Skip to content
ReviewRamble
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

Also inEspañolDeutschFrançais

JiuwenSwarm

Huawei-backed multi-agent workbench in Python: a Leader agent splits tasks across a team, reachable from web, terminal and chat apps.

Reviewed by Sheela · Updated

View on GitHub Open source

Our rating

3.6/5

JiuwenSwarm is the serious multi-agent choice for teams that work in Feishu, DingTalk or WeCom, but the config it ships leaves shell and file tools on allow, whatever the README says.

What is JiuwenSwarm?

JiuwenSwarm, renamed from JiuwenClaw and now released as WorkSwarm (PyPI package workswarm), is a self-hosted agent system from openJiuwen. openJiuwen is an open-source community whose code carries Huawei Technologies copyright notices and whose GitHub organisation is listed in China. A Leader agent breaks a task down and assigns the parts to specialist agents, which can run across processes or machines; a single-agent mode handles everyday requests. It has Work and Code spaces, plus Design since v0.2.8, a web UI, a terminal UI, desktop apps for Windows, macOS and HarmonyOS, and nine chat channels: Xiaoyi, Feishu, DingTalk, WeCom and personal WeChat in China, and Telegram, Discord, Slack and WhatsApp elsewhere.

The repo was created on 5 March 2026 and has about 8,470 GitHub stars, 1,150 forks, 1,246 open issues and 609 open pull requests, with 129 contributors listed. That star count is unremarkable for a project with this much visible activity. It is Apache-2.0 Python. v0.2.8 was tagged on 31 October 2026; the newest version on PyPI is 0.2.7, and workswarm had 871 downloads in the latest week. Models come from Huawei Cloud MaaS, OpenAI, DeepSeek, DashScope, SiliconFlow, OpenRouter, other OpenAI-compatible APIs or a local server.

Is JiuwenSwarm worth it?

JiuwenSwarm is less a personal assistant than a team of agents with a chat front end. OpenClaw and nanobot centre on one agent that remembers you; JiuwenSwarm's default Cluster mode has a Leader decompose the job and hand pieces to specialists. Swarmflow scripts add fixed multi-stage pipelines with a team token budget and human sign-off steps. v0.2.8 adds verifier and challenger agents that vote on a result, and lets Claude Code or Codex join a team as outside experts. Neither rival goes this far on orchestration.

The security story needs close reading. The README says tools ask for approval before they run, and the permission engine documented in the repo is thorough: allow, ask and deny levels, built-in rules for risky shell commands, and a whitelist for paths outside the workspace. Yet in our reading of the config template the package ships, unchanged on this point at v0.2.7 and v0.2.8, permissions.enabled is false and bash, PowerShell, file writes and edits are all set to allow. The docs say a disabled engine allows every tool call. An operator who connects WhatsApp or Telegram without editing that file is handing a model a shell. The optional jiuwenbox sandbox wraps tools in bubblewrap with seccomp and Landlock, but only on Linux and as a separate service.

The data defaults are quieter. OpenTelemetry export is off, and points at localhost when switched on; team traces go to a local file. Prompts go to whichever model provider you configure. The desktop client's trial login, added in v0.2.7, grants a quota of hosted model resources, and the release notes do not say who runs them or what is kept. No GitHub security advisories have been published, which for a project this young more likely reflects less outside scrutiny than OpenClaw receives than a clean record.

Its centre of gravity is China. Five of the nine channels are Chinese platforms, Huawei Cloud MaaS is the first-listed provider, the README recommends a Tsinghua PyPI mirror, and comments in the config template are in Chinese. It has been renamed twice and has published 32 GitHub releases, 14 of them pre-releases, in under eight months. Against nanobot it is far larger and harder to audit. Against OpenClaw it reaches fewer Western chat apps and has a much smaller community, but it goes further on coordinating many agents through one long task.

Where it earns its place

  • A platform team at a Chinese manufacturer that runs on Feishu

    Give staff one agent in the group chat that researches questions and drafts reports

    Enable the Feishu channel in config.yaml with the bot credentials; in groups people @mention the bot. Point models at Huawei Cloud MaaS or DeepSeek, and set permissions.enabled to true with bash on ask before anyone outside the team can message it.

  • An analyst at a strategy consultancy

    Turn a broad brief into a sourced industry report overnight

    Cluster mode has the Leader split research, analysis and writing across specialist agents, with a team token budget capping spend. The bill is model tokens, which multiply with the number of agents, so a cheaper model for teammates keeps it down.

  • A backend engineer comparing multi-agent frameworks

    Run a fixed multi-stage pipeline with a human approval step

    Write a Swarmflow Python script with human stages and watch runs from the TUI with /swarmflows. Run tools inside jiuwenbox on Linux if the pipeline reads untrusted input; it needs bubblewrap, and NET_ADMIN for network isolation.

