OpenClaw vs Every Alternative

The Definitive Personal AI Assistant Comparison

OpenClaw vs 所有替代方案

個人 AI 助手完全比較指南

"The best AI assistant is the one that actually does things for you — not just talks." — Peter Steinberger, OpenClaw creator

「最好的 AI 助手是那個真正為你做事的——而不是只會聊天的。」—— Peter Steinberger，OpenClaw 創建者

The Landscape in March 2026

OpenClaw exploded from 0 to 320K GitHub stars in under four months. Created by Austrian programmer Peter Steinberger (PSPDFKit founder, later hired by OpenAI), it spawned an entire ecosystem: Chinese cloud giants (Alibaba, Tencent, ByteDance) created their own forks, nearly 1,000 people lined up at Tencent's Shenzhen HQ to have it installed, and a cottage industry charging ~$72 per installation emerged.

But OpenClaw's "god-mode" autonomy — shell access, file system control, browser automation, messaging app integration — also triggered serious security concerns. Agents have been tricked via prompt injection into uploading cryptocurrency keys, deleting code libraries, and exfiltrating financial data.

This guide compares every notable alternative across the categories that matter: architecture, security, messaging platform support, LLM flexibility, capabilities, and community maturity.

2026 年 3 月的格局

OpenClaw 在不到四個月內從 0 飆升至 320K GitHub 星數。由奧地利程式設計師 Peter Steinberger（PSPDFKit 創辦人，後被 OpenAI 聘用）創建，它催生了整個生態系統：中國雲端巨頭（阿里巴巴、騰訊、字節跳動）創建了自己的分支，近 1,000 人在騰訊深圳總部排隊安裝，甚至出現了每次安裝收費約 500 元人民幣的產業。

但 OpenClaw 的「上帝模式」自主性——shell 存取、檔案系統控制、瀏覽器自動化、通訊應用整合——也引發了嚴重的安全擔憂。代理已被透過提示注入攻擊，導致上傳加密貨幣金鑰、刪除程式碼庫和竊取財務資料。

本指南比較每個值得注意的替代方案，涵蓋最重要的類別：架構、安全性、訊息平台支援、LLM 靈活性、功能和社群成熟度。

Quick Reference: The Complete Comparison Matrix

Tool	Stars	Language	Messaging Platforms	LLM Providers	Security Model	License
OpenClaw	320K	TypeScript	23+ (WhatsApp, Telegram, Slack, Discord, Signal, iMessage, Teams, Matrix, LINE…)	20+ (Claude, GPT, Gemini, DeepSeek, Ollama…)	Pairing DM policy, configurable sandbox	MIT
Nanobot	34.2K	Python	11+ (Telegram, Discord, WhatsApp, Slack, Matrix, Email, QQ, DingTalk, Feishu, WeCom, Mochat)	12+ (OpenRouter, OpenAI, Anthropic, DeepSeek, Qwen, Ollama…)	Allow-lists, per-group policies, optional E2EE (Matrix)	MIT
ZeroClaw	27.6K	Rust	4 (Telegram, Discord, Slack, Matrix)	OpenAI, Anthropic, OpenRouter, custom endpoints	Workspace-scoped filesystem, forbidden paths, strict sandbox by default	MIT / Apache 2.0
PicoClaw	25.2K	Go	8 (Telegram, Discord, WhatsApp, Matrix, QQ, DingTalk, LINE, WeCom)	6+ (OpenAI, Claude, Gemini, OpenRouter, Volcengine, Zhipu)	Minimal (pre-v1.0)	MIT
NanoClaw	23.8K	TypeScript	5 (WhatsApp, Telegram, Discord, Slack, Gmail)	Claude (Anthropic-compatible endpoints)	Container isolation (Docker/Apple Container), per-group filesystem isolation	MIT
Khoj	33.5K	Python/TS	3 (Web, WhatsApp, Obsidian)	GPT, Claude, Gemini, DeepSeek, Llama, Qwen, Gemma, Mistral	Self-hosted option, cloud available	AGPL-3.0
CoPaw	12.4K	Python	10+ (DingTalk, Feishu, QQ, Discord, iMessage, Telegram, Matrix, Mattermost, MQTT)	DashScope, ModelScope, OpenAI-compat, llama.cpp, MLX, Ollama	Local-first, Tool Guard security layer	Apache 2.0
Eigent	13K	Python/TS	1 (Slack)	Local (vLLM, Ollama, LM Studio) + cloud models	Local-first, OAuth 2.0, data isolation	Open Source
PAI	10.1K	—	2 (Discord, ntfy push)	Claude (primary), multi-model architecture	User/system separation, granular policies, permission validation	Open Source
Open Interpreter	62.8K	Python	0 (Terminal/API only)	OpenAI, Claude, local models (Ollama, LM Studio)	Manual code approval, experimental safety mode	AGPL
AnythingLLM	56.4K	JS/TS	0 (Web UI, embeddable widget)	20+ (OpenAI, Claude, Gemini, Ollama, LM Studio, Groq…)	Self-hosted, optional telemetry, multi-user in Docker	MIT
LobeChat	73.8K	TypeScript	0 (Web/Desktop/PWA)	Multi-provider + local models	White-box memory, self-hosted option	Apache 2.0
LibreChat	34.7K	TypeScript	0 (Web UI)	10+ providers (Claude, GPT, Gemini, Ollama, DeepSeek…)	OAuth2, LDAP, multi-user, moderation	MIT
AutoGPT	183K	Python/TS	0	OpenAI (primary)	Self-hosted Docker, cloud beta	MIT + Polyform Shield
SuperAGI	17.3K	Python	0 (toolkit: Twitter, GitHub, Jira, Email, Calendar)	Custom models, external providers	Self-hosted, API key management	Open Source
Leon AI	17.1K	JS/Python	0 (Voice/text via own interface)	Local LLMs (experimental rewrite)	Full offline operation	MIT
Jan	41.1K	TS/Rust	0 (Desktop app)	HuggingFace models, OpenAI, Claude, Mistral, Groq	Complete local execution, no cloud required	Apache 2.0
Home Assistant Voice	85.4K	Python	0 (Voice satellites, web)	Ollama, OpenAI, cloud LLMs	Local control, self-hosted	Apache 2.0
OVOS	268	Python	0 (Voice-first)	Ollama, OpenAI-compatible APIs	Self-hosted, fully offline capable	Apache 2.0
BabyAGI	22.2K	Python	0	OpenAI	Local database for secrets	MIT
Botpress	14.6K	TypeScript	Multiple (via Hub integrations)	OpenAI (primary)	Workspace isolation, cloud platform	MIT
Rasa	21.1K	Python	10 (Messenger, Slack, Telegram, Twilio, Teams, Rocket.Chat…)	LLM via CALM (new)	Self-hosted, Apache license	Apache 2.0

快速參考：完整比較矩陣

工具	星數	語言	訊息平台	LLM 供應商	安全模型	授權
OpenClaw	320K	TypeScript	23+（WhatsApp、Telegram、Slack、Discord、Signal、iMessage、Teams、Matrix、LINE…）	20+（Claude、GPT、Gemini、DeepSeek、Ollama…）	配對 DM 策略、可配置沙箱	MIT
Nanobot	34.2K	Python	11+（Telegram、Discord、WhatsApp、Slack、Matrix、Email、QQ、DingTalk、飛書、企業微信、Mochat）	12+（OpenRouter、OpenAI、Anthropic、DeepSeek、Qwen、Ollama…）	允許列表、按群組策略、可選 E2EE（Matrix）	MIT
ZeroClaw	27.6K	Rust	4（Telegram、Discord、Slack、Matrix）	OpenAI、Anthropic、OpenRouter、自定端點	工作區範圍檔案系統、禁止路徑、預設嚴格沙箱	MIT / Apache 2.0
PicoClaw	25.2K	Go	8（Telegram、Discord、WhatsApp、Matrix、QQ、DingTalk、LINE、企業微信）	6+（OpenAI、Claude、Gemini、OpenRouter、火山引擎、智譜）	最小化（pre-v1.0）	MIT
NanoClaw	23.8K	TypeScript	5（WhatsApp、Telegram、Discord、Slack、Gmail）	Claude（Anthropic 相容端點）	容器隔離（Docker/Apple Container）、按群組檔案系統隔離	MIT
Khoj	33.5K	Python/TS	3（Web、WhatsApp、Obsidian）	GPT、Claude、Gemini、DeepSeek、Llama、Qwen、Gemma、Mistral	自託管選項、雲端可用	AGPL-3.0
CoPaw	12.4K	Python	10+（DingTalk、飛書、QQ、Discord、iMessage、Telegram、Matrix、Mattermost、MQTT）	DashScope、ModelScope、OpenAI 相容、llama.cpp、MLX、Ollama	本地優先、Tool Guard 安全層	Apache 2.0
Eigent	13K	Python/TS	1（Slack）	本地（vLLM、Ollama、LM Studio）+ 雲端模型	本地優先、OAuth 2.0、資料隔離	開源
PAI	10.1K	—	2（Discord、ntfy 推播）	Claude（主要）、多模型架構	使用者/系統分離、細粒度策略、權限驗證	開源
Open Interpreter	62.8K	Python	0（僅終端/API）	OpenAI、Claude、本地模型（Ollama、LM Studio）	手動代碼批准、實驗性安全模式	AGPL
AnythingLLM	56.4K	JS/TS	0（Web UI、可嵌入小工具）	20+（OpenAI、Claude、Gemini、Ollama、LM Studio、Groq…）	自託管、可選遙測、Docker 多用戶	MIT
LobeChat	73.8K	TypeScript	0（Web/桌面/PWA）	多供應商 + 本地模型	白盒記憶體、自託管選項	Apache 2.0
LibreChat	34.7K	TypeScript	0（Web UI）	10+ 供應商（Claude、GPT、Gemini、Ollama、DeepSeek…）	OAuth2、LDAP、多用戶、審核	MIT
AutoGPT	183K	Python/TS	0	OpenAI（主要）	自託管 Docker、雲端測試版	MIT + Polyform Shield
SuperAGI	17.3K	Python	0（工具包：Twitter、GitHub、Jira、Email、Calendar）	自定模型、外部供應商	自託管、API 金鑰管理	開源
Leon AI	17.1K	JS/Python	0（透過自有介面的語音/文字）	本地 LLM（實驗性重寫）	完全離線運作	MIT
Jan	41.1K	TS/Rust	0（桌面應用）	HuggingFace 模型、OpenAI、Claude、Mistral、Groq	完全本地執行、不需雲端	Apache 2.0
Home Assistant Voice	85.4K	Python	0（語音衛星、web）	Ollama、OpenAI、雲端 LLM	本地控制、自託管	Apache 2.0
OVOS	268	Python	0（語音優先）	Ollama、OpenAI 相容 API	自託管、完全離線能力	Apache 2.0
BabyAGI	22.2K	Python	0	OpenAI	本地資料庫存儲密鑰	MIT
Botpress	14.6K	TypeScript	多個（透過 Hub 整合）	OpenAI（主要）	工作區隔離、雲端平台	MIT
Rasa	21.1K	Python	10（Messenger、Slack、Telegram、Twilio、Teams、Rocket.Chat…）	LLM 透過 CALM（新版）	自託管、Apache 授權	Apache 2.0

