SwarmOps

Orchestrate parallel AI agent workers to build software projects.

SwarmOps manages swarms of AI agents that work together to complete software projects. Built on OpenClaw as its execution engine, it handles the full lifecycle from requirements gathering through code review — spawning workers in parallel, managing git isolation, and coordinating multi-stage review chains.

How It Works

Interview → Specification → Build → Review → Complete

1. Interview — Chat with an AI to gather project requirements
2. Specification — Architect agent designs the solution and decomposes into tasks
3. Build — Multiple builder agents work in parallel, each in isolated git worktrees
4. Review — Sequential review chain (code → security → design) with automatic fixes
5. Complete — Merge to main when all reviews pass

Key Features

• Parallel Execution — Spawn multiple workers simultaneously with dependency-aware scheduling
• Git Isolation — Each worker operates in its own worktree, merged after completion
• Role-Based Agents — Configurable roles with custom models, thinking levels, and prompts
• Review Chain — Multi-stage code review with automatic fix-and-retry loops
• Smart Conflict Resolution — AI-powered merge conflict resolution
• Spawn Safeguards — Circuit breaker, rate limiting, and task deduplication to prevent runaway sessions
• Real-time Dashboard — Monitor workers, view logs, manage projects

Screenshots

Real-time worker monitoring — 24 parallel agents building a project

Task progress view with auto-advance status indicators

Architecture

SwarmOps consists of:

• Dashboard — Nuxt 4 web UI for project management and monitoring
• Orchestrator — Core logic for task scheduling, worker management, and phase transitions
• Gateway Integration — Connects to OpenClaw for AI agent spawning

Quick Start

# Clone and install
git clone https://github.com/siimvene/SwarmOps.git
cd SwarmOps
./install.sh

# Start the server
cd dashboard && node .output/server/index.mjs

The dashboard runs on http://localhost:3000 by default.

Requirements

• Node.js 20+
• OpenClaw Gateway running locally

Project Structure

SwarmOps/
├── dashboard/          # Nuxt 4 web application
│   ├── app/           # Vue components, pages, composables
│   ├── server/        # Nitro API routes and utilities
│   └── public/docs/   # Built-in documentation
├── packages/
│   └── orchestrator/  # Core orchestration logic (TypeScript)
└── data/
    └── orchestrator/  # Runtime data (roles, pipelines, state)

Configuration

Roles

The install script sets up 9 pre-configured roles in data/orchestrator/roles.json:

Role	Model	Purpose
architect	Claude Opus	High-level system design and planning
task-decomposer	Claude Opus	Break down requirements into parallel tasks
builder	Claude Sonnet	Implement features and write code
reviewer	Claude Opus	Code review and quality gates
security-reviewer	Claude Opus	Security-focused vulnerability review
designer	Claude Sonnet	UI/UX design review
researcher	Claude Opus	Deep dives and competitive analysis
workflow-coordinator	Claude Sonnet	Multi-step workflow management
quant-analyst	Claude Sonnet	Trading strategies and market analysis

Each role includes detailed instructions. Customize in the dashboard under Roles.

Environment Variables

Variable	Description
OPENCLAW_GATEWAY_URL	Gateway API URL (default: http://127.0.0.1:18789)
OPENCLAW_GATEWAY_TOKEN	Auth token for dashboard access
PROJECTS_DIR	Path to projects directory
ORCHESTRATOR_DATA_DIR	Path to orchestrator state

Pipeline Flow

%%{init: {'theme': 'neutral'}}%%
flowchart TB
    subgraph Interview["📋 Interview Phase"]
        A[New Project] --> B[Interview Agent]
        B <--> C[User Chat]
        C --> D{Requirements<br/>Complete?}
        D -->|No| B
        D -->|Yes| E[interview.json<br/>complete: true]
    end

    subgraph Spec["🏗️ Specification Phase"]
        E --> F[Architect Agent]
        F --> G[Design Solution]
        G --> H[Decompose into Tasks]
        H --> I[IMPLEMENTATION_PLAN.md]
        I --> J[progress.md]
    end

    subgraph Build["⚡ Build Phase"]
        J --> K{Ready Tasks?}
        K -->|Yes| L[Spawn Parallel Workers]
        L --> M1[Worker 1]
        L --> M2[Worker 2]
        L --> M3[Worker N]
        M1 & M2 & M3 --> O[Merge to Main]
        O --> P{More Tasks?}
        P -->|Yes| K
        P -->|No| Q[All Tasks Done]
    end

    subgraph Review["🔍 Review Phase"]
        Q --> R[Code Review]
        R --> S{Pass?}
        S -->|Fix| R
        S -->|Yes| U[Security Review]
        U --> V{Pass?}
        V -->|Fix| U
        V -->|Yes| X[Design Review]
        X --> Y{Pass?}
        Y -->|Fix| X
        Y -->|Yes| AA[All Reviews Pass]
    end

    subgraph Complete["✅ Complete"]
        AA --> CC[Project Complete]
    end

Loading

Phase Transitions

From	To	Trigger
Interview	Spec	interview.json has complete: true
Spec	Build	specs/IMPLEMENTATION_PLAN.md created
Build	Review	All tasks in progress.md marked [x]
Review	Complete	All reviewers approve

The phase watcher polls every 30 seconds and advances projects automatically when conditions are met.

Documentation

Built-in docs are available at /docs in the dashboard, covering:

• Architecture and system design
• Agent context and prompts
• Operations and monitoring
• Resilience and error handling

License

MIT