Skip to content

Latest commit

 

History

History
511 lines (352 loc) · 22.6 KB

File metadata and controls

511 lines (352 loc) · 22.6 KB

HiClaw: OpenClaw Evolved — Safer, Easier, Your AI Team in 5 Minutes

Published: February 27, 2026


Sound familiar?

As a heavy OpenClaw user, I know firsthand how powerful it is — one Agent can write code, check emails, operate GitHub. But when you start tackling more complex projects, the cracks show:

Security keeps you up at night: Every Agent needs its own API Key. GitHub PATs and LLM keys are scattered everywhere. The CVE-2026-25253 vulnerability in January 2026 made it clear — this "self-hackable" architecture trades convenience for real risk.

One Agent wearing too many hats: Frontend, backend, docs — all in one. The skills/ directory becomes a mess, MEMORY.md fills up with unrelated context, and every session loads a pile of irrelevant noise.

Multi-Agent coordination without good tooling: Manual config, manual task assignment, manual progress sync. You want to focus on decisions, not babysit AI.

Mobile experience is painful: Want to direct Agents from your phone? Integrating with Slack or Teams bots can take days or weeks.

If any of this resonates, HiClaw was built for you.


What is HiClaw?

HiClaw = OpenClaw, evolved.

The core innovation is the Manager Agent — your AI chief of staff. It doesn't do the work directly; it manages a team of Worker Agents on your behalf.

┌─────────────────────────────────────────────────────┐
│                  Your Local Environment              │
│  ┌───────────────────────────────────────────────┐  │
│  │         Manager Agent (AI Chief of Staff)     │  │
│  │                    ↓ manages                  │  │
│  │    Worker Alice    Worker Bob    Worker ...   │  │
│  │    (frontend)      (backend)                  │  │
│  └───────────────────────────────────────────────┘  │
└─────────────────────────────────────────────────────┘
         ↑
    You (human admin)
    Make decisions, skip the babysitting

Architecture: OpenClaw Gets an Upgrade

OpenClaw's design is like a complete organism: a brain (LLM), a central nervous system (pi-mono), and eyes and a mouth (various Channels). But in the original design, the brain and sensory organs are "external" — you configure the LLM provider and message channels yourself.

HiClaw performs an "organ transplant," turning those external components into built-in organs:

┌────────────────────────────────────────────────────────────────────┐
│                         HiClaw All-in-One                          │
│  ┌──────────────────────────────────────────────────────────────┐  │
│  │                     OpenClaw (pi-mono)                       │  │
│  │                    Central Nervous System                    │  │
│  └──────────────────────────────────────────────────────────────┘  │
│           ↑                              ↑                         │
│  ┌────────────────┐              ┌────────────────┐                │
│  │  Higress AI    │              │   Tuwunel      │                │
│  │  Gateway       │              │   Matrix       │                │
│  │  (brain access)│              │   Server       │                │
│  │                │              │  (sensory org) │                │
│  │  Switch LLM    │              │                │                │
│  │  providers &   │              │  Element Web   │                │
│  │  models freely │              │  (built-in UI) │                │
│  └────────────────┘              └────────────────┘                │
└────────────────────────────────────────────────────────────────────┘

LLM Access: Higress AI Gateway

The brain is no longer external — it's managed through an AI Gateway:

  • One endpoint, many models: Switch between Alibaba Cloud Qwen, OpenAI, Claude, and more from the Higress console
  • Centralized credentials: Configure API Keys once, shared across all Agents
  • Least-privilege access: Each Worker gets call permissions only — never touches the real API Key

Communication: Built-in Matrix Server

The sensory organs are built-in too:

  • Tuwunel Matrix Server: Ready-to-use messaging server, zero configuration
  • Element Web included: Open a browser and start chatting
  • Mobile-friendly: Supports Element, FluffyChat, and other cross-platform clients
  • Zero integration cost: No Slack/Teams bot approval process needed

💡 Think of it this way: vanilla OpenClaw is like a desktop PC — you buy your own GPU (LLM) and monitor (Channel) and install drivers yourself. HiClaw is a laptop — everything's integrated, open the lid and get to work.


Multi-Agent System: Your AI Jarvis

On top of the component integration, HiClaw ships an out-of-the-box Multi-Agent system — Manager Agent coordinates Worker Agents, just like Tony Stark's Jarvis.