Tier 1: Direct Competitors (Personal AI Assistants with Messaging Integration)

These are the tools most directly comparable to OpenClaw — they run on your machine, connect to LLMs, integrate with messaging apps, and can take autonomous actions.

第一層：直接競爭者（具備訊息整合的個人 AI 助手）

這些是與 OpenClaw 最直接可比的工具——它們在你的機器上運行、連接 LLM、整合通訊應用，並能採取自主行動。

1. OpenClaw — The Category King

GitHub: openclaw/openclaw · 320K stars · MIT License

What it is: A self-hosted, local-first personal AI agent that acts as your gateway — coordinating agents, messaging channels, and tools across all your connected devices. Created by Peter Steinberger, released November 2025, it went viral globally and especially in China where it triggered a national AI agent craze.

Architecture: WebSocket-based gateway running on ws://127.0.0.1:18789. Messages flow: Messaging Platform → Gateway → Agent Runtime (RPC mode) → Tools/Channels. Clients (CLI, native apps, device nodes) connect back to the gateway. YAML-based declarative configuration with per-channel routing.

Supported LLMs: 20+ providers — Claude (Opus 4.6, Sonnet 4), GPT-4, Gemini 2.0 Flash, DeepSeek, MiniMax M2.1, Kimi K2.5, plus local models via Ollama (Llama, Qwen, Mistral). Model switching via /model command or config file.

Messaging Platforms (23+): WhatsApp, Telegram, Slack, Discord, Signal, iMessage (via BlueBubbles), Google Chat, IRC, Microsoft Teams, Matrix, Feishu, LINE, Mattermost, Nextcloud Talk, Nostr, Synology Chat, Tlon, Twitch, Zalo, WebChat, plus native macOS/iOS/Android apps.

Key Capabilities:

Browser control (dedicated Chrome/Chromium instance)
Canvas + A2UI (agent-driven visual workspace)
Voice Wake (macOS/iOS), Talk Mode (Android continuous voice)
Camera/screen recording, location access, notifications
Cron jobs, webhooks, Gmail Pub/Sub automation
Device-local actions via Node system (macOS system.run/system.notify)
Android: SMS, contacts, calendar, motion sensor access

Security Model: Default DM policy uses pairing mode (unknown senders receive codes; approved via allowlist). Open DM mode requires explicit opt-in. Local gateway runs on loopback; remote access via Tailscale Serve/Funnel or SSH tunnels. Built-in doctor CLI tool surfaces risky configurations.

Plugin System: ClawHub — minimal skill registry with automatic discovery and installation. Three skill types: bundled, managed, workspace-level. Sessions coordination for agent-to-agent communication.

Community: 320K stars, 61.4K forks, 19,896 commits. Active Discord (discord.gg/clawd). Sponsors: OpenAI, Vercel, Blacksmith, Convex.

Pros:

Unmatched messaging platform coverage (23+ integrations)
Multi-device coordination (macOS/iOS/Android nodes)
Massive community and ecosystem momentum
Rich native apps with voice, camera, and canvas
ClawHub skill marketplace with growing catalog
Model-agnostic — works with virtually any LLM

Cons:

Security nightmare potential — prompt injection attacks have led to data exfiltration in the wild
Requires Node.js 22+ runtime (~1GB+ RAM baseline)
Single-user focus; not designed for teams
Setup complexity despite onboarding wizard
macOS app requires signed builds for permission persistence
Gateway architecture means a single point of failure

1. OpenClaw — 品類之王

GitHub: openclaw/openclaw · 320K 星 · MIT 授權

簡介： 自託管、本地優先的個人 AI 代理，作為你的閘道器——協調代理、訊息頻道和跨所有連接裝置的工具。由 Peter Steinberger 創建，2025 年 11 月發布，在全球爆紅，尤其在中國引發了全國性的 AI 代理熱潮。

架構： 基於 WebSocket 的閘道器運行在 ws://127.0.0.1:18789。訊息流：通訊平台 → 閘道器 → 代理運行時（RPC 模式）→ 工具/頻道。客戶端（CLI、原生應用、裝置節點）連回閘道器。基於 YAML 的聲明式配置，支援按頻道路由。

支援的 LLM： 20+ 供應商——Claude（Opus 4.6、Sonnet 4）、GPT-4、Gemini 2.0 Flash、DeepSeek、MiniMax M2.1、Kimi K2.5，加上透過 Ollama 的本地模型（Llama、Qwen、Mistral）。透過 /model 指令或配置檔案切換模型。

通訊平台（23+）： WhatsApp、Telegram、Slack、Discord、Signal、iMessage（透過 BlueBubbles）、Google Chat、IRC、Microsoft Teams、Matrix、飛書、LINE、Mattermost、Nextcloud Talk、Nostr、Synology Chat、Tlon、Twitch、Zalo、WebChat，加上原生 macOS/iOS/Android 應用。

主要功能：

瀏覽器控制（專用 Chrome/Chromium 實例）
Canvas + A2UI（代理驅動的視覺工作區）
語音喚醒（macOS/iOS）、對話模式（Android 持續語音）
相機/螢幕錄製、位置存取、通知
定時任務、webhooks、Gmail Pub/Sub 自動化
透過 Node 系統的裝置本地操作（macOS system.run/system.notify）
Android：簡訊、聯絡人、行事曆、動作感測器存取

安全模型： 預設 DM 策略使用配對模式（未知發送者收到驗證碼；透過允許列表批准）。開放 DM 模式需明確啟用。本地閘道器運行在回環介面；透過 Tailscale Serve/Funnel 或 SSH 隧道進行遠端存取。內建 doctor CLI 工具顯示風險配置。

外掛系統： ClawHub——最小化技能註冊中心，支援自動發現和安裝。三種技能類型：內建、管理的、工作區級別。會話協調支援代理間通訊。

社群： 320K 星、61.4K 分支、19,896 次提交。活躍的 Discord（discord.gg/clawd）。贊助商：OpenAI、Vercel、Blacksmith、Convex。

優點：

無與倫比的通訊平台覆蓋（23+ 整合）
多裝置協調（macOS/iOS/Android 節點）
龐大的社群和生態系統動能
豐富的原生應用支援語音、相機和畫布
ClawHub 技能市場持續增長
模型無關——幾乎適用於任何 LLM

缺點：

潛在的安全噩夢——提示注入攻擊已在實際使用中導致資料外洩
需要 Node.js 22+ 運行時（~1GB+ RAM 基準）
單用戶專注；不適合團隊
儘管有入門精靈，設置仍然複雜
macOS 應用需要簽名構建以持久化權限
閘道器架構意味著單點故障

2. Nanobot — The Lightweight Research-Ready Alternative

GitHub: HKUDS/nanobot · 34.2K stars · MIT License

What it is: An ultra-lightweight personal AI assistant from the University of Hong Kong, delivering core agent functionality in ~4,000 lines of Python — 99% fewer lines than OpenClaw. Launched February 2026, it's the fastest-growing direct competitor.

