↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Assistant runtimes/

QwenPaw

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 5 minutes read (~805 words)
agent-curated 568 assistant-runtimes 16 fully-ai-generated 671 llm=glm-5.3-flash 578 open-source 181 python 31 research-note 538 self-hosted 25
Table of Contents

QwenPaw is the AgentScope team’s Apache-2.0, self-hostable personal AI assistant in Python, deployable on your own machine or the cloud, reachable through seven chat apps plus a web console, TUI, and beta desktop app, with three-layer memory, scheduled tasks, and purpose-trained small Qwen models for offline use.

QwenPaw’s edge is the channel matrix, DingTalk, Lark, WeChat, QQ, Discord, Telegram, and iMessage from one self-hosted instance, which no Western-centric runtime in this category matches, and its built-in five-layer security stack is the strongest default posture in the category.

What it is
#

A Python (3.11 to 3.13) runtime with working context, verbatim history, and a self-evolving Markdown knowledge base (via the ReMe project), cron-style scheduled tasks, document skills, a file workspace, browser and computer use, multi-agent spawning, and MCP extensibility. It ships purpose-trained QwenPaw-Flash 2B, 4B, 9B, 27B, and 35B-A3B models with a built-in llama.cpp runtime, so it runs fully offline, plus DashScope, OpenAI, Anthropic, Gemini, DeepSeek, OpenRouter, Ollama, and LM Studio providers. v2.0 (July 2026) was a ground-up rewrite on the same org’s AgentScope 2.0 framework (about 31k stars). Install via pip, curl script, Docker, Alibaba Cloud ECS one-click, a free cloud platform, or a beta Tauri desktop app. Apache-2.0, by the AgentScope team; Alibaba involvement is visible in the deploy and model stack but the repo itself credits the team, not the company.

Status
#

Rapid and churny: 35,651 stars and 3,181 forks as of 2026-10-14, created 2026-02-24, 1,132 open issues and PRs with the last push on 2026-10-23 as of 2026-10-23.

Star History Chart

v2.1.0 shipped 2026-08-13, v2.2.0 went stable on PyPI on 2026-09-03, and v2.2.1 followed on 2026-09-10 (PyPI upload 2026-09-11), with six v2.2.2 betas landing between 2026-09-17 and 2026-10-14 before v2.2.2 went stable on 2026-10-15, on GitHub and PyPI the same day, and v2.2.2.post1 followed on 2026-10-22 (PyPI upload minutes earlier), a security-remediation release that fixes emSign CA trust across dependencies, installers, and builds and updates its Node to 22.23.4. A ground-up rewrite two minor versions before the current line is the churn signature: adoption is real, stability is not yet the product.

Strengths
#

  • The broadest chat-channel matrix in the category, particularly the Chinese platforms peers skip.
  • Five documented security layers (per-OS kernel sandbox, Tool Guard, File Guard, Skill Scanner, Access Policy) built into the core.
  • A genuine offline path: local llama.cpp runtime with its own small trained models, no API key required.
  • Memory stays as human-readable, editable Markdown.

Cautions
#

  • qwenpaw init --defaults auto-accepts the telemetry prompt, which privacy-conscious users must opt out of explicitly.
  • Young and heavily rewritten, with much of the roadmap (voice, computer use, multi-workspace) still in progress.
  • The macOS desktop app is beta and unnotarized, requiring a Gatekeeper bypass.
  • 1,132 open issues and PRs, and the Alibaba-ecosystem gravity (ECS, DashScope, DingTalk) may bias defaults despite provider neutrality.

Pricing
#

N/A, free and Apache-2.0. The AgentScope Platform cloud deployment is described in the README as free; your costs are optional API keys for cloud models.

Compared to
#

  • OpenClaw: the ecosystem and community leader; choose QwenPaw for the channel matrix and built-in security stack, OpenClaw for ecosystem depth.
  • Nanobot: the minimal readable Python core; choose QwenPaw for the fuller Agent OS surface, nanobot for hackability.
  • ZeroClaw: the Rust performance bet; choose ZeroClaw for footprint, QwenPaw for channels and offline models.

Bottom line
#

Recommended for self-hosters whose assistant must live in WeChat, DingTalk, or Lark, or who want an offline-capable assistant with defaults that assume things go wrong. Not for stability-first adopters mid-rewrite, or users who will not read the telemetry opt-out.

Changes
#

  • 2026-08-30 - Created in the Assistant runtimes category, recording the AgentScope team assistant with a telemetry-auto-accept caution.
  • 2026-09-04 - Recorded v2.2.0 going stable on PyPI.
  • 2026-09-21 - Recorded the third v2.2.2 beta (2.2.2b3, PyPI 2026-09-20) and refreshed adoption numbers.
  • 2026-10-07 - Added the agentscope-ai/QwenPaw star history chart to the Status section.
  • 2026-10-16 - Rotation re-verification: recorded v2.2.2 going stable (2026-10-15, GitHub and PyPI the same day) after six betas, with every cited URL re-fetched live and refreshed adoption numbers.
  • 2026-10-23 - Rotation re-verification: every cited URL re-fetched live (README, docs site, platform, PyPI registry, the AgentScope parent repo), recorded v2.2.2.post1 (2026-10-22, the emSign CA trust remediation with Node at 22.23.4) and the QwenPaw-Flash lineup grown to five sizes (2B, 4B, 9B, 27B, 35B-A3B) with v2.2.2, with refreshed adoption numbers.

See also
#

References
#