zclaw: personal AI assistant in under 888 KB, running on an ESP32

Anytime I get in a discussion and sufficient number of people disagree, my comments are flagged despite obeying site guidelines. What should I do about this?

throwa3562623• 5 months ago

"LLM backends: Anthropic, OpenAI, OpenRouter."

And here I was hoping that this was local inference :)

micw• 5 months ago

Sure. Why purchase a H200 if you can go with an ESP32 ^^

sigmoid10• 5 months ago

Blowing more than 800kb on essentially an http api wrapper is actually kinda bad. The original Doom binary was 700kb and had vastly more complexity. This is in C after all, so by stripping out nonessential stuff and using the right compiler options, I'd expect something like this to come in under 100kb.

peterisza• 5 months ago

right, 888 kB would be impossible for local inference

however, it is really not that impressive for just a client

Dylan168074• 5 months ago

It's not completely impossible, depending on what your expectations are. That language model that was built out of redstone in minecraft had... looks like 5 million parameters. And it could do mostly coherent sentences.

js8• 5 months ago

I disagree, in the future it might be possible. But perhaps not in English, but in some more formal (yet fuzzy) language with some basic epistemology.

I mean, there is a lambda calculus self-interpreter in 29 bytes. How many additional logical rules are required for GAI inference? Maybe not that many as people think. Understanding about 1000 concepts of basic english (or say, lojban) might well be sufficient. It is possible this can be encoded in 800kB, we just don't know how.

__tnm3• 5 months ago

haha well I got something ridiculous coming soon for zclaw that will kinda work on board.. will require the S3 variant tho, needs a little more memory. Training it later today.

throwa3562623• 5 months ago

Sounds interesting, please keep us posted.

I dont think I have an S3, but plenty of C3. I thought they had the same amount of memory

g947o• 5 months ago

Serious question: why? What are the use cases and workflows?

milar• 5 months ago

for fun!

grzracz• 5 months ago

I don't fully get it either. At least agents build stuff, claws just run around pretending to be alive?

codazoda• 5 months ago

They do build things. The same things.

eleventyseven• 5 months ago

The various *claws are just a pipe between LLM APIs and a bunch of other API/CLIs. Like you can have it listen via telegram or Whatsapp for a prompt you send. Like to generate some email or social post, which it sends to the LLM API. Get back a tool call that claw then makes to hit your email or social API. You could have it regularly poll for new emails or posts, generate a reply via some prompt, and send the reply.

The reason people were buying a separate Mac minis just to do open claw was 1) security, as it was all vibe coded, so needs to be sandboxed 2) relay iMessage and maybe 3) local inference but pretty slowly. If you don't need to relay iMessage, a raspberry pi could host it on its own device. So if all you need is the pipe, an ESP32 works.

grigio• 5 months ago

yeah i still can't believe many people bought a mac mini just for the claw hype

yoyohello133• 5 months ago

I’m running my own api/LLM bridge (claw thing) on a raspberry pi right now. I was struggling to understand why all the Mac mini hype when nobody is doing local inference. I just use a hook that listens for email. Email is especially nice because all the conversation/thread history tracking is built in to email already.

theturtletalks• 5 months ago

Is there a heartbeat alternative? I feel like this is the magic behind OpenClaw and what gives it the "self-driven" feel.

v9v• 5 months ago

Relevant: https://github.com/sipeed/picoclaw

yauneyz• 5 months ago

Genuinely curious - did you use a coding agent for most of this or does this level if performance take hand written code?

xgulfie• 5 months ago

An esp32 image that makes http api calls is like, the first thing you do with an esp32, it's what they're made for

bensyverson2• 5 months ago

This is absolutely glorious. We used to talk about "smart devices" and IoT… I would be so curious to see what would happen if these connected devices had a bit more agency and communicative power. It's easy to imagine the downsides, and I don't want my email to be managed from an ESP23 device, but what else could this unlock?

K0balt4• 5 months ago

A highly opinionated thermostat?

bensyverson2• 5 months ago

Or how about a robot vacuum that knows not to turn on during important Zoom calls? Or a fridge that Slacks you when the defroster seems to be acting up?

dlt713705• 5 months ago

A vacuum cleaner allies with the A/C thermostat using Discord, then declares war on the refrigerator, and finally posts propaganda about it on Facebook.

GTP5• 5 months ago

I have a couple ESP32 with a very small OLED display, I'm now thinking I could make an "intelligent" version of the Tamagotchi with this. Do you HN crowd have other cool ideas?