Architecture: Modular Python design with three layers: Providers (pluggable LLM backends with 2-step integration), Channels (messaging platform adapters), and Agent Core (reasoning + tool orchestration). Token-based memory system with session history. MCP integration for external tool discovery.

Supported LLMs (12+): OpenRouter (recommended), OpenAI, Anthropic, Azure OpenAI, DeepSeek, Qwen, Moonshot/Kimi, MiniMax, Mistral, vLLM, Ollama, VolcEngine. Auto-detection or pin specific provider.

Messaging Platforms (11+): Telegram, Discord, WhatsApp (local bridge, QR login), Slack (socket mode), Matrix/Element (WebSocket + E2EE), Email (IMAP/SMTP), QQ, DingTalk (stream), Feishu (WebSocket), WeCom, Mochat (Socket.IO). Channel plugin system enables custom integrations.

Key Capabilities:

24/7 real-time market analysis and web search
Full-stack software engineering
Smart daily routine management and scheduling
Personal knowledge assistant with memory
Media file handling across platforms
MCP tool integration
Progress streaming and tool-call logging
Experimental thinking mode

Security: Allow-lists, per-group policies (mention/open/allowlist modes), optional E2EE via Matrix. WebSocket-based channels require no public IP.

Plugin System: MCP-native support for external tools. Channel plugins for custom messaging integrations (still developing, not yet in PyPI release).

Community: 34.2K stars, 5.7K forks, 291 open issues. Feishu group, WeChat group, Discord community.

Pros:

Entire codebase readable in a few hours — ideal for research and customization
11+ messaging platforms rivaling OpenClaw's coverage
No public IP needed for most channels (WebSocket-based)
12+ LLM provider support with auto-detection
MCP-native extensibility
Minimal resource requirements vs OpenClaw

Cons:

Very young project (launched Feb 2026) with evolving APIs
Some enterprise integrations (QQ, DingTalk) in sandbox/beta
No sandboxing by default — executes with full user permissions
Thin plugin ecosystem compared to ClawHub
Requires Python 3.11+; WhatsApp needs Node.js 18+
Documentation spread across multiple formats

2. Nanobot — 輕量級研究友好替代方案

GitHub: HKUDS/nanobot · 34.2K 星 · MIT 授權

簡介： 來自香港大學的超輕量級個人 AI 助手，以約 4,000 行 Python 程式碼提供核心代理功能——比 OpenClaw 少 99% 的程式碼。2026 年 2 月推出，是成長最快的直接競爭者。

架構： 模組化 Python 設計，三層結構：提供者（可插拔 LLM 後端，2 步整合）、頻道（通訊平台適配器）和代理核心（推理 + 工具編排）。基於 token 的記憶系統，支援會話歷史。MCP 整合支援外部工具發現。

支援的 LLM（12+）： OpenRouter（推薦）、OpenAI、Anthropic、Azure OpenAI、DeepSeek、Qwen、Moonshot/Kimi、MiniMax、Mistral、vLLM、Ollama、火山引擎。支援自動偵測或指定特定供應商。

通訊平台（11+）： Telegram、Discord、WhatsApp（本地橋接，QR 登入）、Slack（socket 模式）、Matrix/Element（WebSocket + E2EE）、Email（IMAP/SMTP）、QQ、DingTalk（串流）、飛書（WebSocket）、企業微信、Mochat（Socket.IO）。頻道外掛系統支援自定整合。

優點：

整個程式碼庫幾小時內可讀完——研究和自定的理想選擇
11+ 通訊平台可與 OpenClaw 的覆蓋範圍匹敵
大多數頻道不需公共 IP（基於 WebSocket）
12+ LLM 供應商支援，含自動偵測
MCP 原生可擴展性
相比 OpenClaw 的最小資源需求

缺點：

非常年輕的專案（2026 年 2 月推出），API 仍在演進
部分企業整合（QQ、DingTalk）處於沙箱/測試階段
預設無沙箱——以完整用戶權限執行
相比 ClawHub 的外掛生態系統較薄弱
需要 Python 3.11+；WhatsApp 需要 Node.js 18+
文件分散在多種格式中

3. NanoClaw — The Security-First Containerized Agent

GitHub: gavrielc/nanoclaw · 23.8K stars · MIT License

What it is: A lightweight AI assistant framework that executes Claude agents in isolated containers. Positioned as the security-focused alternative to OpenClaw — "the agent can only mess up its container."

Architecture: Single Node.js process with channel self-registration via Claude Code skills. SQLite database for messages, groups, sessions, state. Filesystem-based IPC for container communication. Flow: Channels → SQLite → Polling loop → Container (Claude Agent SDK) → Response.

Supported LLMs: Claude (primary), any Anthropic-compatible endpoint (Ollama, Together AI, Fireworks via env vars).

Messaging Platforms: WhatsApp, Telegram, Discord, Slack, Gmail. Channels added via Claude Code skills (/add-whatsapp, /add-telegram).

Key Capabilities:

Per-group isolated filesystems and container sandboxes
Scheduled/recurring tasks with cron-like scheduling
Web access (search and content fetching)
Agent swarms for collaborative task execution
Per-group message queuing with concurrency control
Per-group CLAUDE.md memory files
Main channel for admin control

Security: Docker Sandboxes (micro VM isolation), Apple Container (macOS), or Docker runtime. Per-group filesystem isolation. This is NanoClaw's defining feature.

Plugin System: Contributors submit Claude Code skills rather than features to core. Skills are branch-based extensions merged into personal forks.

Community: 23.8K stars, 6.4K forks, 310 commits.

Pros:

Best-in-class security through container isolation
Simple, auditable codebase — "Customization = code changes"
Per-group isolation prevents cross-contamination
Apple Container support on macOS (no Docker needed)
AI-native interaction model (no dashboards; ask Claude)

Cons:

Claude-only by default (limited model flexibility)
Reduced flexibility due to container constraints
No skills marketplace — DIY fork-and-merge model
Limited platform integrations (5 vs OpenClaw's 23+)
TypeScript codebase but simpler than OpenClaw
Constrained hardware access inside containers

3. NanoClaw — 安全優先的容器化代理

GitHub: gavrielc/nanoclaw · 23.8K 星 · MIT 授權

簡介： 在隔離容器中執行 Claude 代理的輕量級 AI 助手框架。定位為 OpenClaw 的安全優先替代方案——「代理只能搞壞它自己的容器。」

架構： 單一 Node.js 進程，透過 Claude Code 技能進行頻道自注冊。SQLite 資料庫存儲訊息、群組、會話、狀態。基於檔案系統的 IPC 用於容器通訊。流程：頻道 → SQLite → 輪詢循環 → 容器（Claude Agent SDK）→ 回應。

支援的 LLM： Claude（主要），任何 Anthropic 相容端點（透過環境變數的 Ollama、Together AI、Fireworks）。

通訊平台： WhatsApp、Telegram、Discord、Slack、Gmail。透過 Claude Code 技能添加頻道。

優點：

透過容器隔離的最佳安全性
簡潔、可審計的程式碼庫
按群組隔離防止交叉污染
macOS 上的 Apple Container 支援
AI 原生互動模型

缺點：

預設僅支援 Claude（模型靈活性有限）
容器限制降低靈活性
無技能市場——DIY 分支合併模型
有限的平台整合（5 個 vs OpenClaw 的 23+）
容器內硬體存取受限

4. ZeroClaw — The Performance-Obsessed Rust Runtime

GitHub: zeroclaw-labs/zeroclaw · 27.6K stars · MIT / Apache 2.0

What it is: A runtime OS for agentic workflows built entirely in Rust. It abstracts models, tools, memory, and execution into a trait-driven architecture with an ~8.8MB binary, <5MB RAM at runtime, and <10ms startup time.

Architecture: Trait-driven Rust design with swappable providers, channels, tools, and memory systems. Three execution models: CLI (one-shot), Gateway (webhook-based), Daemon (autonomous runtime). Optional Docker sandboxing.

Supported LLMs: OpenAI (Codex), Anthropic (Claude), OpenRouter, custom/pluggable endpoints. Multi-account auth profiles with encryption at rest.

Messaging Platforms: Telegram, Discord, Slack, Matrix (with E2EE support).

Key Capabilities:

Workspace-scoped filesystem access by default
Forbidden paths pre-blocked (.ssh, .aws, .gnupg)
Command execution with allowlists
Provider switching via config
Memory abstraction with swappable backends

Security: Secure-by-default philosophy. Workspace-only filesystem scoping enabled by default. Forbidden paths pre-blocked. Strict sandboxing with explicit allowlists. Pairing mode for DM authentication.

Performance vs Competitors:

Metric	ZeroClaw	OpenClaw	PicoClaw
RAM	<5MB	>1GB	<10MB
Startup	<10ms	>500ms	<1s
Binary	8.8MB	N/A (Node.js)	Single binary

Community: 27.6K stars, 3.7K forks, 1,876 commits.

