QwenPaw is a personal AI assistant harness under the Apache-2.0 licence, aimed at people who want an LLM agent running on their own machine or in their own cloud account and reachable from the chat apps they already use.

What it is

QwenPaw is an open-source personal AI assistant from the agentscope-ai organisation, filed in this registry under AI & Machine Learning / AI Interaction & Interfaces. It ships as the qwenpaw package on PyPI and requires Python 3.11 up to (but not including) 3.14, while this registry entry lists TypeScript as the project language; the README that backs the entry is Python-centric. The project carries 35,051 stars, 3,115 forks and 990 open issues, with the last push recorded on 17 September 2026. It sits inside the AgentScope ecosystem and uses the Model Context Protocol (mcp) as one of its extension surfaces, with tags spanning agent, agent-harness, loop-engineering and personal-ai-assistant.

The concrete problem it solves is the glue work between a local model runtime, a durable memory layer and a separate bot integration for every messaging channel. Rather than assembling an Ollama or LM Studio server, a hand-rolled conversation store and one bridge per chat app, a user runs one harness that already provides the agent loop, the memory, the extension mechanism and the channel connections. It replaces that self-built stack, not a hosted chat product: the README positions the built-in QwenPaw Local runtime as needing no API key and no cloud dependency.

Key capabilities

• Three-layer memory: live working context, full verbatim history, and a self-evolving personal knowledge base powered by ReMe; conversations and resources continuously become readable, editable, searchable and linked Markdown memory.
• Built-in QwenPaw Local runtime serving QwenPaw-Flash models at 2B, 4B and 9B sizes trained for agent tasks, with no API key required.
• Alternative model backends through Ollama, LM Studio and 14 or more cloud providers.
• Extension through Skills and Plugins, described in the tagline as easily extensible capabilities.
• Model Context Protocol support for tool use, reflected in the mcp and llm-tools topics.
• Agent-harness and loop-engineering design, reflected in the agent-harness, harness-engineering and loop-engineering topics.
• Multi-channel reach across chat apps, so one assistant instance answers in the places a user already talks.

Who uses it and how

• Individuals running a private assistant on a laptop with the QwenPaw Local runtime, avoiding API keys and cloud spend.
• Users deploying to a cloud host for an always-on assistant, following the documentation at qwenpaw.agentscope.io.
• Developers who need agent behaviour wired into existing products and extend it with Skills, Plugins and MCP tools.
• Teams already inside the AgentScope ecosystem, where the AgentScope Platform is listed as online 24/7 and free.
• Non-English-speaking users, since the README ships in English, Chinese, Japanese and Russian, with community channels on Discord, X and DingTalk.

Getting started

Install the qwenpaw package from PyPI on Python 3.11 through 3.13, then follow the documentation site at qwenpaw.agentscope.io. For local inference, the bundled QwenPaw Local runtime means the assistant can start without configuring any external provider.

How it compares

The facts name no paid products this project replaces. Among the tools it does name, Ollama and LM Studio are model runtimes rather than complete assistants, and QwenPaw consumes them as backends alongside its own QwenPaw Local runtime. ReMe supplies the memory layer instead of competing with the assistant, and AgentScope provides the upstream ecosystem and platform.

When to use it — and when not to

A self-hoster must run a Python 3.11 to 3.13 environment and accept the operational surface the harness brings, including the local model runtime when no external provider is configured. Anyone wanting a fully managed, zero-operations assistant with a stable vendor contract should not pick this. The registry metadata is also inconsistent, listing TypeScript while the distribution and README are Python-based, and 990 open issues indicate active but unsettled development; the README excerpt stops mid-sentence on the cloud-provider list, so deployment topology must be confirmed from the documentation site rather than the repository front page.