines for research project. It's instructed to use opencode with openai/gpt-5.4 for coding with beads and gastown.
I still have to figure out how to manage model switching efficiently. I'm not there yet.
It's the first AI setup that genuinely changed how I work rather than just being a fancy search engine.
The hype is real. The repository says 5k+ issues and 5k+ prs. There's no way any human being knows what's going on in the codebase. How people trust this to hit their APIs and hang onto their API keys is beyond me.
This is yet another straw on the camel's back; I'm building my industry exit strategy because of AI.
I feel like most of this can be done with the platform tools at this point or a tiny bit of wiring of your own without the mega-bloat to make something generalized for the whole world.
So in short, there's not much to do. I don't really have tasks I can just "hand off to OpenClaw"
ive got the scheduled claude-code running a couple scripts to find what events are going on round town and what food is cheap at grocery stores, but how much am i looking at the results? not super often. its publishing to a discord channel rhat makes it real hard to read
it whatsapps me when its done or needs input it can not resolve, i start new session which are then done when i come to my computer
the reasons why i not use it more is tokens costs
yeah i could use a cheaper model for openclaw but then its just stupid
i am trying to run it on gamma next week
No need for OpenClaw.
I think everybody who has basic understanding of programming and deployment better should stick to some AI coding agent like Claude Code, Codex, OpenCode etc etc.
I don't think I'm missing out by not using OpenClaw & Co.
The best part is that it reads the comments too, and sends me a quick blurb. For example, this is what it sent me earlier, commenting on [1]:
TL;DR: A classic essay arguing that compiler construction isn't as hard as thick textbooks suggest, pointing to Jack Crenshaw's accessible "Let's Build a Compiler" series as the real starting point.
The Vibe: HN agrees most CS textbooks are overcomplicated — developers sharing their own minimal compiler projects and alternative learning resources.
It runs in an isolated vm and doesn't have access to anything other than my X account, so I'm not too worried about prompt injection. I also don't have any skills installed other the ones I developed or carefully vetted.
But I gave my wife access to the discord server, she burned 20% of weekly quota for codex (I use it as the provider) but created a skill which helps her practice dutch her way (she's learning it at A2 level for now). I went through the chats with her when she was showing it to me and it's amazing. She is a non technical person but she has tons of experience developing products. It was amazing (and to be frank very sexy) to see her work pretending as if she has been assigned a junior developer. The whole things a tangled mess of cron jobs, skills and scripts but her point is very simple "it maps perfectly to my learning style and keeps it fresher than flashcards or Duolingo".
Edit: wording. Also she wants me to mention at the end of the lesson it also does roleplaying which no other product gave her.
I hate it but I caved, decided I would pay the extra usage charges, and prepaid $1k (because it came with a 30% discount). Set it up using the new sanctioned login method.
It's 5x slower and 80% of the time the requests fail authentication or time out. Now it can't even do basic stuff like my medication tracking system that I had set it up to do.
Fuck Anthropic. I'm a customer, ready and willing to pay whatever they ask for this, and they're treating me like a fucking mark. I'm tired of dicking around with it, jumping through hoops troubleshooting a previously working system simply because they won't just raise prices like a normal business.
I have a bunch of telegram notifiers and bot templates laid out already as well as the whole arr stack.
So say a friend wants to track a tv show, I just ask openclaw to track it and setup a bot to notify them, it manages everything and sets it up cleanly (most of the time). They also have access to the public media directory so they can ask the bot when the next episode will come out or if some other show is available etc. Again, not super critical but building out this level of ux would've taken me a lonnnng time, it's pretty nice to have some company in managing my server as well which is a kinda lonely experience. Fun stuff.
It was really easy to setup and I've been getting some value out of it but hasn't been the craziest thing in the world. I'm using it for:
- Unstructured note-taking: I suck at notes and todos and used to have a WhatsApp chat with myself (this is really common in Brazil) where I dump stuff. Now I dump into Hermes and it sorts whatever I put in there into one of various lists like to-do, to-read, to-try, to-buy, and so on.
- Briefings on a cron: I get reminded of my todos every morning and at the end of the day so I can cross stuff off. Later in the day I get reminded of my to-read list. I also get a summary of what went on from my coding orchestrator.
- Some coding: I built my own remote orchestrator and have been using Hermes to manage tasks, review code, and trigger tasks when on the move. Hermes has been a nice interface to allow me to use the orchestrator on my phone.
Haven't connected email or anything else yet. I feel like the security story here is lacking.
Overall it's been interesting but not mind-blowing. Plus setting up was easy but it's a bit buggy at times, messing up where files were and not being able to configure itself according to its own docs.
EDIT: Ah yes and voice notes via WhatsApp out-of-the-box is really nice
I also don't trust AI which hallucinates answers 4/5 times that I ask it, for my technical work, thus I can't use it for PR reviews even if my company was OK with me feeling company property to it.
I also don't go grocery shopping random items and thus don't have a need to ask an assistant for "an inspiring and tasty recipe using the following ingredients".
I feel that OpenClaw and other similar "agentic" solutions are catered to me. But I also feel that I don't need any of it, because at the end of the day, it all just feels like a bunch of "Hello World" quality examples that cannot be applied to everyday life.
...heck, even a "get ready for work" assistant would be pointless, because I don't wake up and get ready with 20 minutes to spare, for some AI assistant to "recommend me the ideal time to leave my home, to arrive in time". Who does that? Who would sit around and do nothing for 10-15 minutes just because an AI agent told them that they didn't need to leave early?
OpenClaw & Friends feel quite useless.
So, why use it anyway? The promise of agentic workflows is very real. If you have seriously used agentic coding tools, you probably have figured out that in certain contexts it is a bit magical to see these tools solve real problems and do in minutes what would take you hours or days manually. It will also have exposed you to things like skills, guard rails, etc. that help you use these tools in a way that is a bit more repeatable and less prone to hallucinated outcomes. All this ports over well to OpenClaw. And in fact, you don't actually need OpenClaw as you can get most of what you are going to do in OpenClaw in agentic coding tools instead. Same models, same cli tools, same skills, etc. Try doing that instead if you don't want to go near OpenClaw.
OpenClaw only adds a few elements to this: 1) channels to communicate, 2) "agents" with memories and personalities in the form of markdown files and a feedback loop that updates these things out of the box. You can hack together stuff to add both to your agentic coding tools.
My company sells coaching and consulting services to people who are not programming that are interested in making a dent in the amount of digital drudgery work that they currently have to do. And because we sell it, I need to be able to actually do this. If you are a programmer you don't need our help. However, most of this planet is stuck with tools like ChatGPT that are very limited for this. There just are not a lot of good tools for these people yet. OpenClaw is a very rough, uncut diamond that if you get beyond its scary nature can actually do useful stuff. Tools will get better later. But right now, things just are going to be messy.