Before you commit

Setup
pip install workswarm (the README suggests a Tsinghua mirror in China), then jiuwenswarm-init and jiuwenswarm-start, and open http://localhost:5173. Python 3.11 or later; PyPI's 0.2.7 requires below 3.14. A default model must be set in the web UI or ~/.jiuwenswarm/config/config.yaml before the first reply. Desktop installers cover Windows, macOS and HarmonyOS PCs.
Integrations
Xiaoyi, Feishu, DingTalk, WeCom, personal WeChat, Telegram, Discord, Slack and WhatsApp; MCP servers, A2A and the Swarm Skills Hub; OpenAI- and Anthropic-protocol providers including Huawei Cloud MaaS, DeepSeek, DashScope, SiliconFlow, OpenRouter, Ollama, LM Studio, vLLM and SGLang; Claude Code and Codex as team members from v0.2.8.
Data and privacy
Config, memory and sessions live under ~/.jiuwenswarm; prompts go to your model provider. OpenTelemetry export is off by default, and when enabled it records full message content unless log_messages is turned off. No GitHub security advisories published. Apache-2.0; the README puts EU AI Act compliance on the user.
Cost at scale
No paid edition is published. Cost is model tokens, multiplied by the number of agents in Cluster mode, plus a machine for the gateway; distributed teams run the Leader and teammates as separate processes or hosts.
Limits
The shipped config template turns tool approvals off. 1,246 open issues and 609 open pull requests. Three names (JiuwenClaw, JiuwenSwarm, WorkSwarm) scatter docs and packages. Xiaoyi and WhatsApp are private chat only.

What’s good

  • Leader-and-team orchestration with Swarmflow pipelines and token budgets
  • Native Feishu, DingTalk, WeCom, WeChat and Xiaoyi channels
  • Detailed allow/ask/deny permission engine and an optional bubblewrap sandbox
  • Apache-2.0, with OpenTelemetry export off by default

What’s not

  • Shipped config disables the permission engine and allows shell
  • Renamed twice; PyPI lags GitHub releases
  • Config comments and some messages are in Chinese only
  • About 1,250 open issues and a fast beta cadence

Best for

Engineering teams on Feishu, DingTalk or WeCom who want several agents to split long research or coding jobs and will lock down permissions themselves.

JiuwenSwarm alternatives

OpenClaw

Agent frameworksTypeScript394k stars

4.3/5

Self-hosted personal AI assistant that runs on your machine and answers in WhatsApp, Telegram, Slack, iMessage and 20+ other chat apps.

nanobot

Agent frameworksPython50k stars

4.0/5

Lightweight self-hosted personal agent in Python: WebUI, terminal and chat apps, with memory, schedules, MCP and an OpenAI-compatible API.

CrewAI

Agent frameworksPython60k stars

4.1/5

Python framework for teams of role-playing agents (Crews) plus event-driven Flows for predictable production pipelines.

Hermes Agent

Agent frameworksPython257k stars

4.3/5

Nous Research's always-on personal agent: runs on a server, talks over Telegram or Slack, and writes its own skills as it goes.

JiuwenSwarm: common questions

Is JiuwenSwarm worth it?
JiuwenSwarm is the serious multi-agent choice for teams that work in Feishu, DingTalk or WeCom, but the config it ships leaves shell and file tools on allow, whatever the README says. We rate it 3.6 out of 5.
How much does JiuwenSwarm cost?
Free and open source (Apache-2.0). You pay your model provider, or run a local model.
What are the downsides of JiuwenSwarm?
Shipped config disables the permission engine and allows shell; renamed twice; PyPI lags GitHub releases; config comments and some messages are in Chinese only; about 1,250 open issues and a fast beta cadence.
Who is JiuwenSwarm best for?
Engineering teams on Feishu, DingTalk or WeCom who want several agents to split long research or coding jobs and will lock down permissions themselves.
How much does JiuwenSwarm cost for a team or in production?
No paid edition is published. Cost is model tokens, multiplied by the number of agents in Cluster mode, plus a machine for the gateway; distributed teams run the Leader and teammates as separate processes or hosts.
Is JiuwenSwarm safe for company data?
Config, memory and sessions live under ~/.jiuwenswarm; prompts go to your model provider. OpenTelemetry export is off by default, and when enabled it records full message content unless log_messages is turned off. No GitHub security advisories published. Apache-2.0; the README puts EU AI Act compliance on the user.
What are the best JiuwenSwarm alternatives?
The closest alternatives we have reviewed are OpenClaw, nanobot, CrewAI and Hermes Agent.

Answers checked Oct 31, 2026.