qwenpaw/website/public/docs/comparison.en.md

QwenPaw vs OpenClaw: Feature Comparison

Category	Dimension	OpenClaw	QwenPaw
Tech Stack	Primary Language	TypeScript / Node.js	Python
	Agent Framework	• Pi agent runtime	• AgentScope • AgentScope-Runtime
	Memory System	• Workspace file memory • Session model: group isolation, context compaction (/compact), and session pruning	• Long-term workspace memory powered by ReMe • Layered context: key information and recent turns in memory; history, rolling summaries, and tool outputs persisted • Dynamic compaction before inference: prioritize recent high-signal content; compress older content into structured summaries with indexed recall to originals when needed • Time-tiered compression of tool outputs to save tokens • Hybrid retrieval: vector search + full-text search (e.g. BM25) • Structured summaries and long-term memory files for user preferences and task experience • Per-role memory isolation in multi-agent setups to reduce cross-task interference
User Experience	Installation	• Global install of openclaw via npm / pnpm / bun • openclaw onboard wizard; optional --install-daemon for the Gateway daemon	• .zip / .exe installers • One-line script installation • pip install qwenpaw • Docker installation • One-click cloud deployment
	Supported Platforms	macOS / Linux / Windows	macOS / Linux / Windows (PowerShell/CMD)
	Local Model Support	• Configure Ollama / llama.cpp and other endpoints via config • Models and failover	• Install-time --extras for the underlying inference runner; supports LM Studio, Ollama, llama.cpp • Built-in llama.cpp local model provider and global LLM rate limiting (QPM sliding window) • Optional QwenPaw-Flash series tuned for QwenPaw (Trinity-RFT post-training and OpenJudge evaluation alignment; emphasizes docs, scheduling, memory updates, retrieval, and other high-frequency tasks) • 2B / 4B / 9B and full / Q8 / Q4 variants; hardware-aware recommendations; download, enable, and switch in the Console
	Skills Support	• Local Skills • Bundled / managed / workspace Skills with install gating • Install from ClawHub	• Local Skills • Direct import from multiple public Skills Hubs (skills.sh, clawhub.ai, skillsmp.com, lobehub.com, GitHub, modelscope.cn/skills, etc.) • Two-layer skill pool architecture
	Channel Integrations	WhatsApp, Telegram, Slack, Discord, Google Chat, Signal, BlueBubbles/iMessage, IRC, Teams, Matrix, Feishu, LINE, Mattermost, Nextcloud Talk, Nostr, Synology Chat, Tlon, Twitch, Zalo, WeChat, WebChat, etc.; extensible	DingTalk, Feishu, WeChat, WeCom, QQ, Xiaoyi, Discord, Telegram, iMessage, Mattermost, Matrix, Twilio, MQTT; extensible
Community Ecosystem	Open-source License	MIT	Apache 2.0
Features	Memory System	• Workspace file memory • Session model: group isolation, context compaction (/compact), and session pruning	• Powered by ReMe • Dynamic compaction before inference: prioritize recent high-signal content; compress older content into structured summaries with indexed recall to originals when needed • Time-tiered compression of tool results • Structured summaries combined with long-term memory files • Hybrid retrieval: vector search + full-text search • Per-role memory isolation in multi-agent setups • Multimodal memory fusion; experience distillation & Skill extraction; context-aware proactive delivery (planned)
	Multi-agent	• Route channels / accounts / peers to isolated agents (workspace + per-agent sessions) • sessions_* tools for cross-session coordination	• AgentScope-based multi-workspace isolation and collaboration • Several agents in parallel in one instance; separate config, ReMe memory, skills, and chat history per agent • Concurrent load with locking; per-workspace hot reload and atomic cutover when a new instance is ready • CLI --background and /stop; enable/disable agents in Console and API • Collaborator agents use fresh sessions by default to avoid polluting the main agent context • Async collaboration and multi-agent collaboration Skills for complex tasks; cross-turn state externalized to the filesystem first to limit context growth
	Reliability & operations	• openclaw doctor diagnostics and migrations • Retry policy, model failover, logging	• Daemon Agent for long-horizon tasks and health monitoring • Memory-related and Daemon-related magic commands
	Security	• Default DM pairing and allowlist across channels • Optional Docker sandbox • Security documentation • ClawHub marketplace VirusTotal scanning	• Tool guard • Skill scanning • File guard • Tool sandbox (Seatbelt / Bubblewrap / Landlock / AppContainer)
	Cloud & remote access	• Tailscale Serve/Funnel, SSH tunnels, and remote Gateway control • Docker / Nix deployment	• Extend cloud compute, storage, and services via AgentScope Runtime • Docker deployment
	Large-Small Model Collaboration	• Multi-model configuration and failover • Docs recommend latest-generation strong models to reduce prompt-injection risk	• Optional QwenPaw-Flash series tuned for QwenPaw (Trinity-RFT post-training and OpenJudge evaluation alignment; emphasizes docs, scheduling, memory updates, retrieval, and other high-frequency tasks) • Lightweight local models for privacy-sensitive data; long-context planning and reasoning to cloud LLMs (planned)
	Multimodal Interaction	• Voice Wake / Talk Mode • Media pipeline • Live Canvas (A2UI) • macOS / iOS / Android companion apps	• Multimodal preview in Console chat • Voice and video interaction
	Skills & ecosystem	• ClawHub and built-in Skills continue to expand	• Continuously enrich the AgentScope Skills repository and improve discovery and use of high-quality Skills