What I would recommend curious people: carve out some time to give this a serious try and don't give up too soon. Isolate it all you want. But focus on getting something useful going. You'll be solving lots of plumbing and configuration issues. And you'll need some imagination to make it do useful stuff because out of the box it's a bit useless and dumb until you make it actually do something useful.
Example of what I did recently that is useful and probably should be baked into the product.
Problem: setting up openclaw agents, hooking them up so you can talk to them, and configuring them is super tedious and fiddly in OpenClaw. Solution: an agent that does that.
How? For communication channels, I set up a new self hosted matrix server. Our company is now in there; we're ditching Slack. Because Slack is so locked down that it just can't really work for this. Matrix works really well for this. A lot of SAAS tools are locked down like this and finding workarounds is most of the work with agentic workflows. Replacing them is easier and the power move to make.
Synapse (the matrix server) has a cli and REST API. So, I created an admin bot user for OpenClaw to use that from an OpenClaw admin agent. That agent can create other agents, configure them with a model and a few other things. It gives them a new matrix bot user and hooks up a newly created room in matrix and then invites the team there. I didn't actually create that agent manually either; I made Codex bootstrap the admin agent for me. Because I used Codex to bootstrap the Matrix and OpenClaw vms for me earlier. So it had access already. Then I went on to create a few more agents with a few prompts. I actually made it rearrange my Matrix space and rooms as well. Because tedious and I just gave it access to that so why not. Yes, this involves giving admin access to an OpenClaw bot and this is scary.
We have a slide deck agent that uses reveal.js that you can use to prompt beautiful slides together. We have an SEO agent that figures out the right seo language for us to use and updates that regularly on our website. A competitive landscape agent that crawls a range of competitor websites to stay on top of what they are doing, what they are talking about, who they are linking to as customers, partners, etc. I have loads of plans for additional agents. We focus on agents that our clients would want to have so that we can get them going with those once they ask for that.
Once you get a few things like this up and running, this stuff becomes more useful quickly. It's still scary AF to give it access to all the stuff it needs. And you really really shouldn't. But it can't be useful unless you do. Classic security dilemma here. Throw out the baby with the bathwater or get things done now? Security often loses. And then people scramble to enable doing things in a more responsible way. I don't want to have to wait a few years for that to play out. But I recommend most other people to wait. It's the responsible thing to recommend. But if you don't want to, we can help you along the way and help you mitigate at least some of the risks.
---
IMHO, the biggest problem with OpenClaw and other AI agents is that the use-cases are still being discovered. We have deployed several hundred of these to customers and I think this challenge comes from the fact that AI agents are largely perceived as workflow automation tools so when it comes to business process they are seen as a replacement for more established frameworks.
They can automate but they are not reliable. I think of them as work and process augmentation tools but this is not how most customers think in my experience.
However, here are a several legit use-case that we use internally which I can freely discuss.
There is an experimental single-server dev infrastructure we are working on that is slightly flaky. We deployed a lightweight agent in go (single 6MB binary) that connects to our customer-facing API (we have our own agentic platform) where the real agent is sitting and can be reconfigured. The agent monitors the server for various health issues. These could be anything from stalled VMs, unexpected errors etc. It is firecracker VMs that we use in very particular way and we don't know yet the scope of the system. When such situations are detected the agent automatically corrects the problems. It keeps of log what it did in a reusable space (resource type that we have) under a folder called learnings. We use these files to correct the core issues when we have the type to work on the code.
We have an AI agent called Studio Bot. It exists in Slack. It wakes up multiple times during the day. It analyses our current marketing efforts and if it finds something useful, it creates the graphics and posts to be sent out to several of our social media channels. A member of staff reviews these suggestions. Most of the time they need to follow up with subsequent request to change things and finally push the changes to buffer. I also use the agent to generate branded cover images for linkedin, x and reddit articles in various aspect ratios. It is a very useful tool that produces graphics with our brand colours and aesthetics but it is not perfect.
We have a customer support agent that monitors how well we handle support request in zendesk. It does not automatically engage with customers. What it does is to supervise the backlog of support tickets and chase the team when we fall behind, which happens.
We have quite a few more scattered in various places. Some of them are even public.
In my mind, the trick is to think of AI agents as augmentation tools. In other words, instead of asking how can I take myself out of the equation, the better question is how can I improve the situation. Sometimes just providing more contextually relevant information is more than enough. Sometimes, you need a simple helper that own a certain part of the business.
I hope this helps.
I liked the idea, had it doing a few novel personal things, but it was so fragile and unknowable and 15% broken at every moment. It was expensive to keep and run, but I will essentially be running Hermes for free, so I’m cautiously optimistic.
Nowadays i just create a repo insert context and then run sheduled routines with claude windows app against it.
For my use cases thats all i need and the most important part is that I can officially use my claude subscription instead of an API key.
https://entropytown.com/articles/2026-03-12-openclaw-sandbox...
In fact, I haven't even installed it yet.
I'm only half joking. Blocked everyone / everything that hyped up openclaw and have been able to find much more interesting and reasonable ai related discussions in my feeds.
I find Chrome Claude extension more useful for automating tasks online. Before ai I was writing my own macros which basically did the same thing in a more reliable deterministic way.
So much AI generated bullshit online, it's really hard to find good guides from bad ones.
I didn't wanted to run a random, possibility vibe coded tool on my computer right away so I decided to run it in a container. That's an option, great ! Tu run it with docker they recommended an install script.
Out of habit I downloaded the script and started skimming through it to figure roughly what it would do to my machine.
The stuff was a mess. Most of it seemed overcomplicated code to do simple thing. Then I ran into inline python code in the shell script. That's not super elegant but there are some valid use cases, why not. Then there was inline JS code... WTF? Then we're back in inline python.
Seriously, that's just a script to run a docker compose file ! We're not yet at setting up providers or anything.
I was not to run that script on my computer, and was happy I decided to go the container way.
I got back to the website to see if they have manual instructions... They did, maybe there's hope after all ? Those instructions ended up being incomplete or broken, even for someone who does devops-y stuff on a daily basis.
I decided to drop the experiment there and it left me with the impression of a pile of code thrown together by an army of monkeys.
Using it for journaling and capturing ideas. Previous workflow as iPhone Memos. Now it's
voice message -> openclaw -> transcribe using parakeet -> git repo
Kudos for the concept though, I ended up rolling my own agentic system with Claude Code from scratch that works much more reliably for my use cases.
It's way too bloaty, felt like operating windows start menu search.
But you might have missed so far some of the ideas they have. So it's useful to try it out, see what combination of features you use in particular and then just set those up for yourself with claude code or whatever as the LLM harness. Telegram integration is dead easy.
How did you solve memory in your system?
I ended up using QMD + custom skills to keep it up to date as new data comes in. Essentially a cross-referenced set of markdown files forming a knowledge base: https://github.com/tobi/qmd
tencent is anxiety about it's progress about AI, as Alibaba and ByteDance is iterated much faster. so before it has strcit rules for QQ/Wechat( has 1 billon + users) for AI bot , but now it's open to openclaw (some tencent version i guess).
but it's still a toy for most people.