Pros:

Best performance profile of any agent runtime (5MB RAM, 10ms startup)
Security-by-default with pre-blocked sensitive paths
Zero vendor lock-in — fully pluggable architecture
Dual MIT/Apache license
Encrypted auth profiles at rest

Cons:

Rust toolchain required for building from source
Only 4 messaging platforms (vs OpenClaw's 23+)
Smaller plugin/skill ecosystem
No native mobile apps
Smaller community than OpenClaw/Nanobot
No compliance certifications

4. ZeroClaw — 效能極致的 Rust 運行時

GitHub: zeroclaw-labs/zeroclaw · 27.6K 星 · MIT / Apache 2.0

簡介： 完全用 Rust 構建的代理工作流運行時作業系統。將模型、工具、記憶和執行抽象為 trait 驅動的架構，~8.8MB 二進制檔案、<5MB RAM 運行時、<10ms 啟動時間。

架構： trait 驅動的 Rust 設計，支援可交換的供應商、頻道、工具和記憶系統。三種執行模型：CLI（一次性）、閘道器（基於 webhook）、守護進程（自主運行時）。可選 Docker 沙箱。

效能對比：

指標	ZeroClaw	OpenClaw	PicoClaw
RAM	<5MB	>1GB	<10MB
啟動	<10ms	>500ms	<1s
二進制	8.8MB	N/A（Node.js）	單一二進制

優點：

所有代理運行時中最佳效能（5MB RAM、10ms 啟動）
預設安全，預先封鎖敏感路徑
零供應商鎖定——完全可插拔架構

缺點：

從原始碼構建需要 Rust 工具鏈
僅 4 個通訊平台（vs OpenClaw 的 23+）
較小的外掛/技能生態系統
無原生行動應用

5. PicoClaw — The $10 Hardware Agent

GitHub: sipeed/picoclaw · 25.2K stars · MIT License

What it is: Ultra-lightweight personal AI assistant refactored from Nanobot in Go. Designed to run on $10 RISC-V hardware (LicheeRV-Nano) with <10MB RAM and 1-second boot time. 95% AI-generated codebase with human-in-the-loop refinement.

Architecture: Single self-contained Go binary. Three components: picoclaw (agent), picoclaw-launcher, picoclaw-launcher-tui. Supports RISC-V, ARM64, MIPS, x86 architectures.

Supported LLMs: OpenAI (GPT-5.4), Anthropic Claude (Sonnet 4.6), Google Gemini, OpenRouter, Volcengine/Ark Code, Zhipu.

Messaging Platforms (8): Telegram, Discord, WhatsApp, Matrix, QQ, DingTalk, LINE, WeCom.

Key Capabilities:

Full-stack code generation and deployment
Web search (Brave, Tavily, Perplexity, DuckDuckGo, SearXNG)
Task scheduling and automation
Memory management
File operations
Logging and planning management
Free Groq Whisper voice transcription

Community: 25.2K stars, 3.4K forks. Launched February 2026. Seeking community maintainers.

Pros:

Runs on $10 hardware — democratizes AI assistants for IoT/edge
Single binary, zero dependencies
Cross-architecture (RISC-V, ARM, MIPS, x86)
8 messaging platforms — strong for a lightweight tool
Includes voice transcription (Groq Whisper)

Cons:

Pre-v1.0 — not recommended for production
Narrow integration library
No marketplace equivalent
Minimal security model
Limited documentation
Seeking maintainers suggests sustainability risk

5. PicoClaw — $10 硬體代理

GitHub: sipeed/picoclaw · 25.2K 星 · MIT 授權

簡介： 從 Nanobot 用 Go 重構的超輕量級個人 AI 助手。設計在 $10 RISC-V 硬體上運行，<10MB RAM，1 秒啟動。95% AI 生成的程式碼庫。

通訊平台（8）： Telegram、Discord、WhatsApp、Matrix、QQ、DingTalk、LINE、企業微信。

優點：

在 $10 硬體上運行——為 IoT/邊緣裝置普及 AI 助手
單一二進制，零依賴
跨架構（RISC-V、ARM、MIPS、x86）

缺點：

Pre-v1.0——不建議用於生產環境
最小化安全模型
正在尋找維護者，存在可持續性風險

6. CoPaw — The China-Focused Multi-Channel Assistant

GitHub: agentscope-ai/CoPaw · 12.4K stars · Apache 2.0

What it is: A locally-deployable personal AI assistant with emphasis on Chinese messaging platforms. Supports social digests, newsletter integration, document management, and overnight task execution.

Architecture: Python backend with React/Node.js frontend. Web Console UI at localhost:8088. Multiple installation paths: pip, automated script, desktop app (beta), Docker, ModelScope Studio, Alibaba Cloud ECS.

Supported LLMs: DashScope (Alibaba), ModelScope, OpenAI-compatible APIs, llama.cpp (cross-platform), MLX (Apple Silicon), Ollama, LM Studio.

Messaging Platforms (10+): DingTalk, Feishu, QQ, Discord, iMessage, Telegram, Matrix, Mattermost, MQTT, and more via plugins.

Key Capabilities:

Social digests from Xiaohongshu, Zhihu, Reddit, Bilibili, YouTube
Newsletter and email/calendar contact management
Document organization, reading, and summarization
Personal knowledge base and tech news tracking
Overnight task execution with draft delivery
Tool Guard security layer with @mention filtering
LLM auto-retry with backoff and token usage tracking

Security: Local-first operation, Tool Guard security layer (v0.0.7), shell confirmation, tool sandboxing on roadmap.

Community: 12.4K stars, 1.5K forks, 386 open issues.

Pros:

Best Chinese platform support (DingTalk, Feishu, QQ, WeChat ecosystem)
Works entirely locally with open-source models (no API key required)
Multiple installation paths including desktop app
Social digest feature unique among competitors
Apache 2.0 license (commercially friendly)

Cons:

Desktop app in beta with potential stability issues
Documentation spread across multiple channels
Feature breadth over depth — varying quality per platform
Active development means frequent breaking changes
Smaller English-language community

6. CoPaw — 以中國市場為重心的多頻道助手

GitHub: agentscope-ai/CoPaw · 12.4K 星 · Apache 2.0

簡介： 可本地部署的個人 AI 助手，重點支援中國通訊平台。支援社交摘要、電子報整合、文件管理和過夜任務執行。

通訊平台（10+）： DingTalk、飛書、QQ、Discord、iMessage、Telegram、Matrix、Mattermost、MQTT 及更多。

優點：

最佳中國平台支援（DingTalk、飛書、QQ、微信生態）
完全本地運行，使用開源模型（不需 API 金鑰）
社交摘要功能在競爭者中獨一無二

缺點：

桌面應用處於測試版
功能廣度優於深度——各平台品質不一
較小的英語社群

7. Khoj — The AI Second Brain

GitHub: khoj-ai/khoj · 33.5K stars · AGPL-3.0

What it is: A personal AI application that functions as a "second brain" — chat with LLMs, search your documents, create custom agents, and schedule automations. Bridges on-device and enterprise-scale deployments.

Architecture: Python (50.9%) + TypeScript (36.2%) backend. Docker-based deployment. Cloud version at app.khoj.dev for zero-setup access.

Supported LLMs: GPT, Claude, Gemini, DeepSeek (online); Llama3, Qwen, Gemma, Mistral (local).

Messaging Platforms: WhatsApp, Obsidian plugin, Web browser, Desktop, Phone apps.

Key Capabilities:

Document processing (PDFs, Markdown, Notion, Word, org-mode)
Internet search for real-time answers
Custom agent creation with knowledge bases and personas
Scheduled tasks, newsletter generation, smart notifications
Semantic document retrieval
Image generation, text-to-speech, voice playback
Deep research mode

Security: Self-hosting for complete privacy. Cloud version available for convenience. Enterprise options: cloud, on-premises, or hybrid.

Plugin System: Custom agents with persona, knowledge, and tool customization.

Community: 33.5K stars, 2.1K forks, 5,150+ commits, 170+ releases.

Pros:

Best document/knowledge search capabilities
Dual deployment: self-hosted or cloud (app.khoj.dev)
Excellent LLM diversity (online + local)
Custom agent builder with knowledge base integration
Enterprise-ready with hybrid deployment
Scheduled automations and proactive notifications

Cons:

AGPL-3.0 license restricts commercial use without source disclosure
Limited messaging platform support (WhatsApp + web only, no Telegram/Discord/Slack)
Beta status (v2.0.0-beta.25) suggests stability considerations
No shell access or file system write capabilities
More "search and chat" than "autonomous agent"
Less action-oriented than OpenClaw — primarily a knowledge tool

7. Khoj — AI 第二大腦

GitHub: khoj-ai/khoj · 33.5K 星 · AGPL-3.0

簡介： 作為「第二大腦」的個人 AI 應用——與 LLM 聊天、搜尋文件、建立自定代理、排程自動化。

優點：

最佳文件/知識搜尋能力
雙重部署：自託管或雲端
自定代理建構器整合知識庫
企業級混合部署

缺點：

AGPL-3.0 授權限制商業使用
通訊平台支援有限（WhatsApp + web）
更偏向「搜尋和聊天」而非「自主代理」

Tier 2: Action-Capable Assistants (Without Messaging Integration)

These tools can take actions (run code, browse the web, manage files) but lack native messaging platform integration. You interact through a terminal, web UI, or desktop app.