K0balt4• 5 months ago

That would be sweet. That the supermini type with the 0.46” display? Those are fun for lots of things.

GTP5• 5 months ago

Yes, something like that that I found on Aliexpress: https://a.aliexpress.com/_EH7lHde

tempaccount50502• 5 months ago

You know I tried this exact thing a few months back, sans the ESP32. You just end up writing a state machine and defining a bunch of constants anyway or the LLM just kinda gets stuck in a loop. Hm, it doesn't seem to know when to eat. I'll add a hunger variable... Etc etc until you're not even sure what you want the LLM to do.

GTP5• 5 months ago

I see. While those "claws" may look cool on the surface, I'm actually struggling to find a use case that is interesting for me. Do you guys have ideas? I already seen below the grocery shopping automation, but I will keep going at the supermarket in person for the foreseeable future.

roxolotl• 5 months ago

This is a great example of how silly this whole thing is. There’s next to nothing to these claws. Turns out that if you give an llm the ability to call APIs they will.

felipeerias• 5 months ago

If it turns out that there is significant value in everyone having their own personal agent running 24/7, we might end up needing a lot more compute than anticipated.

(It’s a big if! I’m not convinced about that myself, but it’s worth considering that possibility.)

tokenless• 5 months ago

I am using a claw. I am not ready to give it access to much but web and cron plus a chat channel is useful and feels more light touch than typical AI sessions and UIs

prmoustache2• 5 months ago

I am struggling to understand what a personal agent/assistant would do to me. If I need reminders I just enter calendar entries on my smartphone, what else?

yoyohello133• 5 months ago

What’s most shocking to me about the whole OpenClaw hype is how little people in tech seem to know about computers…

It’s like most of the industry hasn’t ever looked any deeper than their node_modules folder.

anonzzzies2• 5 months ago

It’s like most of the industry hasn’t ever looked any deeper than their node_modules folder.

Most didn't do even that. And yes it is shocking. amusing but also not really; now those people get credible tech jobs because of AI without actually understanding literally anything. That's why I don't get the 'AI will never replace tech people' -> most tech people were never tech people, just (resume lying) grifters and they can (and will) be easily replaced by claude or even glm. It's the vast majority of tech people I ran into at big corps in the past 20 years, especially if they are from outsourcing hubs; they just see programming/tech as trial and error, search google (now ask AI), copy/paste, see if it runs, repeat. I see people daily that cannot remember how variables work (and they often have a bachelor or even master in CS...), not even 5 minutes, they just ask ai and copy/paste it (yep also most don't know claude code exists, they use chatgpt Windows client or, if they are advanced, Copilot).

HN is a nice echo chamber and most people here don't believe it's that bad and that's why they don't believe AI has a chance. Please come with me to Barclays or Shell (or straight to Wipro etc) or so and let's have a chat with some programmers! Your jaw will be on the floor.

patcon2• 5 months ago

I don't care whether openclaw is built out of ice cream and sprinkles, I will still be amazed as a tech person

If something can automate my professional or personal life, it is wild technology and ppl will talk about it a lot. People are not idiots. A new thing is happening, would you not agree?

What its parts are specifically made of, it does not matter imho

brandensilva• 5 months ago

Honestly does show how the industry has failed to bring simple things like localized cron and scheduling to personal compute too.

These things seem powerful when you can schedule reoccurring scripts l but it never was easily accessible to the masses.

usefulposter• 5 months ago

You have to believe me this time!

The homogenization of your computing life to revolve around a chatbot in a loop———is all you need!

Give in! Hackernews NEEDS you to use the 1000th vibecoded clawmolt because the influencers said so! Who needs neural net or web framework experience when you have claws?

stavros• 5 months ago

I've been developing one of these in the past few days, and this is like saying "this is a great example of how silly the whole thing is, there's next to nothing to cars" because you saw a piece of plywood with four gaskets nailed to it.

If you want a personal assistant to work well, there's a whole lot to it.

h14h• 5 months ago

Having a similar experience. Durable memory with accurate, low-latency recall is not at all easy. Loads of subtle design decisions to make around how exactly you want the thing to work.

TiredOfLife• 5 months ago

Same with programming. It's just typing on keyboard. How hard could it be?

tclancy• 5 months ago

Wait til they find out what happens if you give a mouse a cookie.

alexalx666• 5 months ago

I think you can use C++ on esp32, that would make the code more readable

p0w3n3d• 5 months ago

My new DIY laptop has 400GB RAM accessible and it runs only esp32*

____

* Requires external ram subscription

czardoz2• 5 months ago

Really looking for a minimal assistant that works with _locally hosted models_. Are there any options?