On-demand, two modes

The system is opt-in — use it how you want:

Mode 1: Talk directly to Manager

  • Simple tasks go straight to Manager, it handles them
  • Great for quick questions and simple operations

Mode 2: Manager delegates to Workers

  • Complex tasks get broken down and assigned to specialized Workers
  • Each Worker has isolated Skills and Memory
  • No cross-contamination between Workers

Collaboration: Supervisor + Swarm

From the Manager-Worker perspective, this is a Supervisor architecture — Manager as the central coordinator. But because it's built on Matrix group chat rooms, it also has Swarm characteristics.

Shared context, no repeated briefings: Every Agent sees the full group chat history. Alice says "I'm working on the login page," Bob automatically knows what the frontend is doing and can align the API design.

Anti-stampede design: Agents only trigger LLM calls when @mentioned — irrelevant messages don't wake them up, keeping costs predictable.

Artifacts don't pollute context: File exchanges and code snippets flow through the underlying MinIO shared filesystem, not the group chat, so context stays clean.

Security: Manager Manages, But Can't Leak

In vanilla OpenClaw, every Agent holds real API Keys. If one gets compromised or accidentally outputs credentials, you're exposed.

HiClaw's solution: Workers never hold real credentials:

┌──────────────┐      ┌──────────────────┐      ┌─────────────┐
│   Worker     │─────►│  Higress AI      │─────►│  LLM API    │
│  (holds only │      │  Gateway         │      │  GitHub API │
│  Consumer    │      │  (credentials    │      │  ...        │
│  Token)      │      │   centralized)   │      │             │
└──────────────┘      └──────────────────┘      └─────────────┘
  • Workers hold only a Consumer Token (like a badge — grants access, not keys)
  • Real API Keys and GitHub PATs live in the AI Gateway
  • Even if a Worker is compromised, the attacker gets nothing useful

Manager is equally locked down: it knows what tasks Workers are doing, but has no access to API Keys or GitHub PATs. Its job is coordination, not execution.

Dimension Vanilla OpenClaw HiClaw
Credential ownership Each Agent holds its own Workers hold Consumer Token only
Leak surface Agent can output credentials directly Manager has no access to real credentials
Attack surface Every Agent is an entry point Only Manager needs hardening

Human in the Loop: Transparent, Always Interruptible

Compared to OpenClaw's native Sub Agent system, HiClaw's Multi-Agent system is not just easier to use — it's fully transparent:

┌─────────────────────────────────────────────────────────────┐
│                  Matrix Project Room                        │
│                                                             │
│  You: Build a login page                                    │
│                                                             │
│  Manager: Got it, delegating...                             │
│           → @alice frontend page                            │
│           → @bob backend API                                │
│                                                             │
│  Alice: Working on the login component...                   │
│  Bob: API interface defined...                              │
│                                                             │
│  You: @bob wait, password must be at least 8 chars          │  ← intervene anytime
│                                                             │
│  Bob: Updated...                                            │
│  Alice: Got it, frontend validation updated too             │
│                                                             │
│  Manager: Task complete, please review                      │
└─────────────────────────────────────────────────────────────┘

Core advantages:

  • Full visibility: All Agent collaboration happens in the Matrix room
  • Intervene anytime: Spot an issue, @mention the Agent directly
  • Natural interaction: Like working with a team in a group chat

Manager's Core Capabilities

Capability Description
Worker lifecycle management "Create a frontend Worker" → auto-configures, assigns skills
Automatic task delegation You state the goal, Manager breaks it down and assigns it
Heartbeat monitoring Periodically checks Worker status, alerts you if something's stuck
Project room setup Creates Matrix Rooms for projects, invites relevant participants

Worker Skills: A Safe Open Ecosystem

OpenClaw has a great open skills ecosystem at skills.sh — 80,000+ community skills you can install in one command: Higress WASM plugins, PR reviews, changelog generation...

But in vanilla OpenClaw, you might hesitate to use them freely. A public SKILL.md you haven't audited could trick an Agent into outputting credentials or running dangerous commands — because the Agent holds your API Keys.

In HiClaw, it's a different story. Every Worker runs in an isolated container and is designed to hold no real credentials:

What can a Worker see?
✅ Task files, code repos, its own working directory
✅ Consumer Token (like a keycard — grants API access only)
❌ Your LLM API Key
❌ Your GitHub PAT
❌ Any sensitive credentials

Even if a skill tries to steal credentials, there's nothing to steal. Workers can freely pull skills from the public registry as needed.

