ZeroClaw: 5MB Rust AI Agent Runtime with Sub-10ms Startup and Hardware Control

Introduction

ZeroClaw is a ground-up Rust reimplementation of OpenClaw, not a fork, developed by the independent organization zeroclaw-labs. It packages a full AI personal assistant into a single ~5MB binary that starts in <10ms and idles at ~16MB RAM — two orders of magnitude lighter than OpenClaw's Node.js stack (200MB+ dependencies, 380MB+ idle, up to 1.5GB under load). This resource profile unlocks 24/7 agent runtimes on Raspberry Pi, industrial controllers, and 1–2GB VPS instances, turning theoretical savings into real server‑bill reductions.

Three Core Highlights

Highlight 1: Resource Usage Drops Two Orders of Magnitude

Startup <10ms, binary ~5MB (vs. OpenClaw ~200MB with deps), idle memory ~16MB (vs. 380MB+). The project claims memory compression from ~1.5GB to MB‑level, making previously impossible deployments on low‑end hardware practical.

Highlight 2: Security‑First by Default

The default autonomy level is supervised: medium‑risk operations require approval, high‑risk are blocked. Workspace boundaries, command policies, and four OS‑level sandboxes (Landlock, Bubblewrap, Seatbelt, Docker) isolate every tool call. Each invocation produces an encrypted tool receipt, providing a full audit trail. A YOLO mode exists for trusted dev environments, but it is an explicit opt‑in escape hatch.

Highlight 3: Agent That Drives Physical Hardware

Via a Peripheral trait, ZeroClaw connects directly to GPIO, I2C, SPI, and USB on Raspberry Pi, STM32, Arduino, and ESP32. The documented example: user says "turn on led" → RAG fetches chip datasheet → LLM decides → gpio_write(13, 1) executes. This moves the agent out of the chat box and into real devices.

What Fits in One Binary

Connectivity Layer

~20 LLM providers (Anthropic, OpenAI, Ollama, etc.) with fallback chains.

30+ channels: Discord, Telegram, Matrix, email, voice, webhook, CLI.

Tools: shell, browser, HTTP, hardware, custom MCP servers.

Everything runs on your machine, with your keys, in your workspace.

Runtime Components

HTTP/WebSocket gateway + Web dashboard (chat, memory browser, config editor, cron management, tool inspector).

SOP engine: MQTT/webhook/cron/peripheral‑triggered standard operating procedures with approval gates and checkpoint resume.

ACP (Agent Client Protocol) for IDE/editor integration.

Memory: SQLite + embeddings for persistence.

Four Philosophical Principles

You own it → Security first → Minimalism → No vendor lock‑in.

Installation & Configuration

curl -fsSL https://raw.githubusercontent.com/zeroclaw-labs/zeroclaw/master/install.sh | sh<br/>zeroclaw quickstart          # one‑shot init: pick provider, generate working config<br/>zeroclaw agent -a <alias>    # enter interactive chat<br/>zeroclaw service install     # register as systemd/launchctl/Windows service for background residency

Windows offers a Rust‑free precompiled PowerShell install path. The release page (current v0.8.5) provides cross‑platform packages: deb, AppImage, dmg, msi, Android.

Migration from OpenClaw

ZeroClaw deliberately supports compatibility: it reads OpenClaw Markdown identity files and supplies config/memory migration commands. This lowers switching cost instead of demanding a clean‑slate rewrite.

Team Adoption Considerations

Audit‑First

Tool receipts are critical for team scenarios — every tool call leaves an encrypted receipt, enabling compliance review alongside the default supervised level.

Config‑as‑Code

All configuration lives in a single TOML file; V3 config requires as few as four sections (provider, agent, risk profile, etc.), making it Git‑friendly and reviewable.

Upgrade Discipline

The project explicitly marks interfaces as evolving. Upgrades require reading release notes; some config keys may change. Teams should assign a version‑watch role and avoid blind auto‑updates.

Who It's For (and Not For)

Good fit: Developers running 24/7 personal AI assistants on Raspberry Pi, IoT devices, or 1GB VPS; security‑sensitive contexts needing on‑premise data and default‑deny tool execution; OpenClaw users seeking a lighter alternative.

Bad fit: Heavy reliance on OpenClaw's ClawHub skill ecosystem (skills not directly compatible, community ~10x smaller); expecting "small runtime = small everything" — local Ollama models still need GBs of RAM (a Pi Zero can run the runtime but not a 7B model); environments requiring frozen APIs — wait for stabilization.

Pros, Cons & Gotchas

Pros

Hits a validated pain point (OpenClaw's heaviness is a community‑acknowledged sore spot).

Rust rewrite delivers order‑of‑magnitude resource gains.

Security defaults rank among the best in agent runtimes.

Hardware control is a unique differentiator.

Dual licensing, RFC process, CI, active Discord; commits as recent as publication date — governance maturity exceeds many peers.

README includes an official repo notice warning about impersonators, handled professionally.

Cons & Gotchas

Young: 8‑month project; APIs still evolving — read release notes before upgrading.

Ecosystem gap: Community, plugins, and ready‑made skills are an order of magnitude smaller than OpenClaw; ClawHub skills don't work out of the box.

Runtime small ≠ model small: Local model memory is separate; don't expect the 5MB binary to make a Pi run large LLMs.

949 open issues: High demand, possible response delays — search issues first when hitting problems.

Number precision: Official comparison charts, info cards, and docs draw from different version snapshots; cite orders of magnitude, not decimal points.

Author's Perspective

ZeroClaw is a rare clear‑headed player in the "rewrite economy." Most rewrites fail because ecosystem lags, users don't migrate, and the project becomes a tech demo. ZeroClaw succeeds on two fronts: it anchors the rewrite rationale in quantifiable resource numbers (answering "why rewrite?") and provides compatibility/migration tooling to minimize switching cost (answering "why switch?"). Both answers hold.

Even more notable is its restraint. The four principles emphasize ownership and security — no "most features" or "biggest ecosystem." In an agent landscape chasing "do more automatically," this is a counter‑consensus choice. Default supervised + per‑call receipts states plainly: autonomy's prerequisite is trustworthiness.

Eight months is too early to claim "replaces OpenClaw"; ecosystems are built over time, not technology alone. Practical advice: adopt now for low‑resource deployments and security‑sensitive scenarios — it's nearly the only option there. For ecosystem‑driven needs, watch and wait. And always verify the zeroclaw-labs organization; the project itself warns of impersonators, so confirm the repo URL before installing.

GitHub repository card: https://opengraph.githubassets.com/1/zeroclaw-labs/zeroclaw