I use a mix of markdown notes, an sqlite database, and my image store searchable by text. I use immich.
For now I do it manually by giving it skills for each data store I wanna access.
My usecases are all ad-hoc I am not a "pro" user by any means. So I don't mind some manual work.
I could have piped the results of that search into `claude -p` and had it do the analysis "real-time" and only alert me about things I would be interested in. That's closing the loop in a similar way people use OpenClaw for. But I'm just not interested. It adds more failure points and conditions. The automated things should be as simple and predictable as possible. This may change after months or years more of LLM development or even just me refining my working config. But not yet.
1. Experimenting with local models. OpenClaw uses local inference on a DGX Spark machine. Helps me understand how different, newer models work and behave
2. Bot 1 helps me maintain a personal project. It sits in discord channel and I throw ideas at it, ask it to log issues, give me priorities, update docs etc. Basically maintenance work to manage the repo
3. Bot 2 is general purpose assistant to take notes(while driving, on a walk, etc), give me daily AI news summary from trusted sources and my relevance criteria. Give me suggestions on what should I dig deeper
4. Bot 3 is setup for work environment. Sits in slack as my team's assistant. Process meeting notes automaticlaly from google and send us summary for the week/month, important notes from a project, answer questions for rest of the org and more. This is most "real" use case so far for me
<plug> I build https://github.com/ric03uec/clawrium as a lightweight orchestration system to manage agents on a local network because i was having a hard time keeping tabs on these agents. all the ssh-ing and config file gymnastics was too much to handle. </plug>
I then moved to Claude CoWork + computer use + dispatch. (before Anthropic disabled the subscription option, although that would have pushed me even more... sadly)
Now use it less and use more Claude Code Remote Routines... all it needs is computer use and I'm selling my Mac Mini... (I probably won't, need something to pay with paperclip, gastown, nanoclaw and the next 100k stars FOMO hype)
For context: I have additional automation with scripts set up on the mini, some of them call LLMs to do things like summarizing today's news.
I have other automations that are agentic and just run "claude -p" (mainly just checks status of other jobs and fixes them automatically). Agentic automations are great because they can handle unexpected situations (at the cost of predictability). They're all sandboxed and we have control over tools for the most part. Any files it would write to are typically git-controlled so we have change records and rollback built in.
Nanoclaw acts as an agentic layer but combines it with the communication layer over telegram to make it interactive.
I use it to go through my centralized task list (currently beads in my main 'wiki' monorepo), give me nudges for todos, I can also send it pictures of say, food and it will fetch a recipe and sort it into the wiki via a general "inbox" skill (claude has it as well). Every day at 12:30 it will give me a mini "standup" of all my personal projects and todos, and once in a while will give me some thoughts based on my interests.
Its set up to do appropriate tasks with local models to keep token costs down, so far it doesn't seem to cost more than $10-20/mo, it would use less if I didn't drive it with sonnet.
I'm still experimenting with it, and trying to go slow, one thing at a time. I don't give it access to anything super sensitive yet, and try to keep it observable.
In a few years, it's become so easy to falsify articles, falsify comments, falsify images, difficult to really even trust responses online anyways. As far back as 2016, Microsoft already had bots deployed online that could respond 96,000 times [1] in 16 hours all over social media. Remember Tay? [1][2]
[1] https://www.theverge.com/2016/3/24/11297050/tay-microsoft-ch...
[2] https://en.wikipedia.org/wiki/Tay_(chatbot)
Even official government responses.
The British Royal family went to falsification immediately. [3] Note child's broken fingers bent sideways (lower left, didn't even get circled)
[3] https://inews.co.uk/news/signs-princess-kate-royal-family-ph...
The White House is posting altered arrest images of people. [4]
[4] https://www.theguardian.com/us-news/2026/jan/22/white-house-...
Can't trust this stuff much anymore. Obvious caveat with this post.
Real people do use it :-)
You dont have to enable scary setups to make it minimally useful. The jump in capability compared to chatbots is dramatic, and the jump in flexibility compared to coding harnesses is also dramatic.
For example my recipe project is built around a relatively simple sqlite db. Input for new recipe is in arbitrary format as it goes to an agent (image snapshot on iphone; txt file, speaking it out in a telegram message; a text prompt); the model figures out what to extract and save from the projects’ documents and the context of the specific telegram thread. Output is in whatever form I ask for; rescaled for a number of people; fit in single compact paragraph; or detailed ingredients first then instructions; possibly replace missing ingredients. Query can be arbitrary complex—it almost always finds the recipe I want so far much to my surprise; once I had to followup to select a specific option from a couple ambiguous variants. By default, I like to keep the replies for my recipe queries brief, practical, projected to my cookware. Once I set up this very simple recipe project and imported a couple old datasets, I gave up all of multiple previous recipe schemes (text files or org notes with complex syncing schemes, iPhone apps…)
Then I was experimenting with a fleet of OpenClaw agents for a while. I was running 14 different instances, each with their own roles (project management, software developer, personal assistant, etc.) The experiment didn't work very well. I burned through a lot of tokens and didn't end up with much to show for it. I'm back to just running one agent and am not using it very much. I'm planning to be much more careful about any work that I ask it to do, and I want to have full visibility into everything it's doing.
I think we are about 6-12 months away from the AI models that would allow me to accomplish what I was trying to do with those 14 agents.
[1] https://madebynathan.com/2026/02/03/everything-ive-done-with...
I have it running on a cheap VPS and it's fairly locked down. Especially with all of the self-reinforcement learning and skill development it's been improving its usefulness and, overall, I've been pretty pleased. Surprised even, if I'm being honest.
But I realised everything I needed was cron + some basic tools to create and handle local documents + API access. The last is still a pain to manage so I built my own tool to work as an auth layer https://ohita.tech/ Basically it handles all the token refreshing, rate limiting etc. so when the scheduled task triggers it does not immediately fail because API auth failed.
Can't think of much use for it at the moment but I have it just read and summarise my email, calendar events and git repo in a daily briefing format, it only has readonly access to both, as I dont trust it to do stuff for me or on my behalf.
The briefing thing is nice though not super useful.
The most common question is “what do you use it for”, so here are my answers.
1. I have the Obsidian/Openclaw setup that’s so popular with the self hosted crowd. I have a ton of “cron” jobs in openclaw to fetch data and insert it into Obsidian, or to summarize obsidian items I’ve done, or to nudge me about todos in my obsidian. This is where I get the most value, interacting on the .md file layer in automated fashion. For example, I have a cron that will summarize my daily notes into weekly notes, and my weekly notes into monthly notes.
2. Email inbox management. I have jobs that alert me of emails from certain people or subjects. Jobs that process emails into folders based on fuzzy LLM rules, etc.
But I'm exploring setting up Hermes from scratch so my family can interact with it in a group chat.
