◤ THE CLAW INDEX
VOL. 01 MAY 2026 FIELD GUIDE ED. ECOSYSTEM
An open-source agent ecosystem, mapped to silicon

The Claw
Ecosystem.

A Hardware Field Guide
Eight forks. Four tiers. One origin.

OpenClaw spawned an entire family of AI-agent frameworks within ninety days. From a $2 microcontroller speaking to a cloud LLM, to a sealed enterprise sandbox with WASM and TEE — here is where each one runs, and why someone built it that way.

/01 The Origin
JAN 30 · 2026 270K+ STARS 84 DAYS
OpenClaw shipped a 430,000-line TypeScript monster — powerful, viral, heavy. The community responded by porting it onto every class of hardware imaginable.
/02 ─ Tier One

Microcontroller — no OS.

Bare metal. No Linux. No Python. Pure firmware speaking to an LLM through whatever interface the chip can muster — Wi-Fi, Bluetooth, sometimes nothing at all.

01

QClaw.

Arduino Uno Q · Qualcomm QRB2210 + STM32U585

The most radical Claw variant. Arduino Uno Q is dual-silicon — a Linux-capable Qualcomm MPU and an STM32 MCU sharing one PCB. QClaw runs Qwen 3.5 entirely on-device, generates Arduino sketches, compiles them with arduino-cli, and flashes the MCU over SWD via GPIO. No cloud. No API keys. No subscription. An 8-tool agentic surface with a 15-skill pre-router.

Edge

  • True air-gap operation
  • Compiles & flashes itself
  • No ongoing API cost
  • Genuine AI-education vector

Trade

  • Local model only — capability ceiling
  • Requires Uno Q hardware
  • Not production-grade
  • Experimental, community-led
02

MimiClaw.

ESP32-S3 · 240 MHz dual-core · 16 MB Flash / 8 MB PSRAM

OpenClaw-in-C for the ESP32-S3. No OS, no Python, no Node — pure ESP-IDF firmware. Acts as a gateway: Telegram → ESP32-S3 → Claude / OpenAI → GPIO. Core 0 handles Wi-Fi and bot polling; Core 1 runs the agent loop and LLM calls. Persistent memory via SPIFFS. Ideal for battery-powered, always-on IoT AI.

Edge

  • Cheapest AI-agent hardware
  • Days of uptime on a power bank
  • Persistent memory & history
  • Direct sensor / actuator access

Trade

  • Requires Wi-Fi + API key
  • API token burn — watch carefully
  • C expertise for customization
  • Telegram dependency
03

ESPClaw.

ESP32 / ESP32-C3 / ESP32-S3 · Arduino IDE or ESP-IDF

The broader ESP32-family Claw port. Where MimiClaw targets S3 specifically, ESPClaw supports the full ESP32 lineup — including the classic ESP32 and the C3. More accessible (Arduino IDE works), less raw than MimiClaw's pure C. Good for robotics kits, prototyping, and anyone comfortable with the Arduino ecosystem. WebSocket gateway for multi-device coordination.

Edge

  • Widest ESP32 compatibility
  • Arduino-IDE friendly
  • Robotics & maker community
  • Lowest entry cost in the table

Trade

  • Less optimized than MimiClaw
  • API required
  • Classic ESP32 RAM is tight
  • Smaller, scrappier ecosystem
/03 ─ Tier Two

Single-Board — Linux at the edge.

A real kernel and a real filesystem in $10–$100 of hardware. The sweet spot for always-on smart-home agents, sensor nodes, and learning the agent-architecture rope without a server bill.

04

PicoClaw & Nanobot.

Raspberry Pi · LicheeRV Nano · Any Linux SBC ≥ 10 MB RAM

PicoClaw (Go, by Sipeed) — 8 MB binary, < 10 MB RAM, one-second startup. Runs on any Linux SBC, even a $15 LicheeRV Nano or an old Android phone. Nanobot (Python, 4,000 lines) — runs on a Raspberry Pi 3B+ with 191 MB RAM, supports Ollama for local models. Both ideal for always-on smart-home agents, sensor nodes, and learning the architecture.

