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

Original Article ↗ Hacker News Discussion ↗

Binary Size & ESP32 Constraints

  • Some are surprised 888 KB is needed for “just an HTTP wrapper”; comparisons are made to the original Doom binary size.
  • Others point out the ESP32 image includes WiFi stack, TLS, and certificates that an OS would normally provide; app logic itself is described as 25 KB, with WiFi (350 KB), TLS (120 KB), and certs (90 KB).
  • General sentiment that ESP32 libraries (WiFi/Bluetooth/HTTP) are quite bloated.

Local vs Cloud Inference

  • Several commenters hoped for fully local inference; consensus is that meaningful LLM-level capability in 888 KB on ESP32 is currently unrealistic.
  • Tiny LMs (hundreds of thousands of parameters) can run, but are more “babblers” than useful assistants.
  • Debate over theoretical future possibilities (e.g., very compact symbolic/fuzzy systems) remains speculative and unresolved.

What Claws/Agents Actually Do

  • Claw-style agents are described as pipes between LLM APIs and other APIs/CLIs (email, social media, messaging, cron jobs).
  • Example workflows: polling email, generating replies, posting on social networks, grocery-order automation via browser tools.
  • Some see them as trivial “LLM in a loop with tool calls”; others argue building a robust assistant (durable memory, reliable behavior) is non-trivial.

Use Cases, Hype, and Skepticism

  • Supporters view this as a powerful new automation layer, including for non-programmers and families (e.g., shared household agents).
  • Critics say the tech is overhyped, conceptually old (agents), and often poorly engineered (“vibe coded”), with serious security and reliability risks if given broad access.
  • There’s concern many developers don’t understand underlying systems, making agent misuse likely.

ESP32 as Agent Host

  • Proponents highlight ESP32’s always-on, low-maintenance nature versus maintaining a full Linux box.
  • Others counter that relying on cloud APIs adds more failure modes (internet, WiFi, remote service).
  • Some argue the real promise of OpenClaw-style setups is running on a “real desktop” with a real browser to navigate today’s hostile, locked-down web.

Local Models & Minimal Assistants

  • People ask for minimal assistants that work with locally hosted models; responses emphasize that the limiting factor is model quality, not the harness.
  • Many agent frameworks already support pointing at local OpenAI-compatible endpoints, but most small local models struggle with the behavior patterns OpenClaw-style agents expect.

Security & Installation Practices

  • Discussion around curl | bash vs safer patterns (download to temp file, verify, then execute); direct streaming is seen as fragile and risky (e.g., truncated streams causing destructive partial commands).

Miscellaneous

  • Numerous side ideas: synthesizers, desk rovers, Tamagotchi-like pets on ESP32 + OLED, “intelligent” thermostats and cleaning robots.
  • A technical question about ESP32 ADC quality vs STM chips is raised but not really answered in the thread (status: unclear).