Zeroclaw
Fast, small, and fully autonomous AI personal assistant infrastructure, any OS, any platform — deploy anywhere, swap anything 🦀
Last verified:
What is Zeroclaw?
ZeroClaw is a private AI personal assistant infrastructure written entirely in Rust that runs 100% locally on your machine. It is a next-generation AI assistant framework designed as a runtime operating system for agentic workflows, abstracting models, tools, memory, and execution so agents can be built once and run anywhere. Your data never leaves your computer, making it ideal for privacy-conscious users who want full control over their AI assistant.
Key features include ultra-lightweight performance with a 3.4-8.8MB binary size, cold startup under 10ms, and less than 5MB RAM usage. It supports 22+ AI providers including OpenRouter, Anthropic, OpenAI, Ollama, Gemini, DeepSeek, Groq, Mistral, and Perplexity. ZeroClaw connects to multiple communication channels like Telegram, Discord, Slack, WhatsApp, Matrix, Signal, Email, and CLI. The pluggable trait-driven architecture allows you to swap any component (providers, channels, tools, memory backends) with zero code changes. It includes built-in security with pairing codes, sandboxing, filesystem scoping, and encrypted secrets.
ZeroClaw is built for systems engineers, AI researchers, hardware enthusiasts, developers building SaaS/API-first assistants, DevOps/SRE teams needing incident runbook agents, IoT/robotics developers working with drones or rovers, legal/compliance teams requiring audit trails, educators creating classroom labs, and makers working with Raspberry Pi or Arduino. It runs on any hardware starting from $10 devices, supports ARM/x86/RISC-V architectures, and offers both native and Docker sandboxed runtime options.
The tool includes a full-stack search engine for memory using SQLite vector + FTS5 keyword search (no Pinecone or Elasticsearch), 70+ integrations across 9 categories, and supports custom MCP servers. Teams benefit from secure-by-design features including pairing requirements, strict sandboxing, explicit allowlists, and no vendor lock-in with OpenAI-compatible provider support plus custom endpoints.
Zeroclaw pricing
Pricing model: Freemium
ZeroClaw is 100% free and open source under MIT and Apache 2.0 dual licensing. There are no paid plans, subscription fees, or premium tiers. The software itself is completely free - you only pay for the AI provider APIs you choose to use (OpenRouter, Anthropic, OpenAI, Ollama, etc.) and any hardware you deploy it on. Running on $10 hardware costs 98% less than a $599 Mac Mini. Pre-built binaries are available for free download from GitHub releases. Homebrew installation is free. The only costs are your own AI API usage costs and hardware costs, with ZeroClaw itself having zero licensing fees.
Zeroclaw pros
- Cold startup under 10ms for instant response
- Ultra-lightweight 3.4-8.8MB binary size
- Runs on $10 hardware with less than 5MB RAM
- 99% less memory usage than OpenClaw
- 98% cheaper than running on Mac Mini
- 100% Rust-written for memory safety guarantees
- 22+ AI providers supported with flexible switching
- Multi-platform channels including Telegram, Discord, Slack, WhatsApp
- Pluggable trait-driven architecture for easy extension
- Zero runtime overhead with uncompromised performance
- Secure by default with pairing codes and sandboxing
- Filesystem scoped to workspace only by default
- No vendor lock-in with OpenAI-compatible endpoints
- Built-in full-stack search engine without external dependencies
- Supports Docker sandboxed runtime for additional security
- 70+ integrations across 9 categories
- Open source with MIT and Apache 2.0 dual licensing
- Hot-applied config changes without restart
- Pre-built binaries available for Linux, macOS, Windows
- Supports local models via Ollama and LM Studio
Zeroclaw cons
- Requires Rust toolchain for building from source
- Compilation needs minimum 2GB RAM and 6GB disk space
- Gateway binds to localhost by default requiring tunnel for remote access
- Browser tools require opt-in and WebDriver setup
- Composio integration requires separate API key and opt-in
- Claude Code OAuth tokens cannot be used per Anthropic terms
- Channel allowlists use deny-by-default which may confuse new users
- WASM/edge runtimes planned but not implemented yet
- WhatsApp Business mode requires Meta Business API setup
- Custom provider endpoints need manual configuration
- No official Windows native build documentation clarity
- Python companion package needed for inconsistent tool calling providers
- Memory embedding requires separate embedding provider configuration
- Community open-skills sync disabled by default requiring explicit enable
- Original config.toml must exist before certain commands work
Frequently asked questions about Zeroclaw
What is ZeroClaw and what does it do?
ZeroClaw is a fast, small, and fully autonomous AI personal assistant infrastructure written in Rust. It serves as a runtime operating system for agentic workflows that abstracts models, tools, memory, and execution. It connects to 22+ AI providers (OpenRouter, Anthropic, OpenAI, Ollama, etc.), reaches users through multiple channels (CLI, Telegram, Discord, Slack, WhatsApp, Matrix, Signal, Email, Webhook), and acts through tools (shell, file, memory, cron, git, browser, hardware). Everything runs locally on your machine with your keys in your workspace, ensuring your data never leaves your computer.
How do I install ZeroClaw?
