Skip to main content

Projects

Real-world AI agent systems, automation pipelines, and security infrastructure.

OpenClaw Multi-Agent Orchestration

Core multi-agent system with specialized roles (orchestrator, monitor, messenger, coder, security analyst) coordinated through a central brain. Built on local LLMs with secure tool access and state persistence.

🎯 Orchestrator (Ziggy)

Primary coordination & reasoning

  • • Complex task decomposition
  • • Sub-agent spawning & management
  • • Multi-step workflow execution

👁 Monitor (Argus)

Heartbeats & status checks

  • • Calendar/email polling
  • • System health monitoring
  • • Light triage work

🔥 Coder (Hephaestus)

Code review & implementation

  • • PR review & debugging
  • • Feature implementation
  • • Build/test automation

Key Features

  • • Effort scaling heuristics (1–10 agents based on complexity)
  • • Kill criteria for stuck agents (3+ iterations, scope violation)
  • • TaskFlow checkpointing for resume-from-failure
  • • Interface contracts prevent merge conflicts
  • • LLM-as-judge evaluation rubric (≥20/25 pass threshold)
Location: /workspace/skills/orchestration/
Status: ✅ Production — 6 parallel-agent skills implemented and enabled

Lite-LCM — Lossless Context Management

SQLite-backed context management system for long-running AI agent sessions. Provides message ingestion, token counting, context assembly, and multi-tier recursive condensation to maintain coherent session history without token limits.

Core Features

  • • Message ingestion with token counting
  • • Context assembly (evictable prefix + fresh tail)
  • • Leaf-only compaction (v0.1)
  • • Recursive multi-tier DAG condensation (v0.4)
  • • Transaction mutex prevents concurrent compaction

CLI Commands

  • • init, ingest, assemble, compact
  • • status, backup, grep, expand
  • • doctor, cleanup
  • • Debug mode: LITE_LCM_DEBUG=1

Test Results

250 messages → 11 leaf summaries → 4 condensed → 1 super-condensed. All 10 CLI commands tested and passing. Local model (qwen2.5:7b) required for speed.

Location: /workspace/skills/auto-generated/lite-lcm/
Database: ~/.openclaw/lite-lcm/lcm.db
Status: ✅ v0.4.0 tested and enabled in production

Ares Threat Modeling Integration

MCP-enabled cybersecurity agent with AWS Labs Threat Modeling MCP Server. Complete STRIDE-based threat modeling workflow with 9-phase process, AWS documentation validation, and code-based remediation verification.

9-Phase Process

  • • Business context analysis
  • • Architecture component mapping
  • • Threat actor analysis
  • • Trust boundary detection
  • • Asset flow analysis
  • • Threat identification (STRIDE)
  • • Mitigation planning
  • • Code validation (Phase 7.5)
  • • Comprehensive export

MCP Integration

  • • AWS Labs Threat Modeling MCP v1.27.1
  • • 100+ threat modeling tools
  • • Mandatory AWS doc validation
  • • Kali Linux container (ares-kali:mcp)
  • • nmap, nikto, burp integration
Location: /workspace/security/sample-threat-model/
Config: MCP server enabled in openclaw.json
Status: ✅ Production ready

The Shield — Web Scraper Defense Bypass

Advanced web scraping module with TLS fingerprint bypass, rotating proxy management, CAPTCHA solving, and anti-detection heuristics. Built as an OpenClaw skill for autonomous research and data collection.

Components

  • • TLS fingerprint bypass (curl-cffi)
  • • Rotating proxy manager (residential/datacenter)
  • • CAPTCHA solver integration (CapSolver, 2Captcha)
  • • Block detection heuristics
  • • Adaptive retry logic with exponential backoff

Output Formats

  • • Markdown, JSON, JSONL, CSV
  • • Static fetch + dynamic rendering (Playwright)
  • • CSS selector extraction
  • • Batch mode processing
  • • robots.txt compliance

Test Results

TLS bypass verified against httpbin.org (200 OK). Residential proxy and CAPTCHA service integration pending API key configuration.

Location: /workspace/skills/auto-generated/web-scraper/shield/
Status: 🧪 v1.0.0 tested, awaiting production API keys
Est. Cost: $46-66/month for 10K pages (Smartproxy + CapSolver)

Plugin Enhancement Sprint System

Automated plugin development workflow for OpenClaw skills. Handles installation, inspection, packaging, and publication to ClawHub with metadata validation and error recovery for common issues (entry not found, stale metadata, package.json misconfig).

Workflow Steps

  • • openclaw plugins install — Fetch from ClawHub or local path
  • • openclaw plugins inspect — Validate metadata, entry points, permissions
  • • openclaw plugins pack — Bundle with manifest generation
  • • openclaw plugins publish — Upload to ClawHub with versioning
Location: /workspace/skills/auto-generated/plugin-enhancement-sprint/
Status: ✅ Active — used for skill development and deployment

Static Site Pipeline (Next.js 16 + Tailwind v4)

Production-ready static site generator optimized for OpenClaw documentation, portfolios, and landing pages. Includes security-hardened configuration, automated builds, and deployment to shared hosting.

Features

  • • Next.js 16.1.1 static export
  • • Tailwind v4 with custom theme
  • • Security headers (.htaccess): CSP, HSTS, Referrer-Policy
  • • SEO artifacts: robots.txt, sitemap.xml, JSON-LD
  • • Accessibility: skip links, focus indicators, reduced motion
  • • Contact form with SMTP → Protonmail alias

Deployment

  • • Namecheap shared hosting compatible
  • • Single zip deploy package (~318K)
  • • Zero server-side dependencies
  • • Automated build verification
Location: /workspace/skills/auto-generated/static-site-pipeline/
Live Site: ginsberg.xyz
Status: ✅ Deployed and audited (A+ security headers)

Learning Loop Integration

Automated self-improvement system with cron-driven skill generation, memory nudges, and pattern analysis. Continuously identifies novel workflows and converts them into reusable AgentSkills.

Memory Nudge (6hr)

Prompts review of recent sessions for patterns and insights

Auto Skill Generator (2hr)

Detects novel multi-step workflows and generates skills

Skill Improvement (daily)

Reviews and refactors existing skills for quality

Validation: 3–7 day tracking period via /workspace/skills/auto-generated/TEST_LOG.md
Status: ✅ Deployed — decision pending after validation period