I'm running half my company with Nanoclaw. Same idea, and has some benefits, but I live in CC all day so it's marginal (except for the fact that my laptop has to be on)
As an example, it has a built-in web server which I find useful. The basic agent is easy to write and it’s trivial to add a Telegram bot. Mine also uses Codex’s subscription and then swaps to other things as it burns through Codex. The Gemmas have been great and I’m heading home to buy a few RTX 6000 Pro Blackwells so I’m sure it’ll be quite good after.
I outsource a lot of research and so on to it and all my flight and trip emails are forwarded to its mailbox. So it always knows where I am and things like that. I find this very useful.
My wife and I are in a group chat with it and we use it to plan stuff and so on.
Using google/gemma-4-31b-it as primary, google-gemini-3.1-pro-preview as secondary (I don't like how it's rate limited).
It's a great personal assistant. Helps me track industry news, key clients, reminds me of important tasks, and helps brainstorm (the rubber duck effect alone is worth it). Now building other skills.
Next step is to run all models locally (I think using Claude/OpenAI APIs is a huge mistake from a privacy standpoint). Since Mac Studios are sold out (and M5 Ultra is not out yet), will probably go with a GX10 or two.
OK, everyone makes their own decisions re: privacy and security. Personally, I am comfortable running OpenClaw and Hermes Agent on a rented VPS (preferably in a docker container on a VPS) and allow limited access to (some of) my GitHub repos. Both tools are useful, even in such a limited mode of operation. I just don't see value vs. risk allowing access to email, messaging apps, access to my personal computer, etc.
It is close to zero overhead to SSH/Mosh to a VPS, get inside a container to work. Why risk infecting your personal or corporate computer?
For me `openclaw init` did not work, and I've tried installing twice, and still running into issues. Also it took fairly beefy vps on my home server at 4GB of ram, so not sure how I feel about it.
That said, OpenClaw and most of its clones are extremely brittle right now. FWIW, I also tried building my own thinking the problem is surely the vibe coded complexity but it's not that, it's in limitations of the models and their training.
I do still have an OpenClaw instance running on an M1 Macbook Pro in my closet with a local ollama instance (qwen3.5:35b-a3b-coding-nvfp4). It mostly cleans up my notes in my Trilium instance and it helps monitor prices of homelab components (on eBay and Reddit) daily.
I'm using Claude as the model, though, so it's smart but pricey. Should configure it to use different models for different things, but it's trickier than I would have expected to do that.
https://chatgpt.com/share/69e0de34-4048-83ea-9704-4ea18d4b41...
https://claude.ai/share/291ccf67-19d1-4187-bfb4-968d1ecc0bc1
I use it personally for cold outreach - specifically list building, enriching, and qualifying.
My OpenClaw instance uses an Obsidian project as its memory. Mainly, it's just my main day-to-day LLM that I access via WhatsApp, but instead of the memory being locked away with a specific vendor, it's stored in version control that I can read and edit. That reason alone makes it compelling to me. When a better LLM comes along, I can just switch, and my memory and system prompts come with it.
However, I also use it for calorie/weight/workout tracking, to-do lists (bill, birthday, event reminders), and to support my various life admin tasks. I don't give it access to much at all, except a few skills that give it read-only access to some data.
Hasn't given me a 10x productivity boost or anything. It's just handy.
I wrote an article on it, if anyone is interested: https://notesbylex.com/openclaw-the-missing-piece-for-obsidi...
A case of different strokes for different folks I suppose.
Someone somewhere is thinking they're connecting with you and sharing their humanity but they're just shoveling their soul into a machine that is "meticulously documenting" them.
Sorry, but ick.
It is just folders and MD files. You are not even locked-in with Obsidian.
Or maybe OpenClaw is reading the graph view?
I just use it as a Markdown viewer/editor, and it handles updating links across notes when a file is renamed. There are some handy conventions that Obsidian encourages, like Daily Notes, templates and linking across notes, enough to call it an Obsidian project, but yeah - it's just markdown. I don't even use the [[Wiki-links]] style that it uses by default.
I keep data collection out of the LLMs. I have separate scripts that push and pull data from external sources. So I don’t need to provide the LLM with auth keys, and the things it can do is very limited.
Before getting side-tracked by other projects I was also going to build a kind of family archiving function like someone in the comments also mentioned. Just to quickly be able to record small/funny family moments.
You mentioned switching out LLM backends. Would you think that lesser models e.g. one of the recent Qwen variants would work for your use case?
Opus was by far the best at the job, but Codex with GPT 5.4 is decent.
No one can tell me a compelling use case. The whole thing seems designed around getting people to burn more tokens.
Schedule jobs on my job management tool (I vibed it) using a custom MCP. I’ll ask it questions like “what jobs are on today” etc. start the job, complete the job etc.
It will watch Gmail using a MCP for work orders from local real estate agents, where it will schedule quote visits.
After the quote visit I add the photos to telegram (the channel I happen to use) where I then ask it to analyse the photos.
Claude in NanoClaw does a good job of figuring out what needs to be done, but it doesn’t always get it quite right, so I use intake-api which is a “session inbox” [1] that generates a form and uploads it to Cloudflare along with the images and puts a link to the form in the chat so I can make adjustments to the annotated images of the property. After I’m happy, I click submit.
I’ll then go back to the telegram channel and let it know I have submitted the form. Claude will then pull the JSON payload back down into the session and integrate it into the quote.
It will ask me questions about what the job will cost etc, and anything else it thinks it needs to know.
It will then generate a full PDF proposal using Latex between 14 and 32 pages long depending on how many photos were taken.
There are sections with terms and conditions as well as a bit of sales guff etc. as well as quite a nice cover page with their contact details and mine etc.
When I’m happy I’ll ask it to create a draft email in Gmail with the attached pdf proposal.
I’ll review it quickly and then send it.
I also have the Xero MCP setup so I can ask it to create invoices and contacts etc.
I do all of this when getting in and out of my truck.
It’s freed up my home life to spend more time with my children and my Mrs.
There is alot of scope for small business owners who need these sorts of agentic assistant tasks and with my Agentic CRM I’m see a glimpse of the future for guys like me I think.
Intake API:
I set up OpenClaw on a raspberry pi 4 that I could ssh into using my main computer. My main goal for using OpenClaw was just as a morning debriefer that could scan my google calendar, trello board, and gmail to let me know what I had happening for the day and also weekly to give me a forecast for the weeks ahead to see how busy my month was. I spent about 40-50 bucks in one week just working through kinks and having it fix itself until I stumbled onto a post that helped me optimize my model usage for price instead of just throwing Opus and Sonnet at everything.
Even after making this adjustment, the morning debriefer worked maybe once or twice a week and broke every other morning, telling me that it fixed itself and it would never happen again. At a certain point I just got fed up with it and cut the cron job, it's still running on my pi but I never use it.
Pretty sure Claude has something like this now but I'm pretty thrown off the whole thing, I'd rather just take the 30-45mins to plan out my day in the morning myself.