You can install ZeroClaw in several ways: 1) Homebrew: 'brew install zeroclaw' on macOS/Linux; 2) One-click bootstrap: clone the repo and run './bootstrap.sh' or use the remote one-liner 'curl -fsSL https://raw.githubusercontent.com/zeroclaw-labs/zeroclaw/main/scripts/bootstrap.sh | bash'; 3) Build from source: 'cargo build --release --locked' then 'cargo install --path . --force --locked'; 4) Pre-built binaries: download from GitHub releases for Linux (x86_64, aarch64, armv7), macOS (x86_64, aarch64), and Windows (x86_64). After installation, run 'zeroclaw onboard --api-key sk-... --provider openrouter' to set up.
What AI providers does ZeroClaw support?
ZeroClaw supports 22+ AI providers including OpenRouter, Anthropic, OpenAI, Ollama, Gemini, Venice, Vercel, Cloudflare, Moonshot, Kimi, Code Synthetic, OpenCode, Z.ai, GLM, MiniMax, Bedrock, Qianfan, Qwen, Groq, Mistral, xAI, DeepSeek, Together, Fireworks, Perplexity, Cohere, Copilot, LM Studio, and NVIDIA. It also supports custom OpenAI-compatible and Anthropic-compatible endpoints. Local providers include Ollama, llama.cpp server, vLLM, and Osaurus.
How does ZeroClaw security work?
ZeroClaw enforces security at every layer: 1) Gateway not publicly exposed - binds to 127.0.0.1 by default and refuses 0.0.0.0 without tunnel or explicit allow; 2) Pairing required - uses 6-digit one-time code on startup, exchanges via POST /pair for bearer token, all webhook requests require Authorization header; 3) Filesystem scoped - workspace_only = true by default, blocks 14 system dirs + 4 sensitive dotfiles, detects symlink escapes; 4) Access via tunnel only - requires active tunnel (Tailscale, Cloudflare, ngrok) for public bind. Channel allowlists use deny-by-default policy.
What channels/messaging platforms does ZeroClaw support?
ZeroClaw supports CLI, Telegram, Discord, Slack, Matrix, Signal, WhatsApp (both WhatsApp Web mode and WhatsApp Business Cloud API mode), Email, Webhook, Mattermost, iMessage, Linq, IRC, Lark, DingTalk, QQ, and Nostr. Channels require the daemon to be running. Each channel has allowlist configuration for security - empty allowlist denies all inbound messages, '*' allows all, otherwise exact-match allowlist. Telegram, Discord, and Slack are most commonly used.
How does ZeroClaw memory work?
ZeroClaw includes a full-stack search engine for memory with zero external dependencies (no Pinecone, no Elasticsearch). It uses SQLite vector database with embeddings stored as BLOB and cosine similarity search, FTS5 virtual tables for keyword search with BM25 scoring, and a custom weighted merge function for hybrid search. Memory backends include SQLite (default), PostgreSQL, Lucid bridge, Markdown files, or explicit 'none' backend. The agent automatically recalls, saves, and manages memory via tools. Embedding providers can be 'none', 'openai', or 'custom:https://...'.
Can I run ZeroClaw on Raspberry Pi or low-cost hardware?
Yes, ZeroClaw is specifically designed to run on minimal hardware. It runs on $10 hardware with less than 5MB RAM, which is 99% less memory than OpenClaw. The binary is only 3.4-8.8MB and cold startup is under 10ms even on 0.8GHz edge hardware. It supports ARM, x86, and RISC-V architectures with pre-built binaries for ARM64 Linux (Raspberry Pi), making it ideal for IoT, robotics, drones, rovers, and edge/air-gap deployments. Education and maker use cases include 'classroom lab in a box' with Raspberry Pi + Arduino.
What is the difference between ZeroClaw and OpenClaw?
ZeroClaw is written in Rust while OpenClaw is TypeScript, resulting in dramatically better performance: ZeroClaw uses less than 5MB RAM versus OpenClaw's greater than 1GB RAM, starts in under 10ms versus over 500 seconds, and has an 8.8MB binary versus 28MB dist. ZeroClaw is 99% less memory than OpenClaw and 98% cheaper to run (any $10+ hardware versus $599 Mac Mini). ZeroClaw is a complete重新设计 of the agent runtime kernel, not just another AI app. They are separate projects - ZeroClaw is not affiliated with openagen/zeroclaw.
How do I configure ZeroClaw?
Configuration is in ~/.zeroclaw/config.toml created by 'zeroclaw onboard'. Key settings include: api_key, default_provider (e.g., 'openrouter'), default_model (e.g., 'anthropic/claude-sonnet-4-6'), default_temperature. Memory config includes backend ('sqlite', 'postgres', 'lucid', 'markdown', 'none'), auto_save, embedding_provider. Gateway config includes port (default 42617), host (default 127.0.0.1), require_pairing, allow_public_bind. Autonomy config includes level ('readonly', 'supervised', 'full'), workspace_only, allowed_commands, forbidden_paths. Runtime can be 'native' or 'docker'.
Is ZeroClaw open source and how can I contribute?
Yes, ZeroClaw is 100% open source under MIT and Apache 2.0 dual licensing. The official repository is https://github.com/zeroclaw-labs/zeroclaw. To contribute: clone the repo, enable pre-push hook with 'git config core.hooksPath .githooks', run 'cargo build' and 'cargo test --locked'. Run './scripts/ci/rust_quality_gate.sh' and 'cargo fmt' before PR. Extend by trait: new Provider goes to src/providers/, new Channel to src/channels/, new Tool to src/tools/. Update reference docs when behavior changes and keep SUMMARY and README in sync. The community welcomes contributions that follow the architecture and quality bar.