Index / zclaw · updated Oct 2, 2026
zclaw
tnm/zclaw · healthy · rank 21 of 43 by stars
An ultra-minimalist AI personal assistant that squeezes onto an ESP32 microcontroller with an all-in firmware budget of 888 KiB. It trades heavyweight agent frameworks for bare-metal C, GPIO control, cron scheduling, and Telegram/web relay chat.
Facts
Repository
Runtime
Posture
Community
Security breakdown
Composite 72 / 100 · how these are scored
higher is safer
higher is safer
higher is safer
higher is safer
higher is riskier
Model access
Read from the repository, not written by a model · 57 files examined
2 providers · runs locally · gateway support
Pinned models
claude-sonnet-4-6 released 2026-02-17 — 8 months old (from the public model catalogue)
Pin last edited 7 months ago
Several defaults are pinned; the project picks one per provider rather than shipping a single default.
Evidence
Decision
Why choose zclaw over OpenClaw?
Why choose this
- Runs on a $5 ESP32 with an 888 KiB all-in firmware budget
- Direct GPIO, I2C, and DHT sensor control via natural language tools
- No server, Node runtime, or heavy dependencies required
Tradeoffs
- Far smaller ecosystem and no multi-channel/agent routing depth
- Requires embedded toolchain setup and hardware flashing
- Limited LLM throughput due to rate limits and device constraints
Best fit
- Makers wanting an AI assistant on ESP32 hardware
- Developers who want GPIO/sensor control via natural language
- Users who value extreme footprint constraints over feature breadth
Avoid if
- You need a rich plugin ecosystem or multi-agent routing
- You want a desktop/cloud assistant without flashing firmware
- You need strong community support and third-party integrations
Strong evidence from README, docs, and recent commits for architecture and features; community sentiment is uncertain due to zero Reddit matches and only indirect web mentions.
AI layer reviewed Sep 21, 2026 · how this is written
Star activity
2,229 stars today
Overview
zclaw is a radically minimal take on the personal AI assistant: instead of a Node.js or Python runtime, it is written in C and targets ESP32 microcontrollers with a strict all-in firmware budget of 888 KiB — including the ESP-IDF/FreeRTOS runtime, Wi-Fi, TLS, and certificate bundles, with only ~35 KB of actual application code. Despite the tiny footprint, it supports scheduled tasks (daily, periodic, one-shot), persistent memory, GPIO control, and user-defined tools composed through natural language.
Users interact with zclaw via Telegram or a hosted web relay, and credentials (Wi-Fi, LLM API keys, Telegram allowlists) are provisioned via scripts and can be stored encrypted in flash using a secure flash mode. It supports both cloud LLM backends and local Ollama endpoints, with compile-time rate limits (default 100/hour, 1000/day). Recent releases added hardware-oriented tools like DHT sensors and generic I2C, reinforcing its identity as a bridge between LLM agents and physical computing.
Compared to OpenClaw, zclaw sacrifices the rich multi-channel routing, dashboards, and plugin ecosystems for extreme portability and hackability. It is best understood as the embedded edge node of the Claw ecosystem: a fun, hackable assistant you can solder onto a board rather than a full-featured agent platform.