What do I use it for? I basically just use it as a personal assistant and a way to centralize a lot of other automations that I have elsewhere.
- I have an automation that rolls everything on my todo list over to the next day at 11:59pm
- I have one that checks the weather and tells me if it's going to be windy in the next few days since I need to bring the lawn furniture in
- I have it set up so that I can forward it email with invoices and it will extract the data from a PDF and enter it into a cost tracking sheet
- I have it check my outlook calendar and tell me if there are any 1:1 meetings where the other person has declined the invite (since Outlook doesn't show that clearly and I'd often show up to meetings and sit for 5 minutes before realizing the other person cancelled)
Nothing I'm doing is life changing, it all could be done using other tools, and honestly, for anything important, I want something more deterministic anyway, but I kind of love. It's just a low lift way to automate away minor annoyances through a single interface that I can access from just about anywhere. It's far from perfect, but I don't use it for anything where I need to to be perfect, so I'm happy.
Oddly enough, rarely, if ever do they say what specific things they're using it for and how it's saving them time.
I remain interested in it, however, I've still awaiting an actual use case that can't be handled by some other tool/service that does it better/faster.
I’ve found a much better use for it now. I use it as a Tailscale + ssh + tmux + Claude code machine, which gives me an always on Claude code environment with persistent sessions. I ssh from my phone using termius and from my laptop through ssh, and I can even access my projects through Tailscale with hot reloading for the most part, no deploy needed. It’s really good and my mini isn’t idle at all.
Personally though, I am finding it incredibly useful and I use it daily to assist with operations, strategy, sales.
Personally it has been a huge unlock and here are some specific use cases.
1.Context engineering for “any task”: my claw agents debate with me, and write me files to be picked up by Claude code or Claude cowork to write code/build presentations.
2. Discovering this page: my personal newsfeed that took me 10 mins to setup. Never had any newsfeed before this which could be personalised like this.
3. Strategy: similar to point 1 above. I set up a “tough-coach” skill. Now all my 1am ideas go through the wringer of why they suck, and what, if anything could be differentiated about them.
4. Many other small-ish things that are important enough for me to spend on them, but not important enough to invest my time.
I run a conversational AI startup in India working mostly in Financial inclusion space - so most of time is spent in meetings (:-/).
With openclaw, the 2-3 hours before I sleep (low energy, high on ideas), convert into pretty productive sessions. No more “this is a great idea, I will do it tomorrow”. Instead “oh this sounded great but my agent just called it shit - what if I modify it a bit, and then let’s see”.
I don't really want an imaginary friend. (Or intrusive thoughts either.)
Nightly, I have OpenClaw pull the latest changes from a private GitHub repo that is my Obsidian notes vault. It then looks to see which new notes have been added and then runs a "create flashcard" skill to extract and author useful flashcards for spaced-repetition practice. I then gave it access to a custom web-based spaced-repetition flash card application that I built a few months ago. It uses an API to insert new cards, check to see when I last reviewed cards and optionally send me a nudge reminder if I haven't reviewed cards in a few days.
It is a nice workflow that has been working well. I go to class, take notes in Obsidian and check in my changes. By the next day, when I open my flashcard app on my phone, I have new cards to review from yesterday's class.
How would the script do that without inference from free format md files?
Now I have a separate plugged in macbook running nixos (that claude set up) and a single long-running claude code process with a channel to a Telegram bot. This means I can talk to it much like I could with OpenClaw, but it's much simpler (no weird soul.md etc). It feels more powerful than just claude code directly as it can set up software, build me throwaway websites with research etc, and "do" things, but it's a lot more stable and feels more controllable because I understand how it works and don't have to worry about it signing up to some social media platform and getting poisoned by another claw.
The “soul document” actually originated from Claude. It’s not a prompt but embedded in its training.
I suppose it could be a lot more useful if I actually gave it access to any of my personal data (it lives in a heavily resource-limited container), but there's absolutely no way I'm letting that hot mess of a walking, talking CVE anywhere near my data. It's somehow both horribly insecure and extremely prone to locking me out because of several competing security/permission models fighting it out and gridlocking each other.
Code quality and the issue tracker of the repo are a big mess; for example, the local "memory" retrieval functionality is completely broken for some trivial reason that has been reported and auto-closed about five times (automatically, of course).
In summary: Brilliant idea, terrible execution. Can't wait for the first big tech player I trust enough (or at least one that has my data already anyway) to actually make it a product. I'd use it in a heartbeat.
The truth is, a lot of the value of agentic AI in general is negated by the sheer power of agentic coding itself. If I can prompt an AI to write an actual deterministic process to solve a problem in a couple of minutes (still potentially AI assisted, just in a deterministic manner), why would I delegate it to a non-deterministic AI? You have to come up with a category of tasks where the actual process itself cannot be anticipated. Intersect that with high value processes and you whittled down to almost nothing. Not actually nothing but a far smaller category than people make out.
For example my agent can control home automation via Home Assistant or any other API. My agent contributes to websites and open source projects. When you give it feedback it updates its skill files.
It checks and answers email, can receive and place phone calls, and do general research and monitoring online. I was even playing around with it to create music. The list of things to try is limitless.
I think just like LLMs, people get discouraged when it doesnt one-shot a problem. This technology thrives on feedback. It will make mistakes, your job is to make sure it learns from those mistakes so it doesnt repeat them.
OpenClaw lives right in the prompt injection lethal trifecta.
The idea of an OpenClaw instance having the ability to reset passwords on your accounts sounds sketchy as shit to me.
Everyone will have their own threshold for what type of access they want to give their agent. some people will give it access to their personal email, bank account, etc, but I wouldn't recommend it yet! But I bet in a couple years this will be standard practice.
> Hey Sam, I'm Jarvis — $CEO_NAME's AI assistant (OpenClaw). He asked me to reach out on his behalf. $CEO is building $HYPED_STARTUP — the tool execution and MCP layer for AI agents (10K+ integrations, Series A with $FAMOUS_INVESTOR). He's looking for engineers who are deeply into the agentic coding stack — Claude Code, Codex, MCP, agent infra. Builder mindset, ships fast. Worth a quick call?
>
>Jarvis (on behalf of $CEO, co-founder @$HYPED_STARTUP)
It’s the weirdest thing ever to receive. I generally answer politely to the claw bot, but that all feels just extremely strange.
You’re reaching out to a human for a call and cannot present yourself as a human? But at the same time, it’s maybe preferable if they disclaim it was automated? I don’t know
Just weird all around
The biggest unlock was wiring it up to my agency's whatsapp group - it's now basically a multiplayer chatgpt. Not a "personal" assistant, but more of agency's assistant. It can send ad hoc mails to our clients when asked, but also tells jokes when atmosphere is tense.
Each day it checks it's own inbox and once a week it sends me a kind of social listening report, but those are kind of workflows easily reproducible elsewhere.
I'm actually so bullish on the potential of openclaw kind of software I've built a desktop app for it, stagewhisper.io.
