Zclaw – The 888 KiB Assistant

“888 KiB Assistant” but the assistant itself is a multi terabyte rental-only model stored in some mysterious data center.

For people who don't get this: it's a Home Assistant type thing. You don't do inference on it, you send it a message on Telegram and it does things with physical things through GPIO. You could use a $140 Raspberry Pi with 8GB RAM and host a local model on it plugged into 30W AC power... or you could use a $10 ESP32 which can run for weeks on a tiny battery, and your existing Wifi connection with a cheap cloud model (cloud models are as cheap as $0.02/1MTokens). This makes it easier to ramp up on new ESP32 projects. You can just tell it to do things / give you info, rather than having to write code.

Me: "GPIO 5 can be active for a maximum of 100ms, then it needs to cool down for at least 1s. Otherwise the MOSFET is fried."

Zclaw: "GPIO 5 is active now, however the server is not responding so I'm awaiting further instructions."

I fail to understand why 888 KiB matters if it's just a wrapper around a cloud api.

888KiB is quite large, but I see they're including the whole rest of the firmware in that number, fair enough. Their actual application code weighs only 35,742 bytes, compiled.

There are many concerns and areas for improvement with open claw and other similar projects (continuous loop script with broad OS access that manages your agents and interfaces with a standard messaging app)

However, file size I have never seen on that list. I would rather offer for something that is even bigger in file size so it afford certain functionality like better security tighter permissions however it would do that.

Clicked on this expecting to see a crontab file.

Still lightyears from a one-bit AI assistant. Send 1 to save the humankind, 0 to exterminate. And hurry up because it's in undefined state right now!

The domain crashed and burned or something, hopefully this link is correct:

https://github.com/tnm/zclaw

There is the same divide starting to form that NFTs had back in the day. Tech bros instantly like if something has claw in the name, the rest of us will dismiss anything with that naming and philosophy as toxic slop culture. will be interesting to see how far this one will go.

“888 KiB Assistant” but the assistant itself is a multi terabyte rental-only model stored in some mysterious data center.

I'm getting "serverless" flashbacks.

It seems to support connecting to your own LLM on the same LAN

My model is at home... just 16Gb still a lot but just FYI

It seems to support connecting to your own LLM on the same LAN

The point is the agent is still the LLM. No LLM, no agent.

I tried connecting OpenClaw to ollama with a V100 running qwen3.5:35b but it was really, really, really slow (despite ollama itself feeling fairly fast).

These "claw" agents really multiply the tokens used by an obscenely huge factor for the same request.

Me: "GPIO 5 can be active for a maximum of 100ms, then it needs to cool down for at least 1s. Otherwise the MOSFET is fried."

Zclaw: "GPIO 5 is active now, however the server is not responding so I'm awaiting further instructions."

Clicked on this expecting to see a crontab file.

The domain crashed and burned or something, hopefully this link is correct:

https://github.com/tnm/zclaw

888KiB is quite large, but I see they're including the whole rest of the firmware in that number, fair enough. Their actual application code weighs only 35,742 bytes, compiled.

dope so i can hook it up to my appliances, hvac, etc? maybe it'll do rad things with my smart doorlocks, remote car starter. can it send my self-driving waymobile to pick up my family at the airport too? such wow

The model can be taken care of in cloud but hardware also depends upon what we want to do right? If we want to run some lightweight python scripts etc, we cant use ESP32 right?

I made a secure one:

https://github.com/skorokithakis/stavrobot

Everything runs in containers (I run it on a server along with everything else), plugins have a permission system so eg the AI can read emails but not delete or send, etc.

I really like it, I run it as my main agent and it has been extremely helpful.

File size is a legit property to keep in mind if your goal is to create an agent that runs on ESP32 boards. They don't expect you to run Zclaw on Mac Mini.

Still lightyears from a one-bit AI assistant. Send 1 to save the humankind, 0 to exterminate. And hurry up because it's in undefined state right now!

I fail to understand why 888 KiB matters if it's just a wrapper around a cloud api.

I'm getting "serverless" flashbacks.

My model is at home... just 16Gb still a lot but just FYI

Have you seen OpenClaw's codebase? 680.000 LOC.

I care how big it is.

A lot of the *claws emphasize binary size and lines of code. I think for better or worse people treat codebase size as a proxy for "how much of the project is unsupervised, unmaintainable, buggy AI slop?"

Because of resource-constrained environments, the primary deployment target seem to be microcontrollers. You can get ESP32 boards for pretty cheap.

8 is lucky number in China

Because it means you can run it on an ESP32 which is a low power microprocessor package.

Is Clawcoin a thing yet?

I tried connecting OpenClaw to ollama with a V100 running qwen3.5:35b but it was really, really, really slow (despite ollama itself feeling fairly fast).

These "claw" agents really multiply the tokens used by an obscenely huge factor for the same request.

i recently decided to get into this ocean boiling game too, the 32GB V100 seems like a pretty good VRAM/$. if i may ask, do you make any special accommodations for cooling? i've never dealt with a passively cooled card before and i'm curious whether my workstation fans (HP Z840) will be sufficient. i'm going to try 2 cards at first but i think i might be able to squeeze a third in there

The model can be taken care of in cloud but hardware also depends upon what we want to do right? If we want to run some lightweight python scripts etc, we cant use ESP32 right?