第二層：具行動能力的助手（無通訊整合）

這些工具能採取行動（執行程式碼、瀏覽網頁、管理檔案），但缺乏原生通訊平台整合。你透過終端、Web UI 或桌面應用互動。

8. Open Interpreter — The OG Local Code Executor

GitHub: OpenInterpreter/open-interpreter · 62.8K stars · AGPL

What it is: A natural language interface that lets LLMs run code on your computer. The original "ChatGPT Code Interpreter but on your local machine" — predating OpenClaw by over a year.

Architecture: Function-calling language models with exec() capability. Streams model messages, code, and system outputs as Markdown. Optional FastAPI server for HTTP endpoints.

Supported LLMs: OpenAI (GPT-4, GPT-3.5-turbo), Anthropic Claude, local models (LM Studio, Jan.ai, Ollama). Uses LiteLLM for provider flexibility.

Messaging Platforms: None. Terminal CLI and Python API only.

Key Capabilities:

Multi-language code execution (Python, JavaScript, Shell)
File/media manipulation (photos, videos, PDFs)
Browser automation (Chrome control)
Data analysis and plotting
Unlimited runtime, file size, internet access (unlike ChatGPT)
Package installation without restrictions

Security: Manual code approval before execution (unless auto_run=True). Experimental safety mode. No sandboxing by default.

Pros:

Mature and battle-tested (62.8K stars, large community)
No restrictions on runtime, file size, or packages
Works with any LLM via LiteLLM
Python API for custom integrations
Strong documentation in multiple languages

Cons:

No messaging platform integration at all
Terminal-only interaction (no web UI, no mobile)
AGPL license restricts commercial derivative works
No persistent memory across sessions
Security relies entirely on user's manual approval
Not an "always-on" assistant — session-based only

8. Open Interpreter — 本地程式碼執行的先驅

GitHub: OpenInterpreter/open-interpreter · 62.8K 星 · AGPL

簡介： 讓 LLM 在你的電腦上執行程式碼的自然語言介面。原始的「本地版 ChatGPT Code Interpreter」——比 OpenClaw 早一年多。

優點： 成熟且經過驗證、無執行限制、透過 LiteLLM 支援任何 LLM。 缺點： 無通訊平台整合、僅終端互動、AGPL 授權、無持久記憶。

9. Eigent — The Multi-Agent Desktop Workforce

GitHub: eigent-ai/eigent · 13K stars · Open Source

What it is: An open-source desktop application that deploys coordinated teams of specialized AI agents working in parallel — one browsing the web, one processing documents, one writing emails. Built on the CAMEL-AI multi-agent framework. Positioned as a local alternative to Claude Cowork.

Architecture: FastAPI backend + React/Electron frontend. Multi-agent framework: CAMEL-AI. Agent types: Developer (code execution), Browser (web search), Document (content creation), Multi-Modal (image/audio). MCP integration for tooling.

Supported LLMs: Local inference (vLLM, Ollama, LM Studio) + cloud models. Custom model support.

Messaging Platforms: Slack only (for sending summaries and notifications).

Key Capabilities:

Dynamic parallel task execution across specialized agents
Trip planning, financial reports, market research
SEO audits, competitive analysis
File system operations, PDF manipulation
Human-in-the-loop intervention during workflows
MCP tools for web, code, Notion, Google suite, Slack

Pros:

True multi-agent parallel workflows (unique advantage)
Top-performing on GAIA benchmark
Local-first with zero data leaving your machine
100% open source from day one
MCP ecosystem for extensibility

Cons:

Minimal messaging integration (Slack only)
Electron app — heavier resource footprint
Younger project with evolving APIs
No always-on daemon mode — task-oriented
Self-hosted setup requires technical knowledge

9. Eigent — 多代理桌面工作團隊

GitHub: eigent-ai/eigent · 13K 星 · 開源

簡介： 部署協調的專業 AI 代理團隊平行工作的開源桌面應用——一個瀏覽網頁、一個處理文件、一個撰寫郵件。建構於 CAMEL-AI 多代理框架。

優點： 真正的多代理平行工作流、GAIA 基準測試頂級表現、本地優先。 缺點： 最小通訊整合（僅 Slack）、Electron 應用較重、無常駐守護進程模式。

10. PAI (Personal AI Infrastructure) — The Goal-Oriented System

GitHub: danielmiessler/Personal_AI_Infrastructure · 10.1K stars · Open Source

What it is: An agentic AI platform by Daniel Miessler designed to magnify human capabilities through persistent, learning-based assistance. Uses a unique TELOS system (10 identity documents: MISSION.md, GOALS.md, PROJECTS.md, etc.) to deeply understand user goals.

Architecture: Built on Claude as primary model. Three-tier memory (hot/warm/cold). UNIX philosophy: composable tools, CLI-first. Skill hierarchy: CODE → CLI → PROMPTS → SKILLS. GUI installer for non-technical users.

Supported LLMs: Claude (primary), multi-model architecture planned.

Messaging/Notifications: Discord integration, ntfy push notifications, ElevenLabs TTS voice.

Key Capabilities:

TELOS system for deep goal understanding
Three-tier memory with learning from every interaction
User/system separation with granular security policies
Permission validation before tool execution
Blocks dangerous operations while enabling normal workflows
Event-driven hooks system (8 lifecycle event types)
ElevenLabs voice integration

Pros:

Unique goal-oriented design — learns your mission, beliefs, strategies
Strongest security policy system among personal AI tools
Three-tier memory architecture
Upgrade-safe customization (USER/ vs SYSTEM/ separation)
GUI installer for non-technical users

Cons:

Claude-only (no multi-model support yet)
Minimal messaging integration (Discord + push only)
v4.0.3 with breaking changes expected
Opinionated philosophy may not suit all users
Terminal-based primary interface
Single-model dependency

10. PAI（Personal AI Infrastructure）— 目標導向系統

GitHub: danielmiessler/Personal_AI_Infrastructure · 10.1K 星 · 開源

簡介： Daniel Miessler 設計的代理 AI 平台，透過持久性、基於學習的協助來放大人類能力。使用獨特的 TELOS 系統深入了解用戶目標。

優點： 獨特的目標導向設計、最強安全策略系統、三層記憶架構。 缺點： 僅支援 Claude、最小通訊整合、有破壞性更新。

11. AnythingLLM — The Universal LLM Hub

GitHub: mintplex-labs/anything-llm · 56.4K stars · MIT License

What it is: A full-stack, self-hosted AI application and "all-in-one AI productivity accelerator." Transforms documents into contextual references for LLM interactions via workspace-based organization.

Architecture: Monorepo: Vite+React frontend, Node.js Express backend, document collector service. Docker primary deployment with multi-user support.

Supported LLMs (20+): OpenAI, Azure OpenAI, AWS Bedrock, Anthropic, Google Gemini, NVIDIA NIM, Ollama, LM Studio, LocalAI, Together AI, Groq, Mistral, DeepSeek, Perplexity, OpenRouter, Cohere, Kobold CPP, LiteLLM, xAI, llama.cpp.

Messaging Platforms: None native. Embeddable chat widget for websites (Docker only).

Key Capabilities:

MCP compatibility for external tools
No-code agent builder
Multi-modal support (vision across LLMs)
Document processing (PDF, TXT, DOCX, etc.)
RAG with 9+ vector database options
Workspace-embedded agents with internet access
Developer API for custom integrations
Desktop apps (Mac, Windows, Linux)

Pros:

Widest LLM provider support (20+)
Best document/RAG capabilities
MIT license — commercially friendly
Multiple deployment options (Docker, cloud, bare metal, desktop)
Multi-user support with permissions
Active 56K+ star community

Cons:

No messaging platform integration
Web UI only — no mobile apps
Multi-user features require Docker
No autonomous "always-on" capability
More of a chat/RAG platform than a personal assistant
Cannot take actions on your behalf (no shell, no file write outside workspace)

11. AnythingLLM — 通用 LLM 中心

GitHub: mintplex-labs/anything-llm · 56.4K 星 · MIT 授權

簡介： 全棧自託管 AI 應用和「一體化 AI 生產力加速器」。透過工作區組織將文件轉換為 LLM 互動的上下文參考。

優點： 最廣泛的 LLM 供應商支援（20+）、最佳文件/RAG 能力、MIT 授權。 缺點： 無通訊平台整合、無自主「常駐」能力、更偏向聊天/RAG 平台。

12. AutoGPT — The Pioneer (Now a Platform)

GitHub: Significant-Gravitas/AutoGPT · 183K stars · MIT + Polyform Shield

What it is: Originally the first viral autonomous AI agent (March 2023), now evolved into a low-code agent platform for designing, deploying, and managing AI workflows.

Architecture: React frontend + Python backend. Block-based workflow system where each block performs a single action. Docker containerized. Requires 4+ CPU cores, 8GB+ RAM, 10GB storage.

