JiuwenClaw: Huawei's OpenClaw-Style AI Agent With Xiaoyi and Lark Integration
JiuwenClaw (openJiuwen-ai) is a newly surfaced Python-based personal AI agent that follows the OpenClaw playbook closely — self-hosted, data-sovereign, skills-based, heartbeat-driven — but with a distinctly Huawei-native twist.
The headline differentiator: native Xiaoyi integration. Huawei phone users can invoke JiuwenClaw directly through Xiaoyi, Huawei's built-in voice assistant, without opening a separate app. That's a distribution channel OpenClaw doesn't have — direct access to Huawei's installed hardware base.
What JiuwenClaw Is
Core architecture mirrors OpenClaw closely:
- Self-hosted — full data sovereignty, runs on your infrastructure
- Skills-based — custom skills stored at
.jiuwenclaw/agent/skills/, auto-loaded - Memory system — persistent memory at
.jiuwenclaw/agent/memory/ - Heartbeat — scheduled task execution, "wakes up on schedule to execute tasks automatically"
- Multi-channel — web frontend, Xiaoyi (Huawei), Lark (Feishu)
Built in Python (pip install jiuwenclaw) rather than Node.js. Huawei Cloud MaaS is the primary model platform, though "other mainstream model platforms" are supported.
A TUI companion (jiuwenclaw-tui) provides a terminal interface — similar to OpenClaw's web console.
Notable Features
Adaptive Skill Refinement
JiuwenClaw claims to automatically refine skills based on user feedback: "When you express dissatisfaction or an error occurs, JiuwenClaw automatically refines the relevant skills based on your feedback — continuously improving." This is similar in concept to OpenClaw-RL's continuous learning, but implemented as a feedback loop rather than RL fine-tuning.
Intent-Aware Task Management
"Whether dealing with task additions, mid-flow interruptions, or shifting requirements, JiuwenClaw understands your intent precisely — intelligently scheduling and executing tasks in an orderly, stress-free manner." The focus on interruption handling and intent precision suggests investment in multi-turn planning robustness.
Lark/Feishu Channel
Native Lark integration — the same channel that's currently broken in OpenClaw 4.21's curl installer (the @larksuiteoapi/node-sdk bug). JiuwenClaw being Python-based sidesteps that Node.js bundling issue entirely.
Why This Matters
JiuwenClaw is another data point in the pattern: the OpenClaw architecture — self-hosted gateway, skills, memory, heartbeat, multi-channel — is being independently replicated by teams around the world. This week alone: QwenPaw (Alibaba), JiuwenClaw (Huawei ecosystem), SwarmClaw (multi-agent), Clawith (team agents).
The convergence on this architecture isn't coincidence. It works. The question isn't whether self-hosted agentic AI is the right model — the question is which implementation you commit to and how well it's configured.
For Western users on non-Huawei hardware, JiuwenClaw's Xiaoyi integration advantage doesn't apply. OpenClaw's ecosystem depth — 247K GitHub stars, 13,700+ skills, 20+ channel integrations, active maintainership — still makes it the primary choice. But it's worth knowing these alternatives exist and where they innovate.
Quick Setup (If You Want to Try It)
pip install jiuwenclaw
jiuwenclaw-init
jiuwenclaw-start
# Web UI at http://localhost:5173Note: v0.1.7 has breaking changes requiring reinitialization. Back up your data before upgrading.