Every shell in the sea

On a rocket ship full of stars. The one that started the molt.

4,000 lines of Python. Research-friendly, ultra-minimal

Sub-10ms startup, 3.4MB binary, <5MB RAM

Runs on $10 RISC-V boards, <10MB RAM, 1s boot

5 files, one process, OS-level container isolation

Self-described OpenClaw alternative with deep IM platform integration

Universal agent UI — supports Gemini CLI, Claude Code, Codex, and more in one interface

Supports 10+ IM platforms with agent capabilities, knowledge base, and plugin system

Built in Rust from scratch — 137K LOC, 1,767+ tests. One binary, one command, agents live. Not a chatbot wrapper. An actual OS for autonomous agents.

10k+ stars. Memory-first architecture with proactive agent behavior, local-first privacy, reduced LLM token costs.

OpenClaw running serverless on Cloudflare Workers — no persistent infra needed

Rust implementation focused on privacy and security

Earned $10K in 7 hours as a demo — task-oriented agent design

Persistent skill memory that lets agents learn and reuse across tasks

678 KB Zig binary — runs on $5 hardware, 22+ providers, 17 channels, fully pluggable everything

Pure C on ESP32-S3 ($5). No OS needed

Self-improving skill system — agent writes reusable SKILL.md docs as it solves problems, building a personal knowledge base over time

Zero unsafe Rust, built-in voice I/O, MCP support

1.7k stars. Same claw, new vibe — a companion-first agent with persistent memory and personality layered on top of OpenClaw.

888 KiB firmware — runs on ESP32, not a server

Concurrent Channel/Branch/Worker/Cortex architecture — 50 users at once without blocking

Composio's hardened fork with 500+ app integrations out of the box

Polished Web UI with sandboxed agent runtime and built-in IM support

Entire personal AI assistant running in a single browser tab — no server, no Docker, no Node.js

Rust-native chat agent with NanoClaw-inspired security model

Single 15 MB Go binary — 4,500 lines of code, runs on a $4/mo VPS

Chinese ecosystem first — native QQ, Feishu, DingTalk, WeCom channels with DeepSeek, Doubao, Qwen, Kimi, and Zhipu AI providers, all in 16K lines of code

Personality system (Heartware) for self-improving AI

First mobile-native OpenClaw alternative — Kotlin Multiplatform for Android and iOS

Omni-channel computer automation — browser driver, shell control, cron scheduling, voice commands, and the widest channel support including WeChat and DingTalk

No LLM required — deterministic, rule-based, zero token costs

Single binary, zero runtime dependencies — a full personal AI agent with TUI, channels, tools, memory, skills, and cron in one Go executable

Docker-isolated agent fleet — each agent in its own container, API keys in a vault proxy agents never see, per-agent daily and monthly budget caps

Agent-as-repo architecture — agent.yaml, SOUL.md, RULES.md, memory/, tools/, skills/ are all git-committed files you can fork, branch, and diff

Cloud-first OpenClaw — multi-agent with shared soul documents, agent-to-agent delegation, public chat UI, SyncBoard admin dashboard, and Convex-powered persistence

Official Go client for RPi/Linux — voice transcript streaming and local TTS

Hardened Rust core, signed WASM plugins, strict local-first defaults

Multi-agent profiles, minimal footprint — no framework bloat

94 built-in tools turn Telegram into a full AI workstation — from headless browsing and Gmail to Python execution and persistent memory

Multi-channel routing and task orchestration — self-hostable, lightweight Python

Bash-only personal assistant — runs on minimal systems, no Node.js or runtimes needed