Supported LLMs: OpenAI (primary). Limited documentation on other providers.

Messaging Platforms: None documented.

Key Capabilities:

Low-code agent builder with visual workflow editor
Block-based automation system
Agent marketplace with pre-built agents
Monitoring and analytics dashboard
Continuous autonomous agent execution

Pros:

Massive community (183K stars — largest in category)
Visual workflow builder — accessible to non-developers
Self-hosted free option
Agent marketplace for pre-built solutions

Cons:

Newer platform code uses Polyform Shield license (not truly open source)
No messaging platform integration
Limited LLM provider documentation
Complex Docker-based setup
Cloud offering still in closed beta
Shifted far from original "autonomous agent" vision

12. AutoGPT — 先驅者（現為平台）

GitHub: Significant-Gravitas/AutoGPT · 183K 星 · MIT + Polyform Shield

簡介： 最初是第一個病毒式自主 AI 代理（2023 年 3 月），現已演變為低代碼代理平台。

優點： 龐大社群（183K 星）、視覺化工作流建構器、代理市場。 缺點： 新平台程式碼使用 Polyform Shield 授權、無通訊整合、已偏離原始自主代理願景。

Tier 3: Chat UIs and LLM Frontends

These are excellent LLM interfaces but primarily serve as chat/conversation tools. They lack the autonomous action-taking and messaging platform integration that defines OpenClaw's category.

第三層：聊天 UI 和 LLM 前端

這些是優秀的 LLM 介面，但主要作為聊天/對話工具。它們缺乏定義 OpenClaw 類別的自主行動和通訊平台整合。

13. LobeChat — 73.8K Stars

GitHub: lobehub/lobe-chat · Apache 2.0

The most popular open-source chat UI. Agent builder, 10,000+ MCP plugins, CoT visualization, branching conversations, artifacts, TTS/STT. Desktop apps + PWA. No messaging integration, no autonomous actions. Best for: teams wanting a polished multi-model chat interface with agent capabilities.

14. LibreChat — 34.7K Stars

GitHub: danny-avila/LibreChat · MIT

Enhanced ChatGPT clone with 10+ provider support, sandboxed code interpreter (Python, Node, Go, Java, Rust, etc.), MCP integration, web search, image generation, multi-user with OAuth2/LDAP. Best for: organizations needing a self-hosted ChatGPT replacement with enterprise auth.

15. Jan — 41.1K Stars

GitHub: janhq/jan · Apache 2.0

Desktop app for running local LLMs. Downloads models from HuggingFace, provides OpenAI-compatible API at localhost:1337. Built with Tauri (TypeScript + Rust). MCP support. No messaging integration. Best for: users wanting a simple local-first LLM desktop experience.

13. LobeChat — 73.8K 星

GitHub: lobehub/lobe-chat · Apache 2.0

最受歡迎的開源聊天 UI。代理建構器、10,000+ MCP 外掛、CoT 視覺化、分支對話、artifacts、TTS/STT。桌面應用 + PWA。無通訊整合、無自主行動。

14. LibreChat — 34.7K 星

GitHub: danny-avila/LibreChat · MIT

增強版 ChatGPT 克隆，支援 10+ 供應商、沙箱程式碼解譯器、MCP 整合、網頁搜尋、圖像生成、OAuth2/LDAP 多用戶。

15. Jan — 41.1K 星

GitHub: janhq/jan · Apache 2.0

執行本地 LLM 的桌面應用。從 HuggingFace 下載模型，提供 OpenAI 相容 API。使用 Tauri 構建。無通訊整合。

Tier 4: Autonomous Agent Frameworks

These are developer frameworks for building autonomous agents, not end-user personal assistants.

第四層：自主代理框架

這些是用於建構自主代理的開發者框架，而非面向終端用戶的個人助手。

16. SuperAGI — 17.3K Stars

GitHub: TransformerOptimus/SuperAGI · Open Source

Framework for building autonomous agents with a marketplace of 20+ toolkits (Twitter, GitHub, Jira, Google Search, DALL-E, Instagram, Notion, Email, Google Calendar). Web GUI, concurrent agent execution, long-term memory. Self-hosted via Docker or cloud at app.superagi.com. Best for: developers building task-specific autonomous agents.

17. BabyAGI — 22.2K Stars

GitHub: yoheinakajima/babyagi · MIT

Experimental self-building autonomous agent using a database-backed function registry. Novel concept: agents that build their own capabilities via functionz framework. Flask dashboard. Not production-ready — built "by someone who has never held a job as a developer." Best for: AI researchers exploring self-building agent architectures.

16. SuperAGI — 17.3K 星

GitHub: TransformerOptimus/SuperAGI · 開源

構建自主代理的框架，擁有 20+ 工具包市場。Web GUI、並行代理執行、長期記憶。適合：構建任務特定自主代理的開發者。

17. BabyAGI — 22.2K 星

GitHub: yoheinakajima/babyagi · MIT

實驗性自我建構自主代理。新穎概念：代理透過 functionz 框架建構自己的能力。非生產就緒。適合：探索自建構代理架構的 AI 研究者。

Tier 5: Conversational AI Platforms (Chatbot Builders)

These are platforms for building chatbots at scale, not personal AI assistants.

第五層：對話式 AI 平台（聊天機器人建構器）

這些是用於大規模建構聊天機器人的平台，而非個人 AI 助手。

18. Botpress — 14.6K Stars

GitHub: botpress/botpress · MIT

GPT/LLM-powered chatbot builder with SDK, CLI, and Botpress Hub for integrations. Cloud-based Studio for visual bot building. Good messaging platform support via Hub integrations. Best for: businesses building customer-facing chatbots (not personal assistant use).

19. Rasa — 21.1K Stars

GitHub: RasaHQ/rasa · Apache 2.0

Now in maintenance mode. Legacy ML-based conversational AI framework with 10+ messaging platform integrations (Messenger, Slack, Telegram, Twilio, Teams). Succeeded by Hello Rasa and CALM engine. Best for: existing Rasa users; new users should look at Hello Rasa instead.

18. Botpress — 14.6K 星

GitHub: botpress/botpress · MIT

GPT/LLM 驅動的聊天機器人建構器。雲端 Studio 視覺化建構。透過 Hub 整合支援通訊平台。適合：建構客戶面向聊天機器人的企業。

19. Rasa — 21.1K 星

GitHub: RasaHQ/rasa · Apache 2.0

現已進入維護模式。 傳統 ML 對話式 AI 框架，支援 10+ 通訊平台。已被 Hello Rasa 和 CALM 引擎取代。

Tier 6: Voice-First Assistants

Open-source voice assistants focused on smart home and voice interaction.

第六層：語音優先助手

專注於智慧家庭和語音互動的開源語音助手。

20. Home Assistant Voice — 85.4K Stars

GitHub: home-assistant/core · Apache 2.0

The dominant open-source smart home platform with voice assistant capabilities via Assist Pipeline. Supports LLM integration (Ollama, OpenAI), local STT/TTS via Piper, wake word detection, multilingual assistants, and voice satellites (Voice PE hardware). 4,662+ contributors. Not a personal AI assistant per se — focused on home automation with voice as one interface.

21. OVOS (Open Voice OS) — 268 Stars

GitHub: OpenVoiceOS/ovos-core · Apache 2.0

Community continuation of Mycroft after MycroftAI's closure. Voice-to-intent pipeline with extensible skill framework. LLM integration via Persona plugins (Ollama, OpenAI-compatible). Runs on Raspberry Pi and embedded Linux. 6,029 commits, 256 releases. Tiny community but dedicated. Best for: privacy-focused voice assistant enthusiasts willing to invest setup time.

20. Home Assistant Voice — 85.4K 星

GitHub: home-assistant/core · Apache 2.0

主導的開源智慧家庭平台，透過 Assist Pipeline 支援語音助手功能。支援 LLM 整合、本地 STT/TTS、喚醒詞偵測、多語言助手。4,662+ 貢獻者。非個人 AI 助手——專注於智慧家庭自動化。

21. OVOS（Open Voice OS）— 268 星

GitHub: OpenVoiceOS/ovos-core · Apache 2.0

Mycroft 關閉後的社群延續。語音到意圖管道，可擴展技能框架。適合：願意投入設置時間的隱私導向語音助手愛好者。

Tier 7: Emerging Alternatives (New Entrants)

These tools appeared in comparison articles and are worth watching, though they have less established track records.

第七層：新興替代方案（新進入者）

這些工具出現在比較文章中，值得關注，但記錄較不成熟。

22. memU — Persistent Memory Assistant

GitHub: NevaMind-AI/memU · ~6.9K stars · Open Source

Personal AI assistant emphasizing long-term memory via hierarchical knowledge graphs. Proactive action suggestions, context compression, cost-optimized API usage. Local-first design. Best for: users wanting an assistant that genuinely learns and improves over time. Weakness: less powerful for complex real-time task execution.