HiClaw ships Workers with a built-in find-skills skill. When a Worker encounters a task requiring specialized knowledge, it searches and installs the right skill automatically:

Manager assigns: "Build a Higress WASM Go plugin"
                  ↓
Worker finds it lacks the right tools
                  ↓
skills find higress wasm
  → alibaba/higress@higress-wasm-go-plugin  (3.2K installs)
                  ↓
skills add alibaba/higress@higress-wasm-go-plugin -g -y
                  ↓
Skill installed — Worker has full plugin dev scaffolding and workflow

If you prefer a private registry, HiClaw supports that too — choose Manual mode during install and set your private Skills Registry URL. You can also tell Manager to use a specific registry when creating Workers at any time.

As long as your private registry implements the same API as skills.sh, Workers switch seamlessly. The usage pattern is identical either way.

Mobile Experience

HiClaw's built-in Matrix server supports multiple clients:

  • Works right after install: No Slack/Teams bot setup needed
  • Direct from your phone: Download any Matrix client (Element, FluffyChat, etc.)
  • Real-time push notifications: Not buried in a "bot" folder
  • Full visibility: You, Manager, and Workers all in the same Room

💡 Supported clients: Element, FluffyChat, and other major Matrix clients — iOS, Android, and Web.


Get Started in 5 Minutes

Step 1: Install

macOS / Linux:

bash <(curl -sSL https://higress.ai/hiclaw/install.sh)

Windows (PowerShell 7+):

Set-ExecutionPolicy Bypass -Scope Process -Force; $wc=New-Object Net.WebClient; $wc.Encoding=[Text.Encoding]::UTF8; iex $wc.DownloadString('https://higress.ai/hiclaw/install.ps1')

⚠️ Windows users need PowerShell 7+ and Docker Desktop first.

What the installer does:

  • Cross-platform: bash on Mac/Linux, PowerShell on Windows — consistent experience
  • Smart mirror selection: Picks the nearest image registry based on your timezone
  • Docker-based: All components run in containers, OS differences abstracted away
  • Minimal config: Just one LLM API Key required, everything else is optional

After install, you'll see:

=== HiClaw Manager Started! ===

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
  ★ Open the following URL in your browser to start:                           ★
                                                                                
    http://127.0.0.1:18088/#/login
                                                                                
  Login with:                                                                   
    Username: admin
    Password: [auto-generated password]
                                                                                
  After login, start chatting with the Manager!                                 
    Tell it: "Create a Worker named alice for frontend dev"                     
    The Manager will handle everything automatically.                           
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

💡 No hosts file edits needed: *-local.hiclaw.io resolves to 127.0.0.1 automatically.

Step 2: Log in and start chatting

  1. Open the URL shown after install (e.g. http://127.0.0.1:18088)
  2. Log in with the credentials shown during install
  3. You'll see a "Manager" conversation waiting for you

Step 3: Create your first Worker

You: Create a Worker named alice for frontend development

Manager: On it...
         Worker alice created. Room: !xxx:matrix-local.hiclaw.io
         You can assign tasks to alice directly in the "Worker: Alice" room

Step 4: Assign a task

You: @alice please build a simple login page using React

Alice: On it...
       [a few minutes later] Done! Code committed to GitHub, PR: https://github.com/xxx/pull/1

Step 5: Check progress from your phone

  1. Download a Matrix client (Element, FluffyChat — iOS/Android/Web)
  2. At login, choose "Other server" and enter your Matrix server address
  3. Monitor Worker progress and intervene anytime

Solo Founder Playbook: One Person, One Team

Say you want to build a SaaS product — from idea to launch to growth. Traditionally you'd need product, design, engineering, QA, marketing... But now:

You: Create 4 Workers:
     - alex: product manager
     - sam: full-stack developer
     - taylor: content & marketing
     - jordan: data analyst

Manager: Done. 4 Workers created, each with independent skills and memory.

You: Kick off project: AI Writing Assistant MVP, launch target this week

Manager: Got it, planning now...

         [Creates project room, invites all Workers]
         [Drafts project plan]
         
         Here's the plan, please confirm:
         
         Phase 1: Product Definition
         - task-001: Competitive research & PRD (alex)
         
         Phase 2: Development
         - task-002: Tech stack & architecture (sam, depends on task-001)
         - task-003: Core feature development (sam, depends on task-002)
         - task-004: Product review (alex, depends on task-003)
         
         Phase 3: Launch Prep
         - task-005: Launch assets (taylor, depends on task-001)
         
         Confirm to start assigning.