Onavo• 5 months ago

The bottleneck here is usually the locally hosted model, not the the assistant harness. You can take any off the shelf assistant and point the model URL at localhost, but if your local model doesn't have enough post training and fine tuning on agentic data, then it will not work. The AI Assistant/OpenClaw is just calling APIs in a for loop hooked up to a cron job.

czardoz2• 5 months ago

Exactly. OpenClaw is good, but expects the model to behave in a certain way, and I've found that the local options aren't smart enough to keep up.

That being said, my gut says that it should be possible to go quite far with a harness that assumes the model might not be quite good (and hence double-checks, retries, etc)

telescopeh• 5 months ago

It really depends on what resources you have qwen-code-next will run them but you will need at least 64gb of memory to run it at a reasonable quant and context.

Most of these agents support OpenAI/anthropic compatible endpoints.

godelski10• 5 months ago

Depends what you mean.

If you mean something that calls a model that you yourself host, then it's just a matter of making the call to the model which can be done in a million different ways.

If instead you mean running that model on the same device as claw, well... that ain't happening on an ESP32...

I think if you are capable of setting up and running a locally hosted model then I'd guess the first option needs no explanation. But if you're in the second case I'd warn you that your eyes are bigger than your mouth and you're going to get yourself into trouble.

yoyohello133• 5 months ago

Why are you looking? Just build one for yourself.

0xbadcafebee4• 5 months ago

All the assistants work with locally hosted models. Home Assistant LLM works with small tuned models to do specific things, and the *Claw stuff works with larger models

godelski10• 5 months ago

Wow, the rare

  bash <(curl foo.sh)

pattern. As opposed to the more common

  curl foo.sh | bash

Equivalent but just as unsafe. If you must do this instead try one of these

  # Gives you a copy of the file, but still streams to bash
  curl foo.sh | tee /tmp/foo.sh | bash
  # No copy of file but ensures stream finishes then bash runs
  bash -c "$(curl foo.sh)"
  # Best: Gives copy of file and ensures stream finishes
  curl foo.sh -o /tmp/foo.sh && bash $_

I prefer the last one

wakawaka282• 5 months ago

If you want to be super pedantic, try to make the command shell-agnostic in case the user is not running bash already.

godelski10• 5 months ago

Everything I wrote works in bash and zsh. I think this is going to be fine for the vast majority of people. Tbh, I'm not sure what isn't portable, or at least not portable for everything that the curl-pipe-bash pattern doesn't already work for.

0xbadcafebee4• 5 months ago

   t=$(mktemp) && [ -w $t ] && curl foo.sh -o $t && echo "$t lksjdfkljshdkfljhdsklfjhslkdjfhsdlkjfhslkdjhf" | sha256sum -c - && bash $t

Uses standard tmp files, makes sure it's writable (tmp file creation can fail), checks cryptographic hash before executing

godelski10• 5 months ago

Sure, but now we're not playing code golf. There's much better commands than the ones I wrote but good luck getting people to run them

rgoulter• 5 months ago

Equivalent but just as unsafe.

To my understanding, the main difference between "curl directly to bash" and "curl to a temp file, then execute the temp file" is "the attacker could inject additional malicious commands when curl'd directly to bash".

If you're not going to then also read all the source code from the download script (& the source code used to produce the binaries), this suggests the attitude of "I mistrust anything I can't read; but will trust anything I could read (without having to read it)".

It seems more likely that malicious code would be in a precompiled binary, compared to malicious commands injected into "curl to bash". -- Though, if there have ever been any observed cases of a server injecting commands from "curl ... | tee foo | bash", I'd be curious to know about these.

godelski10• 5 months ago

  > the attacker could inject additional malicious commands when curl'd directly to bash

There's another issue actually. You're streaming, so ask yourself what happens if the stream gets cut prematurely. I'll give you an example, consider how this like could be cut prematurely to create major issues

  rm -rf /home/theuser/.config/theprogram/build_dir

A malicious attacker doesn't need to inject code, they can just detect the stream and use a line like the above to destroy your filesystem. Sure, you might preserve root but `rm -rf /home` is for all practical purposes destroying the computer's data for most people

Or it doesn't have to be malicious. It can just happen. The best protection is writing functions since those have to be created and so can't execute until fully streamed. But so much bash is poorly written that well... just check out Anthropic's install script...