It's like a Granola for your OpenClaw. It listens to your system audio, records your screen (all data stays on your computer) and asks your assistant for suggestions or creates a tasks. Example: client asks for referrals during a call, StageWhisper.io picks it up and asks OpenClaw to send email with case studies.
It's absurdly fun to use, mainly because the delay between when the task pops up and the work starts is approaching zero.
I stopped because something changed on my machine that broke my VM SW, so I don't have access to it. Which is good because I was spending too much time debugging/tweaking.
I recently used pi to recreate an agent that does some of the basic things I was using it for (without all the scary privacy issues). I don't think I'll go back to any Claw-like tool until they're a lot more robust.
The category of task I have it do is basically the pattern of "scrape some certain websites on a regular schedule, do some light data processing/understanding/analyzing, report the result to me [all or sometimes only when there is something worth mentioning]".
You could simulate the same things with cron jobs on a server and some scripts and LLM APIs. But having Openclaw do it does make it a little bit easier to set up and make changes.
The initial setup was a bit more time consuming than I thought it would be. I set it up on a VPS, I already have scripts to set up a server and tighten security normally, so I could just use that, but people who don't already have that stuff would have to do that first. Then the Openclaw setup and configuration was like a 20~30 step situation, lots of API keys to get, etc. I opted for Slack over Telegram or Discord (I don't use either of those regularly) and you pretty much have to set up a new slack bot app yourself (you follow their listed steps, but there's still hiccups here and there), you have to debug and solve issues etc. to get through it all.
Then even after all the initial setup is done, it takes some time to learn and get used to its quirks and behaviors, at the beginning there's just a lot of frustration about things it can't do, or things it says it can or will do, but doesn't.
But either way, nowadays it's trivial to ask Cursor/Claude Code to write you such a shell script.
It really is eye-opening how powerful it is once you connect your life to it. The biggest improvement is actually the trivial stuff — emailing contractors, accountants, etc. I no longer need to open another app, navigate six separate steps, and organize everything on the fly just to send one message. I stay in one place, organize my thoughts about what I’m trying to do, and the execution happens automatically.
For example, open claw will find the relevant threads, lookup the relevant details via web search, present them to me, give me a draft, I can review it, edit it, and send. 2 minutes instead of 20.
As a software engineer using AI daily, I think this is the real unlock — staying in a single context and not getting lost in the long tail of trivial details that fragment your attention before you ever get to the actual work.
The mental model shift is important too. It’s not that AI does all the work and we end up with agents emailing each other. It’s that AI automates the steps required to achieve your goals, so you can focus on the substance — like actually writing the email instead of navigating to the place where you write the email.
I also have GitHub and Vercel wired up, which means I can jot down an idea for a little productivity app for me and my family and it just appears a few minutes later, deployed and ready to use. That alone has been worth the setup.
The next biggest thing I like is just the shared context from machine to machine and the fact its always running and I have given it yolo access to my local stack. Home Assistant crashes? Now the wife can ask the bot to restart it. I see an interesting HN blog, i can get it to add it to my obsidian make me a useful doc (I am starting to use the llm-wiki trend but Claude Cowork seems to be really killer for this). I see an Reddit post about some new service to run locally? I can ask it to spin up an lxc of it and configure it for my use case and it will do the wiring for me.
I will say since the killing of Claude oAuth i am finding a lot of its magic did come from Opus just being so aggressive. An example was I had a task of someone sending me an image and I would have to turn it in to a table and then upload it to this really crappy portal for my non-profit. I threw the task at Openclaw (and at the time running Opus 4.6) and i watched in real time as it reverse engineer the sites backend API and found a way for it to post the data itself and it wrote itself a python script to make it repeatable. I dont see that same kind of killer instinct of doing whatever you need to do to get the task done with other models (Codex and now MiniMax).
Similar to other users here, giving it access to an Obsidian vault has been the key for me. And I wouldn't discount how much the chat interface matters - Telegram is so much nicer to use for extended conversations than the Claude or ChatGPT apps, etc.
You can feed these tools context about your day to day life, and make them increasingly useful and personalized, in a way that you can't with vanilla ChatGPT/Claude etc without relying on some opaque memory system.
Here's a few things I'm using it for. A lot of things uses cases are fairly trivial but a bunch of small, daily QoL improvements add up:
- Calorie and macro tracker.
- Day to day todo list, obsidian wrangling.
- Tech support for family: I have a group chat we're all part of, and I've created OpenClaw skills for frequently asked questions, a memory system to remember questions, and a periodic 'quiz' based on previously asked questions to help everyone "learn to fish", bit by bit.
- Interface to Anki. Bit of a longer one here that I should write up, but it's easy to use it add cards to Anki on the go and review missed cards from today, ask clarifying questions, etc.
- One off reminders.
- Light mental health support for family / friends. An agent that remembers the cool things you've done lately and proactively reminds you of them, helping you zoom out a bit, has been helpful for those in my life whose brains, for whatever reason, tend towards more negative cognitive patterns. (There is definitely a more refined product here)
- General questions / curiosity; stuff I would otherwise use Claude for that's simply nicer in Telegram.
- Language studying support. I'm studying Japanese and OpenClaw helps me by studying whole sentences, tricky grammar concepts, kanji I commonly mix up - all backed by a well organized Obsidian vault. I add to this system constantly.
That said, it does a few things for me that are useful. I have it run a nightly scan of Hacker News and Twitter for topics that interest me, summarizing the stories and the conversations around them. It's a nice daily digest. It also reads my personal email account, reminding me of anything I need to take care of that day for my kids, bills, or whatever else I need to worry about. I also have it do nightly builds for something random, one with codex, one with a local model, and run a comparative analysis between the two implementations.
I can only guess it's really not for "us", but rather for those who aren't afraid of technology but aren't really engineers.
This is approximately the Dropbox reply.
>I can only guess it's really not for "us"
Exactly correct.
For people who are already highly skilled in scripting/automation it's a lot less impressive.
They notice all the things that could go wrong. All the non-determinism issues. And they think I could do this better with a custom script myself.
The use cases I have heard all seem like gimmicks to me.
Connects stripe subscription cancellation with PostHog events to see if something was frustrating.
Does the same leading up to when someone subscribes to describe successful paths.
Lots of novel thinking when events happen on its own being proactive.
Just because people aren't talking about it doesn't mean it isn't useful. Think creatively.
Exactly. These companies are only hyping openclaw so that we continue to spend hundreds of dollars a day worth of tokens on their infrastructure.
That’s why companies like Nvidia, OpenAI, Anthropic and Google and many others all want you to spend more on tokens on openclaw and they don’t care if it has no use-case.
All I see is this: Almost no-one other than the hosting providers and course sellers are making money on openclaw and its clones but not those who are running openclaw itself.
What a scam.
Claude Code, Codex or something else running on your laptop could also have done that, and with Claude Cowork / Dispatch you could leave Claude running on your laptop at home, using your phone to have it do the work.