The point is the agent is still the LLM. No LLM, no agent.

Have you seen OpenClaw's codebase? 680.000 LOC.

I care how big it is.

File size is a legit property to keep in mind if your goal is to create an agent that runs on ESP32 boards. They don't expect you to run Zclaw on Mac Mini.

What's the use case for running this on a tiny board? Isn't the whole point that it can use your computer for you?

I made a secure one:

https://github.com/skorokithakis/stavrobot

Everything runs in containers (I run it on a server along with everything else), plugins have a permission system so eg the AI can read emails but not delete or send, etc.

I really like it, I run it as my main agent and it has been extremely helpful.

Part of the usefulness is based on the same thing that makes it so dangerous.

If it can only read but not act, it’s safer but less useful.

Because of resource-constrained environments, the primary deployment target seem to be microcontrollers. You can get ESP32 boards for pretty cheap.

8 is lucky number in China

Because it means you can run it on an ESP32 which is a low power microprocessor package.

Is Clawcoin a thing yet?

What's the use case for running this on a tiny board? Isn't the whole point that it can use your computer for you?

For something like OpenClaw yes, but not for Zclaw. I think the naming is more about riding the current wave of Claw-related interest rather than positioning it as competition or replacement for other clawies.

Zclaw is about running an agent in your embedded system.

The examples seem to suggest it would be chatting with your home automation in natural language.

Part of the usefulness is based on the same thing that makes it so dangerous.

If it can only read but not act, it’s safer but less useful.

I can't restrict OpenClaw if I don't need the extra capabilities. I can restrict this.

Zclaw is about running an agent in your embedded system.

I can't restrict OpenClaw if I don't need the extra capabilities. I can restrict this.

The examples seem to suggest it would be chatting with your home automation in natural language.

Before you know it your smart thermostat will be blogging. The joke is on everyone who thought IoT couldn't get any worse. Just imagine the new landscape of security vulnerabilities this opens up.

My "smart" gas stove can be turned on over the internet (if I allow it to connect)—perfect appliance to put an LLM in charge of.

zclaw docs

chapter 0

The 888 KiB Assistant

zclaw is an ESP32-resident AI agent written in C. It runs as a practical assistant over Telegram or host relay, with scheduling, GPIO control, memory, and a tight firmware budget.

Enjoy with zclaw

"Remind me in 20 minutes."
"Water the plants every day at 8:15."
"Set GPIO 5 high."
"Remember that my office sensor is on GPIO 4."

You send plain language, zclaw maps to tool calls, firmware executes on silicon.

You: In 20 minutes, check the garage sensor Agent: Created schedule #7: once in 20 min -> check the garage sensor

Lobster soldering a Seeed Studio XIAO ESP32-C3

Tested targets: ESP32-C3, ESP32-S3, and ESP32-C6. Other ESP32 variants should work fine.

What "888 KiB" Means

The 888 KiB target is an all-in firmware cap, not just zclaw application logic. It includes app code plus ESP-IDF/FreeRTOS runtime, Wi-Fi/networking, TLS/crypto, and cert bundle overhead.

Current default esp32s3 build (grouped loadable image bytes from idf.py -B build size-components; rows sum to total image size):

Layer	Size	Share
zclaw app logic (`libmain.a`)	35,742 bytes (~34.9 KiB)	~4.1%
Wi-Fi + networking stack	397,356 bytes (~388.0 KiB)	~45.7%
TLS/crypto stack	112,922 bytes (~110.3 KiB)	~13.0%
Cert bundle + app metadata	99,722 bytes (~97.4 KiB)	~11.5%
Other ESP-IDF/runtime/drivers/libc	224,096 bytes (~218.8 KiB)	~25.8%

Total image size from this build is 869,838 bytes; padded zclaw.bin is 869,952 bytes (~849.6 KiB), under the 888 KiB cap.

Read This Manual In Order

Chapter 1 · Getting Started Bootstrap install, flash, provision, and first successful boot. Chapter 2 · Tool Surface Current built-in tools and scheduling behavior, including one-shot jobs. Chapter 3 · Runtime Anatomy Task model, queues, LLM path, and practical constraints. Chapter 4 · Security & Ops Safety defaults, flash encryption, and production handling guidance. Chapter 5 · Build Your Own Tool Design, create, validate, and maintain custom natural-language tools. Chapter 6 · Local Dev & Hacking Practical local iteration loops, provisioning profiles, and debug workflows. Chapter 7 · Use Cases Useful and playful ideas that are only practical when the assistant lives on your device. Chapter 8 · Changelog Release notes and a timeline of what changed across versions.

Project Character

Language/runtime: C + ESP-IDF + FreeRTOS.
LLM backends: Anthropic, OpenAI, OpenRouter, Ollama (custom endpoint).
Interface: Telegram and optional host web relay.
Philosophy: ship useful automation under strict resource bounds.