QwenPaw

You will like this if you want an assistant that answers in DingTalk, Lark, Discord or iMessage without a monthly API bill. The daily trait is the local runtime: QwenPaw-Flash models at 2B, 4B and 9B are trained for agent work, ship in Q4 and Q8 quantisations, and download from a button in the web UI, so the thing works with no key at all.

Where it hurts is that small models are small, and the good answers still come from a cloud key. Pick it for a home or lab assistant on Chinese chat apps. Pick NanoClaw if you want Claude in a container.

The risk is velocity. Version 2.0.0 was a rewrite on a new base in July 2026, and three more releases followed inside eight weeks. The roadmap lists fifteen items as in progress, among them batch preview and approval, multi-location file changes, persistent terminals and running-task steering, which are the parts that keep a coding assistant from corrupting a file halfway through. The desktop app is labelled beta and is not notarised.

The consequence: use the console and the channels, not the coding mode, until the roadmap shrinks. What it does right: File Guard blocks the agent from the SSH directory and its own secrets folder by default.

The context design is the interesting part. 1. Scroll Context: every turn is persisted, and turns evicted from the window are indexed for on-demand recall rather than compressed into a summary. 2. Memory is layered: live working context, verbatim history, and a self-evolving knowledge base kept as linked Markdown by the ReMe component. 3. Loop Engineering supplies agent-loop templates, a Coding Mode and a Mission Mode, with approval gates.

Verification is by gate rather than by test. The purpose-trained models are described as trained for agent tasks, with no published evaluation. The observation: recall on demand is only as good as the retriever, and the retriever is not described.

Follow the deployment options. Pip and Docker are the open path; the others are Alibaba Cloud ECS with a one-click link, ModelScope Studio, and an AgentScope Platform advertised as free and online around the clock. A free hosted assistant is not a business, it is customer acquisition for DashScope keys and ECS instances, and the environment variable the README names first is the DashScope one.

Moat: the parent's model line and the channel integrations Western projects skip. Pricing power: irrelevant; the price is charged elsewhere. Exit: none needed; this is a strategic asset inside a cloud division. Position: long as a product, unpriceable as a company.

The demo is an assistant in a group chat. For sixty seats the price is zero plus model tokens, and the security questionnaire has a rare good page: shell commands run under Seatbelt on macOS, Bubblewrap or Landlock on Linux and AppContainer on Windows, and a self-hosted multi-user Hub arrived in 2.2.0, released the day of this review. SSO and SCIM are not mentioned. Telemetry is a prompt during init, and init --defaults accepts it for you, which a rollout script must avoid.

CI fit is limited to the REST API. Windows LTSC installs need manual path work. Approved with conditions: wait one release on the Hub, then pilot.

Apache-2.0 and Python, with the console built from source by npm ci if I want the web UI from a checkout. Local models are first-class: Ollama with the context length set to 32k or better, LM Studio's local server, or the built-in llama.cpp runtime. Tool Guard is a YAML rule engine that inspects every tool call before it runs, with STRICT, SMART, AUTO and OFF levels, so I tighten or loosen it per install. The connector layer speaks MCP, A2A and ACP with encrypted credentials.

Persona is two files, SOUL and PROFILE, which I can version. Docker keeps secrets in their own volume. This is a fork I would enjoy.