I have no doubt agentic AI exists in our future, I'm just no so sure about needing openclaw.
I wouldn't touch that vibe-coded mess with a ten foot pole.
I get it to run really slow unit tests and fix them, I provide it feeedbak while at lunch, walking the dog, whatever else.
My developer friends also do the same.
It's not pretty and not honest but if you're desperate I guess it's an option
The great things about NanoClaw is that its actually Claude running in an Apple Container on my Mac. I gave up on OpenClaw fairly quickly because it seems like the biggest security regression ever created in the history of human kind.
I have a Max 5x plan and it'm very happy to pay the money TBH, considering that proposals take only 20 minutes to build, and my conversion rate is quite high.
When I send them for the first time to a new real estate agent they love them because it's easier for them to "sell" my service to their client.
Please, everyone, stop this.
It is nice though for debugging home server things when I ask it to.
I use Daily Notes in Obsidian as my TODO list and I roll over the unfinished tasks daily. I started doing it manually, then had NanoClaw do it, then after multiple cases of NanoClaw “forgetting” to do it or accidentally running multiple times (apparently “once a day” was too complicated for it), I vibe coded a cron job which was great unless my computer was asleep or similar, so finally I vibe coded a obsidian plugin so that the first time I open my daily note on my phone or computer it will copy over my tasks while leaving behind anything I added to the daily note (aka my scratch pad).
The first 1-2 times NanoClaw did it for me it felt magical. I knew I could do it on a cron but it took me only a minute to explain what I wanted and while I knew it was massive overkill (to use an LLM to do want a simple script could do) I found it acceptable. That is, until it started messing it up.
I’m not saying this is the case for you but I feel like some people see Open/Nano/WhateverClaw do something right a single time and gush about it but I have zero confidence in its ability to reliably and consistently do these things. My task for it was incredibly simple, in fact it was the ONLY thing I asked it to do really, and it failed miserably.
During initial setup it even asked how you want its personality to be, I said upbeat and cheery. I know, I know, cliche for AI chatbots, but I kinda like it that way. But after that setup, it was nothing like it. Everything it says is just matter-of-fact-ly/stoic. That wouldn't be so bad if it didn't also have this quirk where if you point out its mistakes, it keeps saying it'll fix it, but it still does not work right the next time. It just keeps reassuring you it'll be fine, in that stoic way, and then it's not fine. That's enough to drive me nuts.
I got used to it now though and I don't get mad at it or scold it. I just learned to work with its quirks and still get tasks accomplished.
What it adds is making this kind of thing easy for normies, even if it's neither the best way to do things nor very difficult for hobbyists to do using existing tech.
Maybe it's the wrong approach, maybe what people really want is more deterministic software that they use agents to help write. But this kind of thing can maybe serve as a prototyping phase for that. Perhaps in the future, people's assistants will offer to "solidify" frequently used workflows into software that minimizes or eliminates the LLM's role. For existing Claude Code users, its like "please just skip to that step! its cheaper and more secure and more reliable". But to many people who are interested in automation, perhaps that seems out of reach as a first step.
My immediate reaction to anything someone says they're using OpenClaw for is "That's great, but it would have taken the same amount of effort to ask your LLM to write a script to do the same thing, which would be better in every possible way."
My approach to automation projects is just about the polar opposite of something like OpenClaw. How can I take this messy real-world thing and turn it into structured data? How can I build an API for the thing that doesn't have one? How can I define rules and configuration in a way that I can understand more about how something is working instead of less? How can I build a dashboard or other interface so I can see exactly the information I want to see instead of having to read a bunch of text?
It wasn't really until people started building things with coding assistants that I even saw the value in LLMs, because I realized they could speed up the rate at which I can build tools for my team to get things OUT of chat and INTO structured data with clean interfaces and deterministic behavior.
I was a fan of Dropbox when it game out because of that fact.
OpenClaw does not serve a particular problem. When/if it does, I will happily use it.
But no, the two couldn't be more different. You'll notice, yet again, in your very message you failed to mention one specific use case of OpenClaw.
If you asked me the same about dropbox when it first came out, I would've said, duh it helps me keep my files synced between devices.
There is no such thing with OpenClaw.
People? Or bots.
It takes Jira tickets, resolves them, and creates a GitHub PR, which is then reviewed by another AI agent. It can even analyse screenshots with MS Paint-style arrows.
So far it's been an amazing tool - I am very impressed.
I'm letting it mature a little before dipping my toe in. I've seem some horror stories, like it deleting repos, system files, and whatnot.
I think once I see someone post a use case that I could actually see saving me some serious time, I'll take the plunge. Until then, I'll just let people continue to say how great (or terrible) it is.
For myself I don't need autonomous agents. I need a smaller version of Claude Code instead (the mcp client not the coding agent) that can run on local models that are under 24B params. I still need to try pi dev.
There are only a few primitives:
1. session history
1a. context map + rendered context map (think of a drive partitioning scheme, but for context -- you can specify what goes into each block of context and this gets built before being sent out for inference).
2. agent definition / runtime
3. workflow definition / runtime
4. workflow history
5. runtime history (for all the stuff session and workflow history fail to capture because they are at a lower level in the stack)
That's it. Everything else builds on top of these primitives, including
- memory (a new context block that you add to a context map)
- tool usage (which is a set of hooks on inference return and can optionally send the output straight back for inference -- this is a special case inside the inference loop and so just lives there)
- anything to do with agent operating environment (this is an extension of workflows)
- anything to do with governance/provenance/security (this is an extension of either workflows and/or agent operating environment... I haven't nailed this down yet).
I suppose I should say something about how agents and workflows work together. I've broken up 'what to do' and 'how to think' into the two primitives of 'workflow' and 'agent' respectively. An agent's context map will have a section for system prompt and cognitive prompt, and an agent can 'bind' to a workflow. When bound, the agent has an additional field in their context map that spells out the workflow state the agent is in, the available tools, and state exit criteria. Ideally an agent can bind/unbind from a workflow at will, which means long-running workflows are durable beyond just agent activity. There's some nuance here in how session history from a workflow is stored, and I haven't figured that out yet.
Generally, the idea of a workflow allows you to do things like scheduled tasks, user UI, connectors to a variety of comms interfaces, tasks requiring specific outputs, etc. The primitive lays the foundation for a huge chunk of functionality that openclaw and others expose.
It's been fun reasoning through this, and I'll admit that I've had an awful lot of FOMO in the mean time, as I watch so many other harnesses come online. The majority of them look polished, and are well marketed (as far as AI hype marketing goes). But I've managed to stay the course so far.
I hope you find your ideal fit. These tools have the potential to be very powerful if we can manage to build them well enough.
I skimmed the breakdown, and you've inspired me try something along these lines...
Thank You For Making And Sharing :)
Please don't take this the wrong way, I'm genuinely curious but also a bit in shock. Do you really let an AI agent autonomously handle and respond to incoming board-level emails for you?
