A developer has built a tiny AI coding agent that runs entirely offline on an ESP32 microcontroller with just 512 kB of RAM, letting users issue plain-English commands to control hardware and fetch live data without cloud dependencies.

What happened: A PySpell system runs on an ESP32 chip, accepting English-language instructions that a ~0.45 M-parameter language model converts into code, which then executes in a sandboxed environment on the device itself. The model, tokenizer, and runtime are all served from the chip—no cloud, no API key required—and can drive a screen, LED, and make HTTP requests to allowlisted hosts.

Why it matters: Developers can now prototype and deploy AI-driven automation on tiny, resource-constrained devices without internet dependency or third-party infrastructure. The approach demonstrates that meaningful inference is possible within severe hardware limits by combining a small model with clever engineering (client-side inference in WebAssembly, frozen embeddings, and semantic-directive architecture rather than token copying).

What to watch: The system supports up to 8 parallel PySpell processes on the same half-megabyte of RAM and is accessible over Tailscale, making it deployable across a network of devices. The developer has published a retraining pipeline so others can adapt the model for different languages by translating instruction phrasings and swapping the embedding model.