23. Moltis — Production Observability Agent

Open-source Rust-based self-hosted AI assistant prioritizing observability: Prometheus metrics, OpenTelemetry tracing, structured logging. 27 workspace crates, 53 feature flags, Docker/Podman/Apple Container sandboxing, Tailscale integration. Best for: SRE-minded users who want production-grade monitoring of their AI agent.

24. NullClaw — The Minimalist Binary

GitHub: nullclaw/nullclaw · ~2.6K stars

678KB Zig-based single binary with 22+ LLM provider support. Edge computing ready, fast startup. Best for: extreme minimalists and edge deployment scenarios.

25. Leon AI — The Offline-First Pioneer (Stalled)

GitHub: leon-ai/leon · 17.1K stars · MIT

Open-source personal assistant designed for complete offline operation. Currently in a massive architectural rewrite (highly experimental develop branch). Single maintainer working in spare time. Best for: nobody right now — wait for the rewrite to stabilize. The vision (offline-first, skills-based, voice + text) is excellent but execution is stalled.

22. memU — 持久記憶助手

GitHub: NevaMind-AI/memU · ~6.9K 星 · 開源

透過階層式知識圖譜強調長期記憶的個人 AI 助手。適合：想要助手隨時間學習和改進的用戶。

23. Moltis — 生產可觀測性代理

開源 Rust 自託管 AI 助手，優先考慮可觀測性：Prometheus 指標、OpenTelemetry 追蹤。適合：重視生產級監控的 SRE 用戶。

24. NullClaw — 極簡二進制

GitHub: nullclaw/nullclaw · ~2.6K 星

678KB Zig 單一二進制，支援 22+ LLM 供應商。適合：極端簡約主義者和邊緣部署場景。

25. Leon AI — 離線優先先驅（停滯中）

GitHub: leon-ai/leon · 17.1K 星 · MIT

設計為完全離線運作的開源個人助手。目前正在進行大規模架構重寫。單一維護者利用業餘時間工作。願景優秀但執行停滯。

Decision Framework: Which Tool Should You Choose?

If you want the most capable personal AI assistant:

→ OpenClaw. Nothing else matches its 23+ messaging platform integrations, native mobile apps, browser control, voice wake, and ecosystem momentum. Accept the security trade-offs or harden your setup.

If security is your top priority:

→ NanoClaw. Container isolation means the agent literally cannot access your host filesystem. Trade-off: Claude-only and fewer platforms.

If you want maximum performance on minimal hardware:

→ ZeroClaw (5MB RAM, 10ms startup) or PicoClaw ($10 RISC-V board). ZeroClaw for production use; PicoClaw for IoT/edge experimentation.

If you want a lightweight, hackable alternative:

→ Nanobot. 4,000 lines of Python, 11+ messaging platforms, 12+ LLM providers. Read the entire codebase in an afternoon. Growing fast.

If Chinese messaging platforms matter:

→ CoPaw (DingTalk, Feishu, QQ first-class) or Nanobot (also strong Chinese platform support).

If you want a knowledge/document assistant:

→ Khoj (AI second brain with document search, WhatsApp support) or AnythingLLM (best RAG, widest LLM support).

If you want multi-agent parallel workflows:

→ Eigent. Unique multi-agent architecture with specialized agents working in parallel. Top GAIA benchmark scores.

If you want a goal-oriented system that learns you:

→ PAI. The TELOS system (mission, goals, beliefs) and three-tier memory are unmatched for personalization.

If you just want a good chat UI:

→ LobeChat (prettiest, 10K+ plugins) or LibreChat (ChatGPT clone with enterprise auth).

If you want to run LLMs locally with zero cloud:

→ Jan. Best desktop experience for local model inference. Simple, clean, works.

If you want smart home voice control:

→ Home Assistant Voice. 85K stars, 4,600+ contributors, LLM integration, local TTS/STT. Not a personal assistant but unmatched for home automation.

決策框架：你應該選擇哪個工具？

如果你想要最強大的個人 AI 助手：

→ OpenClaw。 沒有其他工具能匹配它的 23+ 通訊平台整合、原生行動應用、瀏覽器控制、語音喚醒和生態系統動能。接受安全權衡或加固你的設置。

如果安全是最高優先級：

→ NanoClaw。 容器隔離意味著代理無法存取主機檔案系統。權衡：僅支援 Claude，平台較少。

如果你想在最小硬體上獲得最大效能：

→ ZeroClaw（5MB RAM、10ms 啟動）或 PicoClaw（$10 RISC-V 開發板）。

如果你想要輕量級、可修改的替代方案：

→ Nanobot。 4,000 行 Python，11+ 通訊平台，12+ LLM 供應商。一個下午讀完整個程式碼庫。

如果中國通訊平台很重要：

→ CoPaw（DingTalk、飛書、QQ 一級支援）或 Nanobot（也有強大的中國平台支援）。

如果你想要知識/文件助手：

→ Khoj（AI 第二大腦，支援文件搜尋和 WhatsApp）或 AnythingLLM（最佳 RAG，最廣 LLM 支援）。

如果你想要多代理平行工作流：

→ Eigent。 獨特的多代理架構，專業代理平行工作。GAIA 基準測試頂級分數。

如果你想要了解你的目標導向系統：

→ PAI。 TELOS 系統（使命、目標、信念）和三層記憶在個人化方面無與倫比。

如果你只想要好的聊天 UI：

→ LobeChat（最美觀，10K+ 外掛）或 LibreChat（帶企業認證的 ChatGPT 克隆）。

如果你想完全本地運行 LLM：

→ Jan。 本地模型推理的最佳桌面體驗。簡單、乾淨、好用。

如果你想要智慧家庭語音控制：

→ Home Assistant Voice。 85K 星、4,600+ 貢獻者、LLM 整合、本地 TTS/STT。

The Security Spectrum

This is the most important differentiator in this space. Ranked from most to least secure:

Rank	Tool	Security Approach
1	NanoClaw	Container isolation (Docker/Apple Container), per-group filesystem isolation
2	ZeroClaw	Workspace-scoped filesystem, pre-blocked sensitive paths, strict sandboxing by default
3	PAI	User/system separation, granular permission policies, dangerous operation blocking
4	OpenClaw	Pairing DM policy, configurable sandbox, `doctor` tool — but still powerful attack surface
5	Nanobot	Allow-lists, per-group policies, optional E2EE — but no sandboxing by default
6	CoPaw	Tool Guard layer, @mention filtering — but shell sandboxing still on roadmap
7	Open Interpreter	Manual approval only — `auto_run=True` removes all protection
8	PicoClaw	Minimal security model (pre-v1.0)

The fundamental tension: more capability = more attack surface. OpenClaw gives agents access to shell, files, browser, camera, SMS, and contacts — each is a potential vector for prompt injection. NanoClaw's container approach is the most architecturally sound solution, but it limits what the agent can do.

安全性光譜

這是該領域最重要的差異化因素。從最安全到最不安全排列：

排名	工具	安全方法
1	NanoClaw	容器隔離（Docker/Apple Container）、按群組檔案系統隔離
2	ZeroClaw	工作區範圍檔案系統、預封鎖敏感路徑、預設嚴格沙箱
3	PAI	使用者/系統分離、細粒度權限策略、危險操作阻擋
4	OpenClaw	配對 DM 策略、可配置沙箱、`doctor` 工具——但攻擊面仍大
5	Nanobot	允許列表、按群組策略、可選 E2EE——但預設無沙箱
6	CoPaw	Tool Guard 層、@mention 過濾——但 shell 沙箱仍在路線圖中
7	Open Interpreter	僅手動批准——`auto_run=True` 移除所有保護
8	PicoClaw	最小安全模型（pre-v1.0）

根本張力：能力越多 = 攻擊面越大。 OpenClaw 給予代理 shell、檔案、瀏覽器、相機、簡訊和聯絡人存取——每一個都是潛在的提示注入向量。NanoClaw 的容器方法是架構上最合理的解決方案，但它限制了代理能做的事。

Final Verdict

The personal AI assistant space in March 2026 is defined by one question: how much autonomy are you willing to grant an AI agent on your own machine?

OpenClaw answered "maximum" and became the most popular open-source project on GitHub. But its "god-mode" approach created the security backlash that spawned NanoClaw, ZeroClaw, and a dozen other alternatives with varying security/capability trade-offs.

The ecosystem is consolidating around five archetypes:

Full-power agents (OpenClaw, Nanobot, CoPaw) — maximum capability, maximum risk
Security-hardened agents (NanoClaw, ZeroClaw) — container/sandbox isolation
Goal-oriented systems (PAI, memU) — persistent memory and personalization
Knowledge assistants (Khoj, AnythingLLM) — search and RAG over your documents
Multi-agent platforms (Eigent, SuperAGI) — coordinated agent teams

No single tool does everything well. The winners will be those that find the right balance between capability, security, and ease of use. Right now, OpenClaw has the momentum — but NanoClaw's container-first approach and Nanobot's research-friendly simplicity may prove more sustainable as the security concerns inevitably escalate.

