Zippy — ZeptoClaw mascot
v0.7.1 — Available on crates.io

Ultra-lightweight
Personal AI Assistant.

We studied the best AI assistants — and their tradeoffs. Then we built one binary that gets it all right.

⚡ ~6MB binary
🔌 9 providers
📡 9 channels
🧪 3,100+ tests
Install ZeptoClaw Read Docs
zeptoclaw — agent mode
$
Features

Everything. In One Binary.

OpenClaw's integrations without the 100MB. NanoClaw's security without the TypeScript bundle. PicoClaw's size without the bare-bones feature set.

🤖

Multi-Provider LLM

Claude and OpenAI with automatic retry on failures and fallback between providers. Exponential backoff, token budget tracking, and cost estimation built in.

⚡

Streaming Responses

Real-time SSE streaming from both Claude and OpenAI. Watch responses appear token-by-token in CLI, gateway, or batch mode.

🐝

Agent Swarms

Delegate subtasks to specialized sub-agents with role-specific prompts and tool whitelists. Recursion blocking prevents infinite loops.

🔌

Plugin System

Extend with JSON manifest plugins — define custom tools with command templates, parameter schemas, and validation. Auto-discovered at startup.

📋

Agent Templates

4 built-in templates (coder, researcher, writer, analyst) plus custom JSON templates. Override system prompt, model, tokens, and temperature.

📦

Container Isolation

Native runtime by default. Opt into Docker or Apple Container isolation with --containerized for full sandboxing.

💬

5-Channel Gateway

Telegram, Slack, Discord, WhatsApp, Lark, Email, Webhook, and CLI channels. Channel factory with per-channel auth, allowlists, and unified message bus.

🧠

Memory & History

Long-term key-value memory with categories and tags. Conversation history with session discovery, search, and cleanup. Persistent across restarts.

📊

Telemetry & Cost Tracking

Prometheus and JSON metrics export. Per-model cost estimation for 8 models. Token budget enforcement with atomic lock-free counters.

🔐

Tool Approval Gate

Policy-based tool gating — require approval before dangerous tools execute. SSRF prevention, path traversal detection, shell blocklists.

📂

Batch Mode

Process hundreds of prompts from text or JSONL files. Template support, text or JSONL output, stop-on-error control.

🦀

Written in Rust

Memory safety without garbage collection. 3,100+ tests. Async-first with Tokio. Minimal resource usage even under heavy load.

The Claw Ecosystem

The Best of Every Claw

We studied what works — and what doesn't. OpenClaw proved an AI assistant can handle 12 channels and 100+ skills. But it costs 100MB and 400K lines. NanoClaw proved security-first is possible. But it's still 50MB of TypeScript. PicoClaw proved AI assistants can run on $10 hardware. But it stripped out everything to get there.

ZeptoClaw took notes. The integrations, the security, the size discipline — without the tradeoffs each one made.

OpenClaw

TypeScript

The comprehensive one. Every feature, every channel, every integration.

  • 52+ modules
  • 12 chat channels
  • Voice + Live Canvas
  • Full production stack

NanoClaw

TypeScript

The hackable one. Fork it, read it, make it yours.

  • ~5,000 LOC
  • Agent swarms
  • Skills-based extensibility
  • Container-native

PicoClaw

Go

The tiny one. Runs on hardware that costs less than lunch.

  • <10MB RAM
  • ARM64 + RISC-V
  • 7 LLM providers
  • $10 hardware ready

ZeptoClaw

Rust

The one that took notes. Security, integrations, and minimalism — without the tradeoffs.

  • ~6MB binary, 3,100+ tests
  • 32 tools + plugin system
  • 9 channels + agent swarms
  • Container isolation
Provider Stack

Resilient by Default

Claude (Anthropic)
GPT-5.1 (OpenAI)
Auto-Retry (429/5xx)
Auto-Fallback
Architecture

Message-Driven Design

Built on a MessageBus architecture for loose coupling and easy extensibility.

Channels

Telegram, Slack, Discord, WhatsApp, Lark, Webhook, CLI — unified message interface

Agent Loop

Approval gate → Token budget → Tool execution → Streaming response

Provider Stack

Base provider → Fallback → Retry with cost tracking

Who It's For

Built for Builders

🏠

Self-Hosters

Run your own AI assistant on a VPS or home server. Single binary, no runtime dependencies, Docker-ready.

🔒

Security-Conscious Teams

SSRF prevention, path traversal detection, and shell blocklists by default. Optional container isolation for full sandboxing.

🏢

Multi-Tenant Operators

Deploy one binary for many users. Isolated workspaces, per-tenant config, health endpoints, and usage metrics.

~6MB
Release Binary
32+
Tools + Plugins
9
Channels
3100+
Tests Passing
Deploy

Deploy Anywhere

One click for managed platforms. One command for any VPS.

DigitalOcean Railway Render Fly.io

Any VPS — Hetzner, Contabo, AWS, Vultr, Linode

$ curl -fsSL https://zeptoclaw.com/setup.sh | bash
Hetzner Contabo AWS EC2 Vultr Linode OVH Any Ubuntu/Debian VPS
Getting Started

Up and Running in 60 Seconds

1. Install

# One-liner (macOS / Linux)
$ curl -fsSL https://raw.githubusercontent.com/qhkm/zeptoclaw/main/install.sh | sh

# Homebrew
$ brew install qhkm/tap/zeptoclaw

# Docker
$ docker pull ghcr.io/qhkm/zeptoclaw:latest

# Build from source
$ cargo install zeptoclaw --git https://github.com/qhkm/zeptoclaw

2. Run interactive setup

$ zeptoclaw onboard
# Walks you through provider keys, channels, and workspace config

3. Start your assistant

# CLI mode with streaming
$ zeptoclaw agent --stream -m "Hello, set up my workspace"

# Use a built-in template
$ zeptoclaw agent --template researcher -m "Analyze this API"

# Gateway mode — Telegram/Slack/Discord/Webhook
$ zeptoclaw gateway

# Batch mode — process prompts from file
$ zeptoclaw batch --input prompts.txt --format jsonl