Personal AI assistants keep asking you to live inside their app, on their servers, with their memory of your life. QwenPaw is the opposite pitch from the AgentScope team: install it on a machine you control, wire your own models, and talk to the same agent from a browser console, a terminal, DingTalk, Discord, Telegram, or a desktop app.
It is Apache 2.0, Python 3.11+, and the GitHub star count is silly high for something you still have to operate yourself. That popularity matters because plugins, skills, and channel guides show up fast when a project is this visible.
What you are actually running
QwenPaw 2.x is built as a small agent OS, not a single chat window. You get:
- A web Console at
http://127.0.0.1:8088afterqwenpaw app, plus a full-screen TUI if you prefer the terminal - Channels so one agent answers on Discord, Lark, WeChat, QQ, iMessage, and the rest without retraining anything per app
- Skills and plugins (scheduling, browser use, document conversion, news, MCP tools) with a marketplace mindset
- ReMe memory: conversations and files get folded into a personal knowledge base you can search and edit, not a black-box chat log
- Multi-agent setups with sub-agents and Agent Communication Protocol (ACP) when one brain is not enough
- Optional QwenPaw-Flash local model sizes (2B through 27B) if you want agent tasks without shipping prompts to a cloud API
The security story is louder than most hobby projects: kernel-style sandboxing per OS, a YAML tool guard, file guard, skill scanner, and access policies that can require human approval before scary commands run. That is the right instinct for software that can touch your filesystem and shell.
How it differs from a lighter agent like nanobot
We covered nanobot recently. nanobot is a slim Python gateway: bring a model, add channels, go. QwenPaw is the furnished apartment. More UI surfaces, built-in local model paths, richer memory (ReMe), and more opinion about governance. Choose nanobot when you want to understand every line; choose QwenPaw when you want AgentScope’s full stack and you are willing to feed it RAM, disk, and update attention.
Why self-host
Your prompts and files stay on your metal. Even when you call a cloud LLM, the workspace, skills, scheduled jobs, and ReMe store live under directories you own (~/.qwenpaw by default), not on a vendor’s multi-tenant cluster.
Channels are identity. Hooking Discord or corporate Lark to a SaaS bot means their terms, their retention, their breach surface. Running the bridge on your VPS keeps the integration under your access rules.
Local models are optional but real. QwenPaw-Flash plus Ollama or LM Studio is a credible path for drafts, cron digests, or internal tools where outbound API calls are a non-starter.
Canadian hosting angle. If you are explaining data residency to a client or your own privacy officer, “the agent runs in Beauharnois or Toronto, keys in our vault, logs on our disk” beats “we use a US assistant product and hope the DPA holds.”
Getting it running
Fastest try on a laptop: the one-liner install script from qwenpaw.agentscope.io, or pip install qwenpaw, then qwenpaw init --defaults and qwenpaw app. Cloud LLM users must set an API key in Console → Settings → Models before chat works.
On a server, Docker is the path you will want for restarts and backups:
docker pull agentscope/qwenpaw:latest
docker run -p 127.0.0.1:8088:8088 \
-v qwenpaw-data:/app/working \
-v qwenpaw-secrets:/app/working.secret \
-v qwenpaw-backups:/app/working.backups \
agentscope/qwenpaw:latestPut nginx or Caddy in front for HTTPS if you expose the Console beyond localhost. Ollama on the same host needs --add-host=host.docker.internal:host-gateway (or host networking on Linux) and a base URL like http://host.docker.internal:11434 in model settings, not localhost inside the container.
Resource planning: the agent process is not the whole story. A 2B local model might idle on a small VPS; a 27B quant or GPU-backed Ollama wants a bigger box with fast disk for model weights. Schedule cron and channel bots on a machine that stays up, not your laptop.
Who should bother, and who should wait
Good fit: developers who already live in AgentScope or Qwen tooling, teams that want one agent on multiple chat apps, and anyone who needs sandboxed tool use with audit-friendly guardrails.
Maybe wait: if you only need a shared ChatGPT-style UI over Ollama, Open WebUI is less moving parts. QwenPaw’s desktop app is still beta; the roadmap is wide and the project moves fast, so pin versions in production and read release notes before you docker pull latest on a Friday.
Bottom line
QwenPaw is a serious self-hosted personal assistant: multi-channel, extensible, memory-aware, and designed for people who expect agents to run commands, not just chat. It asks for more server than a static blog, but you get to own the whole loop.
Want help sizing a Canadian VPS for QwenPaw plus a local model? Talk with Swift Host. We will split “agent container” and “GPU or big-RAM inference” instead of pretending one tiny plan covers both.