Edge

  • Runs where OpenClaw can't
  • Nanobot — easiest to learn
  • PicoClaw — fastest startup
  • Old hardware finds new purpose

Trade

  • No direct GPIO (via OS only)
  • Fewer features than OpenClaw
  • PicoClaw — no formal security layer
  • Nanobot's local model is slow
/04 ─ Tier Three

PC, Server & Cloud.

Where the original lives, and where the security-first forks landed. More RAM, more features, more attack surface — and, in IronClaw's case, more sandbox.

05

OpenClaw.

PC · Mac · Linux Server · 1.5 GB RAM minimum · Node.js / TypeScript

The original. 430,000 lines of TypeScript, 700+ skills, 22+ messaging channels, full computer-use, web browsing, code execution. Most powerful but heaviest. Palo Alto Networks flagged it as a “security nightmare” — documented CVEs and malicious skills found in ClawHub. Always run with Docker isolation. Best for power users on dedicated machines or VPS.

Edge

  • Most complete feature set
  • Largest plugin ecosystem
  • Best community support
  • Multi-platform · Win · Mac · Linux

Trade

  • Documented security CVEs
  • Huge codebase — hard to audit
  • Won't run on SBC or MCU
  • RAM-hungry
06

NanoClaw & IronClaw.

PC / Server · Docker required · TypeScript / WASM sandbox

NanoClaw — 700 lines of TypeScript. Each agent runs in a separate Docker container, no host filesystem access. Built on the philosophy: treat AI agents as untrusted. Supports WhatsApp via container-isolated Playwright. IronClaw adds a WASM sandbox + TEE (Trusted Execution Environment) + encrypted credential vaults — tool code inside the sandbox never touches raw credentials. Best for regulated industries and enterprise deployments.

Edge

  • Strongest security model in the family
  • Tiny, auditable codebase
  • IronClaw — cryptographic proof
  • Multi-agent isolation by default

Trade

  • Requires Docker
  • Fewer features than OpenClaw
  • IronClaw — complex setup
  • PC / server only
8
Forks shipping
in 84 days
$2
Cheapest entry
(ESPClaw / C3)
430K
Stars on the
original repo
1.5×10⁵
RAM gap from
MCU to OpenClaw
/05 ─ Decision

Pick your tier.

A three-step shortcut. Start from what you actually care about — privacy, cost, or capability — and the right Claw is one column away.

If your priority is
privacy.
→ QClaw or Nanobot+Ollama. The model runs on the device. No API key ever leaves the building. Pay once, run forever — at the cost of capability.
→
If your priority is
cost.
→ ESPClaw or MimiClaw. $2–$10 of silicon plus an API key. Watch your token burn — the chip is cheap, the cloud calls aren't always.
→
If your priority is
capability.
→ OpenClaw, NanoClaw, or IronClaw. Full feature set lives on a PC or VPS. Audit the code, run in Docker, and if you're regulated — go IronClaw.
→
/06 ─ Side-by-Side

Cross-tier matrix.

One table. Eight forks. Six dimensions that actually change a buying decision.

Project Hardware Min Cost Cloud API GPIO Security Complexity Best For
QClawArduino Uno Q~$45NoneDirect flashAir-gapMediumAI education · offline agents
MimiClawESP32-S3$5 – $10RequiredNativeWi-Fi exposureMedium · CIoT AI agents · low power
ESPClawESP32 / C3 / S3$2 – $8RequiredNativeWi-Fi exposureEasy · ArduinoMakers · robotics · prototyping
PicoClawRPi / any Linux SBC$10 – $15OptionalVia OSBasicEasy · GoSmart-home edge nodes
NanobotRPi 3B+ / PC$35+OptionalVia OSBasicEasiest · PythonLearning · local Ollama LLM
OpenClawPC / Server$0 (VPS)RequiredNoneCVEs · needs DockerComplex · 430K LOCPower users · full features
NanoClawPC / Server + Docker$0 (VPS)RequiredNoneContainer isolationSimple · 700 LOCSecurity-conscious deployments
IronClawPC / Server + Docker$0 (VPS)RequiredNoneWASM + TEE + vaultComplexEnterprise · regulated industries