zclaw: A Personal AI Assistant in Under 888 KB, Running on an ESP32
Original: zclaw: personal AI assistant in under 888 KB, running on an ESP32 View original →
AI on a Microcontroller
zclaw, shared on Hacker News with 161 upvotes, is a personal AI assistant that runs on an ESP32 microcontroller in under 888 KB. The ESP32 is a $5 microcontroller with a 240MHz dual-core processor and roughly 520 KB of RAM — far below what most people imagine as the minimum for running AI.
Part of the Claw Ecosystem
zclaw is one of the emerging implementations in the "Claw" category — the term Andrej Karpathy recently coined for a new generation of AI agent orchestration systems. While most Claw implementations run on personal computers or servers, zclaw pushes the concept to its most extreme form: an embedded microcontroller.
- Total footprint under 888 KB
- Runs on ESP32 (240MHz dual-core, ~520 KB RAM)
- Personal AI assistant with Claw architecture
- Open source on GitHub
Edge AI's New Frontier
As interest grows in reducing cloud dependency and preserving privacy through edge computing, zclaw demonstrates an extreme example of what's possible. Running an AI assistant on a few-dollar microcontroller points toward a future where AI is embedded in virtually any device.
The project joins a growing ecosystem of lightweight Claw implementations, from NanoClaw to nanobot, that collectively form a decentralized AI agent landscape.
Related Articles
Long-document OCR is bottlenecked by page chunking and growing KV cache. A widely shared post says Baidu’s Unlimited-OCR uses 3B total parameters, 500M active parameters, and a 32K context window to read 40-page documents in one pass.
NVIDIA’s SIGGRAPH update shifts physical AI from cloud demos toward edge deployment. The package includes the 4B Cosmos 3 Edge world model, a Synthetic Video Detector NIM microservice, and a DGX Station agent stack built around Nemotron 3 Ultra.
A March 19, 2026 Hacker News post about Kitten TTS reached 512 points and 172 comments at crawl time. KittenML says its 15M, 40M, and 80M ONNX speech models target CPU inference with eight English voices and 24 kHz output.