February 21, 2026
Microchip, macro meltdowns
zclaw: personal AI assistant in under 888 KB, running on an ESP32
Tiny chip assistant triggers big arguments: cloud vs local and “do we even need it”
TLDR: zclaw packs an AI assistant into under 888 KB on a tiny ESP32 board but relies on cloud models. Comments split between privacy-focused calls for local AI, skeptical “why?” energy, and a self-driven “heartbeat” wish list, with [picoclaw](https://github.com/sipeed/picoclaw) floated as an alternative—small device, big debate.
Meet zclaw: a pint‑sized “personal AI assistant” squeezed under 888 KB on a cheap ESP32 microcontroller (think tiny hobby board, not a laptop). It chats via Telegram or a web page, can set schedules, control pins to toggle gadgets, and remembers things even after a reboot. Cool party trick. But the comments turned it into a reality show.
The loudest reaction? A shrug. One early voice dropped, “I don’t really need any assistance…” while another sighed that the brain isn’t on the device at all: zclaw taps cloud AI providers like Anthropic, OpenAI, and OpenRouter, which ignited the local vs cloud flame war—privacy, latency, and “is it still tiny if the mind lives elsewhere?” Then came the practical crowd asking, “Serious question: why?” pushing for real workflows beyond the nerd flex of fitting it under a megabyte. A wonky subplot emerged around a “heartbeat” loop—fans want a self-driven assistant that acts on its own like OpenClaw—while someone dropped a “Relevant:” link to picoclaw, turning the thread into a lightning round of alternatives.
Jokes flew about the 888 KB flex and the idea of every appliance needing cloud keys. Enthusiasts imagine plant watering, garage doors, and reminder bots; skeptics call it a cloud chatbot on a chip. Verdict: tiny firmware, huge drama, and a community obsessed with whether small should also mean local.
Key Points
- •zclaw is a C-based AI assistant firmware for ESP32 targeting a ≤888 KB default build.
- •Features include timezone-aware scheduling, GPIO control with guardrails, persistent memory, and natural-language tool composition.
- •Interaction is supported via Telegram or a hosted web relay, with LLM providers Anthropic, OpenAI, and OpenRouter.
- •Tested on ESP32-C3/S3/C6, with Seeed XIAO ESP32-C3 recommended; other ESP32 variants may need manual ESP-IDF setup.
- •Comprehensive scripts cover build, flash (including secure), provisioning, monitoring, QEMU emulation, web relay, benchmarking, docs, and tests; licensed MIT.