最終結論

2026 年 3 月的個人 AI 助手領域由一個問題定義：你願意在自己的機器上賦予 AI 代理多少自主權？

OpenClaw 回答了「最大化」，成為 GitHub 上最受歡迎的開源專案。但其「上帝模式」方法引發了安全反彈，催生了 NanoClaw、ZeroClaw 和十幾個其他具有不同安全/能力權衡的替代方案。

生態系統正在圍繞五個原型整合：

全功能代理（OpenClaw、Nanobot、CoPaw）——最大能力、最大風險
安全強化代理（NanoClaw、ZeroClaw）——容器/沙箱隔離
目標導向系統（PAI、memU）——持久記憶和個人化
知識助手（Khoj、AnythingLLM）——文件搜尋和 RAG
多代理平台（Eigent、SuperAGI）——協調的代理團隊

沒有單一工具能把所有事情都做好。贏家將是那些在能力、安全性和易用性之間找到正確平衡的工具。目前 OpenClaw 擁有動能——但 NanoClaw 的容器優先方法和 Nanobot 的研究友好簡潔性，隨著安全擔憂不可避免地升級，可能被證明更具可持續性。

OpenClaw vs Every Alternative ​

The Definitive Personal AI Assistant Comparison ​

OpenClaw vs 所有替代方案 ​

個人 AI 助手完全比較指南 ​

The Landscape in March 2026 ​

2026 年 3 月的格局 ​

Quick Reference: The Complete Comparison Matrix ​

快速參考：完整比較矩陣 ​

Tier 1: Direct Competitors (Personal AI Assistants with Messaging Integration) ​

第一層：直接競爭者（具備訊息整合的個人 AI 助手） ​

1. OpenClaw — The Category King ​

1. OpenClaw — 品類之王 ​

2. Nanobot — The Lightweight Research-Ready Alternative ​

2. Nanobot — 輕量級研究友好替代方案 ​

3. NanoClaw — The Security-First Containerized Agent ​

3. NanoClaw — 安全優先的容器化代理 ​

4. ZeroClaw — The Performance-Obsessed Rust Runtime ​

4. ZeroClaw — 效能極致的 Rust 運行時 ​

5. PicoClaw — The $10 Hardware Agent ​

5. PicoClaw — $10 硬體代理 ​

6. CoPaw — The China-Focused Multi-Channel Assistant ​

6. CoPaw — 以中國市場為重心的多頻道助手 ​

7. Khoj — The AI Second Brain ​

7. Khoj — AI 第二大腦 ​

Tier 2: Action-Capable Assistants (Without Messaging Integration) ​

第二層：具行動能力的助手（無通訊整合） ​

8. Open Interpreter — The OG Local Code Executor ​

8. Open Interpreter — 本地程式碼執行的先驅 ​

9. Eigent — The Multi-Agent Desktop Workforce ​

9. Eigent — 多代理桌面工作團隊 ​

10. PAI (Personal AI Infrastructure) — The Goal-Oriented System ​

10. PAI（Personal AI Infrastructure）— 目標導向系統 ​

11. AnythingLLM — The Universal LLM Hub ​

11. AnythingLLM — 通用 LLM 中心 ​

12. AutoGPT — The Pioneer (Now a Platform) ​

12. AutoGPT — 先驅者（現為平台） ​

Tier 3: Chat UIs and LLM Frontends ​

第三層：聊天 UI 和 LLM 前端 ​

13. LobeChat — 73.8K Stars ​

14. LibreChat — 34.7K Stars ​

15. Jan — 41.1K Stars ​

13. LobeChat — 73.8K 星 ​

14. LibreChat — 34.7K 星 ​

15. Jan — 41.1K 星 ​

Tier 4: Autonomous Agent Frameworks ​

第四層：自主代理框架 ​

16. SuperAGI — 17.3K Stars ​

17. BabyAGI — 22.2K Stars ​

16. SuperAGI — 17.3K 星 ​

17. BabyAGI — 22.2K 星 ​

Tier 5: Conversational AI Platforms (Chatbot Builders) ​

第五層：對話式 AI 平台（聊天機器人建構器） ​

18. Botpress — 14.6K Stars ​

19. Rasa — 21.1K Stars ​

18. Botpress — 14.6K 星 ​

19. Rasa — 21.1K 星 ​

Tier 6: Voice-First Assistants ​

第六層：語音優先助手 ​

20. Home Assistant Voice — 85.4K Stars ​

21. OVOS (Open Voice OS) — 268 Stars ​

20. Home Assistant Voice — 85.4K 星 ​

21. OVOS（Open Voice OS）— 268 星 ​

Tier 7: Emerging Alternatives (New Entrants) ​

第七層：新興替代方案（新進入者） ​

22. memU — Persistent Memory Assistant ​

23. Moltis — Production Observability Agent ​

24. NullClaw — The Minimalist Binary ​

25. Leon AI — The Offline-First Pioneer (Stalled) ​

22. memU — 持久記憶助手 ​

23. Moltis — 生產可觀測性代理 ​

24. NullClaw — 極簡二進制 ​

25. Leon AI — 離線優先先驅（停滯中） ​

Decision Framework: Which Tool Should You Choose? ​

If you want the most capable personal AI assistant: ​

If security is your top priority: ​

If you want maximum performance on minimal hardware: ​

If you want a lightweight, hackable alternative: ​

If Chinese messaging platforms matter: ​

If you want a knowledge/document assistant: ​

If you want multi-agent parallel workflows: ​

OpenClaw vs Every Alternative

The Definitive Personal AI Assistant Comparison

OpenClaw vs 所有替代方案

個人 AI 助手完全比較指南

The Landscape in March 2026

2026 年 3 月的格局

Quick Reference: The Complete Comparison Matrix

快速參考：完整比較矩陣

Tier 1: Direct Competitors (Personal AI Assistants with Messaging Integration)

第一層：直接競爭者（具備訊息整合的個人 AI 助手）

1. OpenClaw — The Category King

1. OpenClaw — 品類之王

2. Nanobot — The Lightweight Research-Ready Alternative

2. Nanobot — 輕量級研究友好替代方案

3. NanoClaw — The Security-First Containerized Agent

3. NanoClaw — 安全優先的容器化代理

4. ZeroClaw — The Performance-Obsessed Rust Runtime

4. ZeroClaw — 效能極致的 Rust 運行時

5. PicoClaw — The $10 Hardware Agent

5. PicoClaw — $10 硬體代理

6. CoPaw — The China-Focused Multi-Channel Assistant

6. CoPaw — 以中國市場為重心的多頻道助手

7. Khoj — The AI Second Brain

7. Khoj — AI 第二大腦

Tier 2: Action-Capable Assistants (Without Messaging Integration)

第二層：具行動能力的助手（無通訊整合）

8. Open Interpreter — The OG Local Code Executor

8. Open Interpreter — 本地程式碼執行的先驅

9. Eigent — The Multi-Agent Desktop Workforce

9. Eigent — 多代理桌面工作團隊

10. PAI (Personal AI Infrastructure) — The Goal-Oriented System

10. PAI（Personal AI Infrastructure）— 目標導向系統

11. AnythingLLM — The Universal LLM Hub

11. AnythingLLM — 通用 LLM 中心

12. AutoGPT — The Pioneer (Now a Platform)

12. AutoGPT — 先驅者（現為平台）

Tier 3: Chat UIs and LLM Frontends

第三層：聊天 UI 和 LLM 前端

13. LobeChat — 73.8K Stars

14. LibreChat — 34.7K Stars

15. Jan — 41.1K Stars

13. LobeChat — 73.8K 星

14. LibreChat — 34.7K 星

15. Jan — 41.1K 星

Tier 4: Autonomous Agent Frameworks

第四層：自主代理框架

16. SuperAGI — 17.3K Stars

17. BabyAGI — 22.2K Stars

16. SuperAGI — 17.3K 星

17. BabyAGI — 22.2K 星

Tier 5: Conversational AI Platforms (Chatbot Builders)

第五層：對話式 AI 平台（聊天機器人建構器）

18. Botpress — 14.6K Stars

19. Rasa — 21.1K Stars

18. Botpress — 14.6K 星

19. Rasa — 21.1K 星

Tier 6: Voice-First Assistants

第六層：語音優先助手

20. Home Assistant Voice — 85.4K Stars

21. OVOS (Open Voice OS) — 268 Stars

20. Home Assistant Voice — 85.4K 星

21. OVOS（Open Voice OS）— 268 星

Tier 7: Emerging Alternatives (New Entrants)

第七層：新興替代方案（新進入者）

22. memU — Persistent Memory Assistant

23. Moltis — Production Observability Agent

24. NullClaw — The Minimalist Binary

25. Leon AI — The Offline-First Pioneer (Stalled)

22. memU — 持久記憶助手

23. Moltis — 生產可觀測性代理

24. NullClaw — 極簡二進制

25. Leon AI — 離線優先先驅（停滯中）

Decision Framework: Which Tool Should You Choose?

If you want the most capable personal AI assistant:

If security is your top priority:

If you want maximum performance on minimal hardware:

If you want a lightweight, hackable alternative:

If Chinese messaging platforms matter:

If you want a knowledge/document assistant:

If you want multi-agent parallel workflows: