Get the top HN stories in your inbox every day.
_pdp_
splatzone
What about heartbeats, cron etc? Seems like a major part of the 'claw' appeal is that it can work autonomously, monitor your email inbox for stuff and take action automatically...
mh2266
I hear a lot about people doing this but it really seems like it is prompt injection as a service. eventually the things that can happen when you give the world write access to an unattended LLM that can access both your browser and password reset mechanism will happen.
or someone will just make it email lewd pics to people’s bosses for the lols
crimsonnoodle58
That theory is being tested. So far no prompt injection has broken in:
dmix
I would never use it on my MacBook or any machine but I understand why technical people would want to experiment with something dangerous like that. It’s novel, exciting, and might inspire some real practical products in the future (not just highly experimental alpha software).
furyofantares
I'd love if someone with experience can correct me if I'm wrong but in my experience it can do all of that really, really badly. I find the happy and most likely case for any sort of autonomous thing is that it totally fails to do anything. The sad case is it does the wrong thing. There's just no case where these things make good judgement calls or understand what you think is important.
I do still find some things useful about my nanoclaw setup - convenience and easy scheduling of LLM related tasks. Well, promising actually, not useful yet. But autonomy is not one of those things.
avaer
You could literally set up a heartbeat or a cron. It's faster than setting up the claw.
And if you don't know how, CC does.
aqme28
Ask your claude to make a cron to wake itself up. Done.
_pdp_
You can do both with the cron daemon. But pantalk can also trigger the agent after some notifications are buffered too. So that also is a trigger. You don't really need one massive library. All operating systems have native ways to do all of these things and more.
I don't know. You can even use systemd if you like.
zackify
Hahaha a year ago I did this. Crontab -e
Run Claude -p and Claude already has mcp,'s configured so it can do anything I wanted.
phanimahesh
Crontab entry to read a file and run a prompt?
bookaway
It is truly odd in a way. You had posts here about Google managers or execs saying AI coded something solid in a few days what their own team were working on for months or years, or something along those lines. But people seem to ignore that creating a clone of your favorite "Claw" product seems like an ideal first project for the sea of mid or senior engineers that haven't dipped their toes into the vibe-coding ocean.
You have people talking about the tired topic of the lack of moat for AI businesses. But people should be calling out the moat that most tech businesses take for granted. Forget the moat that prevents other businesses, what about the moat that prevents your own users from creating your own product "from scratch"?
dmix
This is basically just telling people to learn to code
Which IMO they should anyway if they are doing advanced automation
_pdp_
No coding is required. You can literally ask your agent to install and configure it. It is only 2 small binaries and no external dependencies. It cannot be any easier than that.
fermisea
Shameless plug: https://www.supyagent.com - we basically want to give away the integrations for free. For Claude users you just need to run `supyagent skills generate` and you get all the integrations. Works well with cursor and codex as well, and if you want a UI to go with it that can be tinkered, just run `npx create-supyagent-app`
yoyohello13
Seriously, for anyone that knows how to code it’s super easy to setup your own thing. I set up an cloudflare email worker that just forwards emails to my server and Claude can send me emails back. It’s super nice because email already has all the functionality for threads and nice formatting.
Since I control the server and all the code it’s very simple to setup up schedules or new tools.
bredren
I can’t see myself using most of these because I don’t want them having my conversations.
I really want a native iOS chat client that connects directly to my home server.
runjake
You’re missing the point. None of those have the same integrations into other software and APIs that the OpenClaw plugins provide. And not everyone wants to write their own minimal implementation. This is why OpenClaw is popular.
qudat
I’ve been using happy cli, works great
stavros
For my version of the AI assistant, I used a Docker container and Unix permissions:
https://github.com/skorokithakis/stavrobot
All plugins run in one Docker container, but they're isolated from each other by different *nix users, so they can't read each other's files. That's much more lightweight, and you don't have to run one container per plugin.
Crucially, plugins can't read each other's secrets or modify each other's code. I even have a plugin configuration webpage that doesn't go through an LLM, so the LLM never sees your secrets if you don't want to.
undefined
foo42
I've been taking a similar approach with my own exploration.
botusaurus
> But NanoClaw isn't just my personal project anymore. Thousands of people are using it. People are running production workloads on it. Businesses are building on it. There's a real community now.
as OpenClaw and now NanoClaw became "enterprise", now we need a new FemtoClaw to pick up the indie/boutique place
Tt6000
How is this "becoming enterprise"? If anything it now defaults to millions of Linux users being able to access it
andai
How's 100 lines? :)
wat10000
We need to go the other direction. GigaClaw eats $100,000/month in tokens and requires a Threadripper with 256GB of RAM on a gigabit connection just to handle the orchestration.
daemonologist
I'm sure whatever LLM FemtoClaw calls out to will also write a blurb about its growing adoption in production enterprise applications. This sentiment is probably very well represented in the training data.
Someone
Could also make the other part ‘smaller’ and use nail, hoof or dewclaw (https://en.wikipedia.org/wiki/Dewclaw)
undefined
arcanemachiner
Well, there was Picoclaw, but I think it was renamed to Clawlet.
Rapzid
MicroClaw.. No fear of it becoming corporate LOL.
amelius
Putting these NanoClowns inside a container will not protect you from all kinds of safety hazards.
andai
That's the fun part! You spend all day hardening it... run it in docker in a vm on a separate machine. And then you hook it up to your gmail and give it unrestricted internet access :)
plagiarist
An exciting bet on whether the prompt injection will come from the open web or via email!
motoboi
Let’s be honest. The whole thing is just the prevent Claude from “rm -rf / “.
It’s it someone is trying to avoid the thing talking to the internet or reading your emails, it’s just that it sometimes has the strange itch to change some files outside of the project.
arcanemachiner
Wearing a seatbelt will not protect you from all kinds of car accidents.
amelius
Yes. That's why you don't put a Clown behind the steering wheel.
weinzierl
It is more like getting in the car with Stuntman Mike. The risk is not that the driver might make a mistake but that it actively turns against you and a container is not a security boundary against an adversary.
bdcravens
Tesla Robotaxi says hold my beer
InsideOutSanta
Wearing a helmet will not protect you from all injuries caused by jumping off a cliff.
Point is, don't jump off a cliff.
troupo
The nature of these tools is that you tell them not to jump off a cliff, so they ride the bicycle over it. Or a car. Or "you're completely right. I assumed it was possible to fly". Or...
yoyohello13
I’m not sure what docker is helping with that an unprivileged Linux user account doesn’t already do. The scary stuff with claws is unrelated to process isolation.
einarfd
I’ve been building sandboxing for Claude code workloads. So I can let it run wild without breaking my computer. Originally I used docker, but I’m now in the process of jettisoning that, and switching to qemu.
For my use case I want ssh access and being able to use docker in docker. This allows for things like test containers and docker compose. You can get all of that working with docker. But you kind of have to fight docker the whole way.
NanoClaw might have different needs, and docker could work better for it, and I hope so for their sake. But I’m not optimistic.
brcmthrowaway
Can someone explain the special sauce of the claws compared to just use claude.ai etc
lm28469
There is no special sauce, it's mass hysteria driven by fake adoption metrics and people who don't know anything about computers who let "agents" run free on theirs. It's the equivalent of showing a magician cut a women in a box in half to a 5 years old kid... Put them in the same category as the neckbeards getting a hard on every 3 weeks for the past 2 years when they get to see the new version of ThE PeLiCaN On A BiCyCle... I wonder how long the circus will keep on going, at least it's funny to witness from the outside
Zetaphor
I've found real utility in it, but the hype definitely exceeds the current capabilities
samrus
I agree. Its similar to the dotcom bubble. Alot of those websites had real utility as well. Its just the hype created a liability in investment that wouldnt be fulfilled
dmix
This is a classic early internet style snarky comment you used to find all over forums
lm28469
> Some comments were wrong about one thing long ago hence every comments remotely similar about any other remotely similar thing will always be wrong
That's how this argument sounds... And it really isn't a strong argument
enraged_camel
Damn son, you sure sound salty!
lm28469
Better salty than tricked by smoke and mirrors into thinking the singularity will happen in two release cycles and that chatgpt will cure cancer and poverty by 2028 lmao
fooster
It is a huge unlock. Ignore this snark and try it yourself. Any agentic use case you can imagine you can start to tackle.
Openclaw itself is buggy but the idea is amazing.
Gormo
> Any agentic use case you can imagine you can start to tackle.
If "agentic use case" is shorthand for "use case that would benefit from giving non-deterministic systems blanket access to private local data and external accounts" than I can't imagine any such use cases.
juleiie
You really should open your mind to what this tech can achieve. Sooner or later it will click in a way that permanently alters the reality
stavros
They're "always" running, so they can notify you out of the blue, without you having to initiate a conversation. It's really nice UX to get a message from my assistant saying "hey, it's time to leave for the gym, and don't forget the supermarket bag because you're picking up milk on the way back, as you've run out".
mpweiher
Dunno, my calendar reminds me "out of the blue", without me having to initiate a conversation, that it's time to leave for the gym, no "claw" or "ai" involved.
I always have my backpack with me, so if I need milk I can pick it up on the way back. And I am pretty sure that I have to notice if I need milk myself.
The tech sounds cool, but whenever I hear about actual applications, I don't see the point.
signatoremo
That's because you just lack of imagination. Imagine if you have a human personal assistant, what would you ask them to do? Examples:
"Find me the cheapest ticket to Las Vegas for the first week of June. Buy one at anytime that you think is reasonable. Wait until no later than two months from now before buying. Get two tickets if my brother can also go".
"Email me if anyone posts a Sega multi mega for sale. But only if it's in black color".
I have no idea if OpenClaws can already do such a task or not, I don't have one, but it opens up new possibilities. If it isn't there yet, it will be.
stavros
If you don't have a need for a personal assistant, that's fine, not everyone does. That doesn't mean nobody does.