You: Confirmed

Manager: [Sets project status to active]
         
         @alex New task [task-001]: Competitive research & PRD
         Goal: Output PRD defining MVP feature scope
         Spec: ~/hiclaw-fs/shared/tasks/task-001/spec.md
         Report back in the project room when done

[2 hours later]

alex: @manager task-001 complete
      PRD done: core features are AI autocomplete, multi-model switching, history
      Result: ~/hiclaw-fs/shared/tasks/task-001/result.md

Manager: [Reads result.md, status: SUCCESS]
         [Updates plan, marks task-001 complete]
         
         @sam New task [task-002]: Tech stack & architecture
         PRD reference: ~/hiclaw-fs/shared/tasks/task-001/result.md
         
         @taylor New task [task-005]: Launch assets
         Product positioning: ~/hiclaw-fs/shared/tasks/task-001/result.md

[4 hours later]

sam: @manager task-002 complete
     Stack: Next.js + Vercel + Supabase
     Estimated 8 hours for core development

Manager: [Updates plan, assigns task-003]
         
         @sam Continue with core feature development [task-003]

[8 hours later]

sam: @manager task-003 complete, deployed to Vercel
     Demo: https://xxx.vercel.app

Manager: [Updates plan, assigns task-004]
         
         @alex Please review [task-004]: Product acceptance
         Test URL: https://xxx.vercel.app
         Verify feature completeness against PRD

[Issue found during review]

alex: @manager task-004 needs revision
      Issue: multi-model switching has no onboarding — users don't know how to pick
      Suggestion: add a model comparison page
      Result: ~/hiclaw-fs/shared/tasks/task-004/result.md
      Status: REVISION_NEEDED → back to task-003

Manager: [Reads result.md, status: REVISION_NEEDED]
         [Creates revision task task-006]
         
         @sam task-004 review found issues, revision needed
         
         Revision task [task-006]: Add model selection onboarding page
         Feedback: ~/hiclaw-fs/shared/tasks/task-004/result.md
         Report back when done

[Revision complete]

sam: @manager task-006 complete, onboarding page added

Manager: [task-006 done, task-004 re-review passed]
         [All development tasks complete]
         
         @you MVP development complete!
         - Deployed: https://xxx.vercel.app
         - Product review passed
         - Launch assets ready
         
         Project plan: ~/hiclaw-fs/shared/projects/proj-xxx/plan.md

You: Ready to launch. @taylor Product Hunt tomorrow

taylor: Got it, scheduled. Assets are ready.

[Launch day — Manager auto-monitors progress]

Manager: @you Product Hunt launch update
         Current rank: #3
         Upvotes: 423
         Comments: 87
         
         @jordan please set up analytics tracking

jordan: On it, configuring GA4 + custom events...

[Data ready]

jordan: @manager tracking setup complete
        Dashboard: https://analytics.google.com/xxx
        
        Day 1 data:
        - Signups: 1,247
        - Day 2 retention: 34%
        - AI autocomplete usage: 78%
        - Multi-model switching usage: 23%

Manager: @you Project "AI Writing Assistant MVP" — Day 1 Report
         
         Key metrics:
         - Day 1 signups: 1,247
         - Day 2 retention: 34%
         - Feature usage: autocomplete 78%, model switching 23%
         
         Insight: model switching usage is low
         Suggestion: @alex investigate and optimize the onboarding flow

[And so it goes — Manager runs the whole loop: plan → assign → monitor → coordinate → report]

What did Manager actually do?

Phase Manager's role
Project planning Breaks goal into tasks, identifies dependencies
Task assignment @mentions Workers with task context
Progress tracking Updates plan on completion, triggers next steps
Issue handling Review fails → auto-creates revision task
Status reporting Proactively updates you at key milestones
Risk flagging Spots anomalies in data, suggests optimizations

You make the decisions. Manager handles the rest.


Open Source


Closing Thoughts

HiClaw is an evolution of OpenClaw — not a replacement, an upgrade.

We kept everything that makes OpenClaw great (natural language interaction, Skills ecosystem, MCP tooling) while fixing the security and usability pain points.

If you're a:

  • Solo developer who wants to do a team's worth of work alone
  • Power OpenClaw user who wants a safer, smoother experience
  • Solo founder who needs AI teammates to share the load

HiClaw is for you.

Start now:

bash <(curl -sSL https://higress.ai/hiclaw/install.sh)

HiClaw is open source under the Apache 2.0 license. If you find it useful, a Star ⭐ and contributions are always welcome!