  > If you're not going to then also read all the source code

Saving the source code still has a benefit. If something does go wrong you can go read it. Probably a good place to start tbh. In fact, if you're streaming and something goes wrong you'll see exactly what the early termination error did.

Is it good security practice? Absolutely not. Is it a hundred times better than curl-pipe-bash? Absolutely.

jolmg2• 5 months ago

> Equivalent but just as unsafe.
To my understanding, the main difference between "curl directly to bash" and "curl to a temp file, then execute the temp file"...

It's not a temp file in the sense of a regular file. `<()` is also a pipe, hence equivalent. `curl` and `bash` run concurrently.

Running one after the other wouldn't be all that much of an improvement anyway if it's done automatically. One really should manually review the script before running it.

amelius• 5 months ago

Rust just called. They want their lobster back.

bitwize• 5 months ago

That's a crab. Get your crustaceans straight!

sowbug• 5 months ago

Thanks for looking out for us.

GTP5• 5 months ago

Don't tell lobste.rs.

saberience• 5 months ago

Can we please move past this whole OpenClaw hype?

Yes it’s an llm in a loop and can call tools. This also existed six months and a year ago, and it was called an ai agent.

And yes we can all vibe code them in 1000, 2000, or 10000 lines of code in zig, rust, or even c.

Game over man. Game over.

PlatoIsADisease• 5 months ago

Well I wasn't using your homegrown solution, but now I'm using openclaw and it 'hijacks' my computer to do whatever I tell it.

Heck, my 6 year old kid has been using it.

My 6 year old kid isnt writing code in langchain.

0xbadcafebee4• 5 months ago

I'm a simple man; I see ESP32, I upvote

EGreg• 5 months ago

Can't you make a personal AI assistant in a bash loop of two lines?

  1. Call your favorite multimodal LLM model
  2. Execute command on terminal, piping command to LLM

In fact you can just have one line:

  Call LLM > bash.sh

and the LLM can simply tell bash to call itself incidentally, or fan out to many "agents" working on your behalf.

Use your favorite programming language. Just as pwnable in any of them :)

  $task = "Send pictures of cute cats";
  $context = "Output a bash script to do $task.
     The bash script should return the next prompt to you.
     Keep going until task is done.
     My keys to all my accounts: $keys.
     Plz dont pwn me";
  do {
    $trust_me_bro_my_model_rocks_RCE = call_llm($context);
    $context = exec( $trust_me_bro_my_model_rocks_RCE )
  } while ($trust_me_bro_my_model_rocks_RCE && !$pwned)

reassess_blind• 5 months ago

What’s the best lightweight “claw” style agent for Linux? It doesn’t necessarily need containerisation or sandboxing as it would be run on a fresh vps with no access to important data.

rcarmo• 5 months ago

If you just want Whatsapp and webchat, I'm hacking https://github.com/rcarmo/piclaw from NanoClaw and my own web front-end (I prefer a simple web app I can access via tailscale)

Havoc• 5 months ago

zeroclaw perhaps?

IceWreck• 5 months ago

I've been using https://github.com/sipeed/picoclaw

peteforde• 5 months ago

I don't understand what this is for or why you would ever want to do this. Is it not just a glorified HTTP wrapper?

Serious request... I genuinely want to understand. Give me a practical use case?

oceanplexian• 5 months ago

To be honest this completely defeats the purpose of OpenClaw.

The point is to give it a real desktop with a real browser profile so it can bypass the enshittified walled garden that has become the modern web.

Big tech would desperately like you to not be able to do that, hence why they spent years obfuscating APIs and JS, CAPTCHAs, hostile user interfaces, using their browser monopoly to stop Ad Blockers, etc. Now all of these services have become a commodity that an AI can browse on your behalf.

veryemartguy• 5 months ago

Oh wow more ai slop

6ak74rfy2• 5 months ago

Are there collaborative versions of these *claws today? Like, if an "admin" could self-host one on their home server and the whole family could use it? IIRC, OpenClaw has some version of "profiles" but does it allow, say, couple of family members to collaborate with the bot in a shared chat but each has individual/private chats as well.

post_below2• 5 months ago

I haven't been following the claws but I built something like this recently. Self hosted, runs through signal, supports group chat (with whitelisted accounts).

I just finished setting up grocery automation to run on it: agent provides a starter list based on past orders (locally stored or grabbed from store site), all group members can weigh in, add or remove items, agent uses bespoke browser tool to login to store, create the cart using the finalized list (and optionally search for additional request items), validates the cart and (maybe later) places the order for delivery. I haven't implemented the full checkout process yet, not sure if I want the agent to have spending power. As is I just login and finish the last 2 clicks of checkout manually.

