Key takeaways
- All-in firmware under 888 KiB including ESP-IDF/FreeRTOS runtime, Wi-Fi, TLS, and cert bundle — the smallest AI personal assistant
- GPIO read/write control with guardrails turns any ESP32 into a natural-language-controlled IoT device
- Persistent memory across reboots with timezone-aware scheduling (daily, periodic, one-shot)
- Supports Anthropic, OpenAI, OpenRouter, and Ollama — no cloud lock-in
FAQ
What is zclaw?
zclaw is a personal AI assistant written in C that runs on ESP32 microcontrollers with an all-in firmware budget of 888 KiB. It supports Telegram chat, GPIO control, cron scheduling, custom tools, and persistent memory.
What hardware does zclaw run on?
ESP32-C3, ESP32-S3, ESP32-C6, and ESP32-WROOM. The recommended starter board is the Seeed XIAO ESP32-C3 (~$5).
How much does zclaw cost?
Free and open-source. You need an ESP32 board (~$5-15) and an LLM API key (or local Ollama).
Who competes with zclaw?
MimiClaw (also ESP32, C), PicoClaw (Go, edge hardware), and NullClaw (Zig, smallest x86 footprint).
Executive Summary
zclaw is a personal AI assistant written in pure C that runs on ESP32 microcontrollers with a strict all-in firmware budget of 888 KiB. It brings natural language AI to the cheapest possible hardware — a $5 ESP32 board — with GPIO control, timezone-aware scheduling, persistent memory, and Telegram integration. [1]
| Attribute | Value |
|---|---|
| Author | tnm |
| Language | C |
| License | MIT |
| GitHub Stars | 2,159 (as of June 2026) |
| Firmware Size | ≤ 888 KiB budget (~830 KiB actual) |
| Latest Release | v2.13.0 (March 22, 2026) |
Product Overview
zclaw targets the absolute minimum viable AI assistant: a microcontroller with Wi-Fi, an LLM API key, and natural language control over physical hardware. The 888 KiB budget includes everything — zclaw logic, ESP-IDF/FreeRTOS runtime, Wi-Fi/networking, TLS/crypto, and certificate bundles. [1]
Key Capabilities
| Capability | Description |
|---|---|
| Telegram Chat | Chat via Telegram or hosted web relay |
| GPIO Control | Read/write with guardrails, bulk gpio_read_all |
| Scheduling | Timezone-aware daily, periodic, and one-shot tasks |
| Custom Tools | Built-in + user-defined tools via natural language |
| Persistent Memory | Survives reboots, stored in flash |
| Personas | Neutral, friendly, technical, or witty |
| Multi-Provider | Anthropic, OpenAI, OpenRouter, Ollama |
| Encrypted Credentials | Secure flash mode for NVS storage |
| Sensors & I2C | First-class DHT11/DHT22 reads; generic i2c_scan/read/write tools |
Hardware
Tested on ESP32-C3, ESP32-S3, ESP32-C6, and ESP32-WROOM. Recommended starter: Seeed XIAO ESP32-C3 (~$5). [2]
Recent Development (since February 2026)
Four releases shipped between March 1 and March 22, 2026 (v2.8.2 through v2.13.0). v2.13.0 added first-class dht_read support for DHT11/DHT22 temperature/humidity sensors, generic I2C tools (i2c_scan, i2c_write, i2c_read, i2c_write_read), and TLS optimizations that hold the binary at 830 KiB — under the 888 KiB budget. [3] The repo remains maintained as of June 2026 (last commit May 17, 2026), though release cadence has slowed since the March burst. [1]
Strengths
- Smallest AI assistant — 888 KiB all-in firmware, runs on $5 hardware
- Physical world control — GPIO read/write enables real IoT automation via natural language
- No cloud lock-in — Supports Anthropic, OpenAI, OpenRouter, and local Ollama
- Persistent memory — Context survives reboots, unlike most microcontroller projects
- One-line install — Bootstrap script handles ESP-IDF setup
- Active community — 2,159 stars as of June 2026, well-documented with dedicated docs site
- Growing hardware reach — DHT sensor and generic I2C support landed in v2.13.0 [3]
Cautions
- ESP32 only — No x86, ARM server, or general Linux support
- Requires LLM API — The ESP32 can't run models locally (Ollama needs a separate host)
- Limited channels — Telegram and web relay only; no Discord, Slack, or WhatsApp
- C codebase — Higher barrier to contribution than TypeScript/Python alternatives
- Rate limits — Default 100/hour, 1000/day (configurable at compile time)
- Thin-wrapper skepticism — Critics note the intelligence lives in cloud APIs; the ESP32 is essentially an always-on HTTP client, so "smallest assistant" claims depend on what you count [4]
- Slowed release cadence — No releases since v2.13.0 (March 22, 2026), though commits continued through May 2026 [1]
What Developers Say
The Hacker News launch thread (130+ points) split between hardware enthusiasm and claw-fatigue. [4]
"I'm a simple man; I see ESP32, I upvote" — 0xbadcafebee, Hacker News [4]
"This is a great example of how silly this whole thing is. There's next to nothing to these claws." — roxolotl, Hacker News [4]
"Can we please move past this whole OpenClaw hype? Yes it's an llm in a loop and can call tools." — saberience, Hacker News [4]
Defenders countered that the value is an always-on device with no OS maintenance overhead, not local computation. Hackaday also covered the project shortly after launch. [5]
Pricing & Licensing
| Tier | Price | Includes |
|---|---|---|
| Open Source | Free | Full firmware, MIT license |
Hardware cost: ~$5-15 for an ESP32 board. Hidden costs: LLM API usage.
Funding: Not publicly disclosed — an independent open-source project by tnm, with no announced commercial backing. [1]
Competitive Positioning
| Competitor | Differentiation |
|---|---|
| MimiClaw | MimiClaw also targets ESP32 in C but focuses on ESP32-S3 with WebSocket; zclaw has stricter size budget and GPIO focus |
| PicoClaw | PicoClaw runs Go on edge Linux; zclaw runs bare-metal C on microcontrollers |
| NullClaw | NullClaw is smallest x86 binary (678 KB); zclaw is smallest microcontroller firmware (888 KiB) |
Bottom Line
zclaw pushes the personal AI assistant to its absolute minimum: a $5 microcontroller running under 888 KiB of firmware that can chat via Telegram, control GPIO pins, read sensors, schedule tasks, and remember context across reboots. As of June 2026 it sits at 2,159 stars and remains maintained, though the rapid-fire release pace of March 2026 has cooled. It's the proof that AI assistants don't need gigabytes — or even megabytes — provided you're comfortable with the intelligence living in a cloud API.
Recommended for: Hardware hackers, IoT enthusiasts, and anyone who wants an AI assistant on the cheapest possible device.
Not recommended for: Users wanting multi-channel support, rich integrations, or a general-purpose personal agent.
Research by Ry Walker Research • methodology