A true AI agent for pixel-perfect web cloning. Multi-agent architecture built on Claude Agent SDK with 40+ specialized tools. Clones from CSS & structured blocks—not screenshots—enabling perfect reproduction where single-model tools like Cursor/Claude Code fail.
A true AI agent — not just a wrapper around an LLM. Multi-agent collaboration with real tools, self-correction loops, and a complete sandbox environment to build production-ready code from scratch.
Others guess from screenshots. We extract the real code — DOM, styles, components, interactions. that screenshot-based tools simply cannot achieve.
This entire multi-agent system is open source. Learn from it, use it, build upon it.
Why Cowork?
Working with a single AI is like asking one coworker to handle everything alone — they get overwhelmed. Traditional single-model approaches hit a wall:
Context window overflow on large pages
Hallucinations when juggling too many responsibilities
Slow, sequential processing
Our solution: A team of specialized agents working together, like coworkers each focused on what they do best. Less back-and-forth, more getting things done.
Why Not Just Use Cursor / Claude Code / Copilot?
We tried. Even with the complete extracted JSON — full DOM tree, all CSS rules, every asset URL — single-model tools struggle:
| Challenge | | Multi-Agent |
|-----------|-------------------------------|---------------------|
| 50,000+ line DOM tree | ❌ Context overflow, truncates critical parts | ✅ Task distributed to worker agents |
| 3,000+ CSS rules | ❌ Loses specificity, misses variables | ✅ Parallel processing by workers |
| Component detection | ❌ Guesses boundaries, creates monoliths | ✅ Structured workflow ensures consistency |
| Responsive breakpoints | ❌ Often hardcodes single viewport | ✅ Extracts all media queries |
| Hover/animation states | ❌ Cannot see, cannot reproduce | ✅ Browser automation captures all |
| Output quality | ❌ "Close enough" approximation | ✅ Pixel-perfect, production-ready |
The core problem: A 200KB extracted JSON exceeds practical context limits. Even if it fits, a single agent can't maintain coherence when context gets polluted with too much information. The solution isn't smarter agents — it's task distribution with reliable workflows.
Honest limitation: Complex animations are still hard to extract perfectly — but that's a crawler problem, not an agent problem. The multi-agent architecture itself is capable of far more than web cloning. Imagine: automated refactoring, codebase migration, documentation generation, or any task too large for a single agent to handle.
This pattern — Agent + Tools + Sandbox — is reusable for any AI agent product:
| Component | Purpose | In Nexting |
|-----------|---------|------------|
| Main Agent | Coordinates workflow and distributes tasks | Analyzes page, spawns workers, merges results |
| Worker Agents | Execute subtasks with scoped permissions | Process sections in parallel, isolated context |
| Tools | Capabilities agents can invoke | File I/O, Browser automation, API calls |
| Sandbox | Safe execution environment | BoxLite - Embedded micro-VM runtime |
BoxLite: Hardware-level isolated micro-VMs for AI agents. No root access needed, runs OCI containers with true kernel isolation. → github.com/boxlite-ai/boxlite
Connect With Me
Building something with this architecture? Have questions? Reach out:
Built on Claude Agent SDK — the same foundation as Claude Code. This isn't a chatbot with API calls; it's a real agent that thinks, plans, executes, and self-corrects in an isolated sandbox.
The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.