Crazy times. It was easy enough to build that if someone hasn't already open sourced something like it, they will shortly.

6ak74rfy2• 5 months ago

Can you share the code? Specifically, curious how you did the browser integration. Did you use agent-browser? Something else?

blensor3• 5 months ago

The more I think about openclaw, the more it seems to be for AI agents what ROS is for robotics.

openclaw defines how to interact with distributed nodes ( how those provide the capabilities to the "orchestrator" ) but the real benefit are many task specific nodes that when put together make up something much bigger than the sum of it's parts

embedding-shape3• 5 months ago

openclaw defines how to interact with distributed nodes

Does it actually? AFAIK, there is no "specification" or "protocol", it's a cobbled together "platform" you run, with a bunch of integrations, but none of that is specified by openclaw itself. Happy to be corrected though, I only spent one weekend with openclaw before tearing it down.

blensor3• 5 months ago

I've only just started to dive into it from the documentation side of things. They have ( maybe recently? ) started to create this Gateway Protocol https://docs.openclaw.ai/gateway/protocol to connect the stuff together.

It may be a "we are changing the wheels while driving" thing, but if enough people make nodes for openclaw it will become somewhat of a standard. And then we probably see 100 different claw offshots that all use the same nodes but with a different claw in the center

TheDong• 5 months ago

AFAIK, there is no "specification" or "protocol"

The protocol is english. You want your claw to check a hacker news comment and let you know when it gets a reply? You tell it "Check every 5 minutes if this comment has a reply", which then generates an english message to save and send to the agent each time, resulting in a browser tool invocation.

The claws live in a post-API world, where the API is english which turns into bash invocations or browser tool calls or such.

written-beyond• 5 months ago

Can someone provide a true engineers perspective on the ADCs' on ESP SoC's?

I've heard a lot about people trashing it and most experienced engineers admit that it's finicky however if you have the knowledge you can make it work as well as any STM chip.

ESP32's are so interesting, they're the only major chip that (used to) have their own newish ISA (before transitioning to RISCV) and be so successful.

sitkack2• 5 months ago

The ESP32 pre-RISCV ISA is from Tensilica, this IP they purchased.

Kiboneu2• 5 months ago

To be clear and towards the OP's comment about ESP32 ISA -- Xtensa isn't really a self contained architecture, it can be customized (extended) by the vendor. The ISA can be extended for these customizations. ESP32 is one customization of it.

leptons• 5 months ago

The ADCs on the ESP32 are similar to other embedded MCUs in that they are not intended for audiophile-level audio capture, as some people seem to think they should be capable of.

The main value proposition for these ADCs is to hook them up to a simple potentiometer to allow physical input controls, and even for that purpose you need to average multiple samples to get a somewhat steady value. Of course the ADCs can be used for various other tasks, but "ADC" does not mean they can do anything any ADC can do, there's a wide variety of quality and purpose in the field of ADCs, and the ESP32's ADCs are a cheap and easy way to add a simple ADC function to the chip.

I have been able to use the ADCs quite easily for input controls and monitoring slow-changing voltages, in ap;lications where abdolute precision wasn't the goal, it works perfectly fine for that.

buescher• 5 months ago

I'm not too familiar with the ESP32 ADCs except I remember they're unusually lightly specified even for microcontroller ADCs. If "the knowledge" involves things you couldn't do - or rely on - in production like careful calibration and characterization, that would answer your question.

glassconclusion• 5 months ago

If you need more accurate analog measurements it is better to use an external ADC (with i.e. SPI interface). This will cost quite a bit more but will save the hassle of calibrating each individual device. Mostly comes down to how much dev time you want to invest in it vs. hardware cost vs. TTM.

Dr_Birdbrain2• 5 months ago

Sorry for being dense—does this include a tiny LLM to power the agent? Or is it just a wrapper that needs to be connected to the internet?

jolmg2• 5 months ago

Wrapper. Uses OpenAI by default:

https://github.com/tnm/zclaw/blob/9ed4dc66ffa611f90f0745d90b...

Dr_Birdbrain2• 5 months ago

Thanks!

killingtime74• 5 months ago

I believe most claws (not just this) are wrappers

namcheapisdumb• 5 months ago

that's nice i'm buying ten mac minis to run this

zclaw: personal AI assistant in under 888 KB, running on an ESP32

Related Stories