The milk thing was just an example of a tool that can intelligently combine things for you, not a literal "it's a calendar with a milk function".
This is a bit like "if I want to call my friends, I have a phone a home, why would I need a mobile?" which somewhat betrays a lack of imagination.
dgellow
Everything I’ve seen about it feels so over engineered
netsharc
Hmm, Google Gemini has access to my Google Tasks and can set reminders. It's also asked me if I want it to check something at "tomorrow 9am", and when I said yes, it managed to do that.
stavros
Yeah, that's kind of like it. Agents just have many many more integrations, so they can do many more things. For example, it knows all my preferences, and can search for flights and say things like "this one is more expensive, but skipping the morning wakeup is worth the $20".
caminante
But have you had consistently good experience with Google Gemini and Google apps? Or read the mixed reviews?
For me, Gemini has been hit or miss and somehow less useful than Assistant was 2+ years ago.
dimitri-vs
How would it know you've ran out of milk?
stavros
I told it when I noticed. I made a little pendant with a mic I can speak into and it goes to the bot.
brcmthrowaway
How do people afford this?
andoando
Claude max $100 is way more usage than I need. And yeah its not running all the time, just has a heartbeat file telling it how to check something and run
stavros
A subscription, really. It doesn't actually run all the time, it just has a cron job that makes it feel that way.
qudat
I just tell CC to create a cron job systemd unit
jesse_dot_id
Haven't you ever wanted to create a gigantic attack surface for your digital life that is always running and just aching to be pwned?
samrus
Whatsapp api integration + cronjob
Thats it. Its just pets.com
Edit: although as a counterpoint to my cynicism, just the intgreations and deamon-ness can be a game changer if the user experience is good. Thats the critical thing. If you can actually delegate tasks to it and not worry about them then it'd be great. But if your gonna have to worry if its done properly, or that it would delete your emails, then it doesnt work yet. But the dream of a robot assistant is inviting. I just dont think the underlying AI is there yet
gas9S9zw3P9c
It can schedule stuff and run in a loop, so it's like claude combined with cron. Truly amazing technology.
saberience
There is no special sauce. They are claude or codex in a loop. The loop is facilitated by basic cron jobs. That's it.
Ai Agent as it has been for months, plus skills, plus a cron job to prompt it to do things every 20 minutes or 2 hours or however often you want.
sailfast
Crons. A local daemon. System access as a user with the ability to listen to changes. Some idea of shared “memory” between sessions. Provider agnostic about AI. Multi-model.
dimitri-vs
It's for people that don't know how or don't want to be bothered with setting up a messenger integration and a scheduler.
Xx_crazy420_xX
I can't believe the solution is creating uncompatibile branch and forcing users to use cladue for resolving merge conflits. Why not bake in the dual compatibility?
samrus
See your not thinking with agents yet. You need to be more like me. I have claude chew my food and wipe my ass. Saves a ton of time
jimmydoe
you may slot in podman, but apple container is not very good atm.
sergiotapia
I installed nanoclaw last night funny to see it here on HN.
It was easy to install it, and get it running. I could @Andy message it on whatsapp but after that it fell apart fast.
I asked it to login to Facebook and check my notifications, and it started saving credentials and random things in the repo as json files. And din't work. It was hard to even figure out what was happening and why it didn't work.
Then I tried messaging it again and it didn't respond to me.
These things are extremely brittle despite the enourmous amount of github stars. I think it's just normies starring things trying to get on the train unfortunately. The promise of an AI Jarvis is unrealized still.
arsalanb
I'm surprised that the developer experience around sandboxing on macOS is generally so bad. Seatbelt is in limbo and apple containers are just a pain to work with as some have highlighted in this thread
ed_mercer
Is there any developer experience on Apple products that isn’t bad?
avaer
Ironically, the whole "claw" thing reminds me of the time everyone was scrambling to get on the container orchestration bandwangon.
nsonha
TBH I'm not sure why there is a whole debacle around security with openclaw (obviously you should run it in a sandbox) and if it makes sense for these bots to tackle sandboxing themselves. Now I have to trust their sanboxing vibe code? If not, then I have to run them in another sandbox and deal with nested virtualization.
seertaak
If you use Arch Linux, there's now an AUR package to install it -- made by yours truly.
Get the top HN stories in your inbox every day.
You don't need to use OpenClaw, NanoClaw or any of these new variants. You can literally use Codex, Claude Code, Gemini, OpenCode for the same thing. The only thing that it is missing from all of them is the communication channels because none of them come with native communication tools like OpenClaw.
But this is not such a big deal.
I made an open-source lightweight daemon in Go that fills that gap. All it does is to provide the means to connect to popular messaging systems like Slack, Discord, WhatsApp, Telegram, etc. and expose this all through the CLI.
The project is hosted here: https://github.com/pantalk/pantalk
My personal realisation recently has been that the unix way is the best way. We just need to go back creating daemons and lightweight composable CLIs and let agents do their thing. They are increasing being trained to operate the command-line and they are getting pretty good at it.