What is ZeroClaw?

Introduction

ZeroClaw is a runtime operating system for agentic workflows — a lightweight, trait-driven infrastructure for building and deploying autonomous AI agents. Written entirely in Rust, it abstracts models, tools, memory, and execution so agents can be deployed anywhere with minimal overhead.

The project tagline says it best: "Fast, small, and fully autonomous AI assistant infrastructure — deploy anywhere, swap anything."

Unlike heavier frameworks that require gigabytes of RAM, ZeroClaw runs on $10 hardware with less than 5MB RAM — 99% less memory than OpenClaw. Its single-binary Rust runtime (8.8MB) enables near-instant cold starts in under 10 milliseconds.

Why ZeroClaw?

The AI agent landscape is growing rapidly, but most frameworks are built in Python or TypeScript — which struggle with performance on edge hardware and at scale. ZeroClaw takes a fundamentally different approach:

• Lean by default: Runs on $10 boards with <5MB RAM, 99% less than comparable frameworks.
• Secure architecture: Pairing-based gateways, strict sandboxing, explicit allowlists, workspace scoping, and encrypted secrets at rest.
• Fully swappable: Every core subsystem implements traits — providers, channels, tools, memory, tunnels, runtime, and observability can all be substituted.
• Fast cold starts: Single-binary Rust runtime starts in under 10ms — no warm-up, no JIT, no garbage collector.
• Cost-efficient: Designed for Raspberry Pi, embedded systems, and small cloud instances.
• Provider agnostic: OpenAI-compatible endpoints plus custom provider support — connect to OpenAI, Anthropic, OpenRouter, or any custom endpoint without vendor lock-in.

Trait-Driven Architecture

ZeroClaw's defining feature is its trait-driven design. Every subsystem is swappable through a common trait interface:

| Subsystem | Trait | Default Implementations | |-----------|-------|------------------------| | AI Models | Provider | OpenAI, Anthropic, OpenRouter, custom endpoints | | Channels | Channel | CLI, Telegram, Discord, Slack, Matrix, Email, Webhook, 9+ more | | Memory | Memory | SQLite hybrid search, PostgreSQL, Markdown files, none | | Tools | Tool | Shell, file ops, git, browser, HTTP, screenshots, hardware | | Runtime | RuntimeAdapter | Native, Docker sandboxed | | Observability | Observer | Logging, multi-observer | | Tunnel | Tunnel | Cloudflare, Tailscale, ngrok, custom |

This means you can swap out any piece — use Anthropic instead of OpenAI, PostgreSQL instead of SQLite, Tailscale instead of Cloudflare — without changing your agent logic.

Memory System

ZeroClaw includes a custom full-stack search engine with zero external dependencies — no Pinecone, no Elasticsearch, no LangChain required:

• Vector storage: Embeddings stored as SQLite BLOB with cosine similarity search.
• Keyword search: FTS5 virtual tables with BM25 scoring for text-based retrieval.
• Hybrid merge: Weighted combination of vector and keyword results for best accuracy.
• Embedding provider: Trait-based — supports OpenAI embeddings, custom URL endpoints, or disabled mode.
• Chunking: Line-based markdown chunking with heading preservation for context-aware splits.
• Caching: LRU eviction in SQLite embedding_cache table for fast repeated queries.

Alternative memory backends include PostgreSQL, Markdown files, or no memory at all — chosen via configuration.

15+ Messaging Channels

ZeroClaw supports deployment to a wide range of messaging platforms:

• Chat: CLI, Telegram, Discord, Slack, Mattermost
• Encrypted: Signal, iMessage, WhatsApp
• Enterprise: Matrix, Lark, DingTalk, Nostr
• Web: Email, IRC, Webhook, QQ, Linq

Each channel implements the Channel trait, making it straightforward to add new integrations.

Installation

Install ZeroClaw using one of these methods:

Homebrew (macOS/Linux):

brew install zeroclaw

One-click bootstrap:

git clone https://github.com/zeroclaw-labs/zeroclaw.git
cd zeroclaw
./bootstrap.sh

From source:

cargo build --release --locked
cargo install --path . --force --locked

Pre-built binaries are available for Linux (x86_64, aarch64, armv7), macOS (x86_64, aarch64), and Windows (x86_64) from GitHub Releases.

Prerequisites

• Linux/macOS: Build essentials, pkg-config, Rust toolchain
• Windows: Visual Studio Build Tools (Desktop C++ development), Rust via rustup
• Compilation: Minimum 2GB RAM + swap, 6GB free disk; recommended 4GB+ RAM, 10GB+ disk

Configuration

ZeroClaw is configured via ~/.zeroclaw/config.toml (TOML format). Configuration covers:

• Provider selection and authentication
• Channel setup and allowlists
• Memory backend (SQLite, PostgreSQL, Markdown, none)
• Runtime kind (native or Docker)
• Security policies and rate limits
• Custom tool definitions
• Tunnel and identity configuration

Authentication Modes

• API keys: Stored encrypted in ~/.zeroclaw/auth-profiles.json
• OpenAI Codex OAuth: Device-code or browser callback flow
• Claude Code/Anthropic: Setup tokens and subscription auth
• Multi-profile support: Profile format <provider>:<profile_name>

Core CLI Usage

# Onboarding (interactive or scripted)
zeroclaw onboard --interactive
zeroclaw onboard --api-key sk-... --provider openrouter

# Chat interaction
zeroclaw agent -m "Your message"
zeroclaw agent  # Interactive mode

# Start gateway/daemon
zeroclaw gateway
zeroclaw daemon

# System diagnostics
zeroclaw status
zeroclaw doctor

Performance Benchmarks (February 2026)

| Metric | ZeroClaw | PicoClaw | NanoBot | OpenClaw | |--------|----------|---------|---------|----------| | Language | Rust | Go | Python | TypeScript | | Runtime RAM | <5MB | <10MB | >100MB | >1GB | | Startup (0.8GHz) | <10ms | <1s | >30s | >500s | | Binary size | 8.8MB | ~8MB | N/A | ~28MB |

Startup times normalized for edge hardware; comparisons assume normalized load.

Security Model

ZeroClaw takes security seriously with multiple layers of protection:

• Pairing-based gateways for secure multi-user access
• Workspace scoping: Agents confined to designated directories
• Encrypted secrets: At-rest encryption for credentials
• Allowlists: Explicit permission rules for channels and tools
• Rate limiting: Built-in protection against abuse
• Sandbox runtime: Optional Docker-based isolation

Use Cases

ZeroClaw is ideal for:

• Autonomous coding agents that can read, write, and test code
• Research pipelines that gather, summarize, and synthesize information
• Customer support bots with tool access to internal systems
• DevOps automation agents for infrastructure management
• Edge deployment — agents running on embedded hardware and IoT devices
• Multi-channel bots — one agent across Telegram, Discord, Slack, and more

Next Steps

• Check out the latest releases to download ZeroClaw
• Read the ZeroClaw vs OpenClaw comparison
• Visit the official documentation for detailed references