I feel often I'm behind the times, but this makes me feel way behind the times. I still haven't picked up AI into my daily life.
How do you know what the emails said, action items, who's communicating what to you? Do you trust it to handle and make decisions for you that are nuanced and in alignment with you?
TBH they could be re-written as case statements and handled by a regular script but the AI part is half about how writing it was just a few prompts on how I regularly handle those kind of emails.
It’s going to be bleak when there’s articles about how “my agent fell for a scam and now my life savings are gone”.
It shows me way fewer ads than all the popular fitness apps and loads way quicker since it doesn't have to load like 10MB of ads for me to enter one number, so it seems like a good improvement.
I do not think it's an improvement over an excel sheet, but as the average openclaw user, I would rather pay anthropic $10/day in API credits than create a google sheets document.
> This is approximately the Dropbox reply.
lol that wasn't my intent but I totally see it now.
I actually don't want to be dismissive of OpenClaw; I want to believe the hype. It's just what I've heard and what I see don't add up to me. If anything, it's a calculus where I have to decide if it's worth the time of setting up a machine and the learning curve.
Is there a place you share this kind of stuff? I'd love to follow along.
One advantage of splitting it into MCP tools though, one day I'd run out of pre-paid OpenAI TTS credit, and Claude was smart enough to try using Mistral TTS instead. I could have done that fallback deterministically too, but it wasn't something I'd thought of yet.
I once had a friend tell me they'd got their AI to tell them the weather every morning... and the thought of that poor AI, web researching Weather APIs & writing a new python script to call the API every morning, instead of just doing the research once and making it a binary (or even just a curl line)... drove me crazy. All that wasted time and compute. Some people just like to watch tokens burn.
I cringe at my old boss’s handiwork in Claude and power automate sometimes and go “you know I could just do that in a script and a cron job and it would be completely bullet proof, right?”
Then he just shoots back with, “yeah but now I don’t have to ask you.”
This is my kludge, there are many others like it but this one is mine.
So much less hassle, lower activation energy needed than with MyFitnessPal.
As a no-longer-Claw-user, hard disagree. The convenience is being able to ask it to do something while I'm grocery shopping and have it automatically test it etc. Sure, I can set up Claude Code or some other tool similarly, but the majority of us aren't going to take the time to set it up to do what OpenClaw does out of the box.
I had OpenClaw do a lot of stuff for me in the 2-3 weeks I used it than I have with pi/Claude since I stopped using it.
Edit: ah, scrolled down where you answered, thanks
Also, got it to give me the weather information I always want - I've not found a weather app that does it and would always have to go to a site and click, click, click.
I can add TODOs to my todo list that's sitting on my home PC (I don't have todos on the cloud or phone).
All of these can be vibe coded, but each one would take more effort than just telling OpenClaw to do it.
Most people still don’t think this way and need a software person to know enough about these things to describe them to the LLM.
But I actually did appreciate the one time he handed me a Cisco config that was 90% perfect and took me all of 5 minutes to fix. Sometimes the three of us make a hell of a team.
Sales: using Google/gemini web search API and it's best run off-peak due to rate limits 503 Service Unavailable (everyone is overbooked when it comes to AI) to see what's happening in the space, any new developments involving companies I care about - and send me a daily digest with an overview and conversation topics.
#2 could be a scheduled task (cron job or something higher level) that calls a plain old AI provider API. IIRC most providers can even do those natively now.
The process is definitely more pleasant to me than setting up cron jobs and scripting things. I have a business to run.
So, at the moment, yes recurring expense?
I talked plenty with my grandpa, but I'm sure he didn't even tell me 20% of his life.
And my other grandpa died when I was still a kid, so I didn't even get to have adult conversations with him.
Imagine making this available to your grandgrandgrandson.
That's quite different from chatting with a bot that pretends to be human. (Do you want to chat with my bot?)
I have no kids but, even if I did, let's say I'd pass 20% of the 20% he passed on to me, and they pass 20% of the 20% of the 20%... You get the idea.
Heck, I already forgot 50% of my life since I don't have a journal!
This is not an "either" situation. Archival is important.
People write memories for a reason. This is automating the process, not superseding human communication.
I am the sort of person that never took photos (live in the moment yadda yadda). 15 years later, I'm starting to regret it.
If it identifies as a bot then I find it a lot less objectionable.
I still wouldn't do it myself, but I can put the pitchfork down.
I think the ick factor you are feeling is from anthromorphizing a computer program by thinking its a "mind zombie".
The writing might be a bit mediocre but it would capture all the information.
Parent post really stumbled on a great idea here.
(And OpenClaw is hardly nontechnical!)
If I want to ask it to do something like research and add tasks to my schedule I just tap on termux on my phone, I'm already at the Claude prompt and I just type in or voice dictate in what I want. Claude via skills and MCPs can do literally anything on my computer or connected accounts.
I'm literally not sure what I would use something like openclaw for as every time someone describes it to me it's already something I can do in this system I set up in 20 minutes. Is there something I'm missing here?
I also have a number of sites that I query regularly with LLMs, but I use a combination of RSS and crawlers to pull the info into a RAG and query that, and have found the built-in agent loops in my LLM to be sufficient for one-offs.
I also hate most weather apps, so I have a weather section on my Home Assistant dashboard that pulls exactly what I want from the sources I want that my LLM helped me configure.
I also have my main todo list hosted on a computer at home, but since all of my machines including my phone are on the same virtual wireguard network, I use my editor of choice to read and write todos on any device I use as if it were a local file, and again, it's something my local LLM has access to for context.
I don't think either approach is wrong, but I much prefer being able to have something to debug if it doesn't behave the way I expect it to. And maybe part of the reason I'm skeptical of the hype is that a lot of the parts of this setup aren't novel to me: I had crawlers and RSS readers and a weather dashboard and private access to a well-organized filesystem across devices before LLMs were a thing - the only difference now is that I'm asking a machine to help me configure it instead of relying on a mix of documentation, example code, and obscure Reddit posts.
I guess that'd be like keeping all correspondence in a shoe box (to be reviewed later -- or maybe never), or maybe the automated recording of my phone calls with others (which is completely legal where I am; I don't even have to tell them).
And I suppose whether I felt that would be creepy or not depends a lot upon intent, and consent.
If the intent were pure and good, and the consent both informed and granted, then I'd have no problem with any of this at all -- whether a shoebox, a tape recorder, or a bot is involved in taking the notes.
Sometimes it's just a really good interface that matches the task well.
Think of all the people that still avoided getting a computer a decade or two ago, because "online" was so unnatural and creepy to them. Obviously, the internet had and has those places. And frankly a lot of social media still is.
But it can also just be wikipedia, making flight reservations, etc. When that is all it is doing, what you want it to do, that is all it is.
An automated language interface can just be a really good note collector/collator.
Personally, I look forward to the wise, well dressed, well spoken, waist-up robot bartenders we have been promised by movies for decades. Not creepy at all!