What This Stack Does
Building production-ready AI agents requires more than just an LLM and a prompt. Agents need safe execution environments, efficient web access, structured data extraction, and secure infrastructure connectivity. This stack combines purpose-built tools for each layer of the agent infrastructure, all open-source and designed to work together through standard protocols like CDP and MCP.
Sandboxed Agent Execution
CUA provides the sandbox layer where agents operate. Its cross-platform virtual machines give agents full desktop control — screenshots, mouse/keyboard, shell commands — in isolated environments that protect the host system. The MCP integration means agents built with CUA can be orchestrated from Claude Desktop, Cursor, or any MCP client.
Web Access and Data Extraction
Lightpanda handles the web browsing layer with 11x faster execution and 9x less memory than Chrome. When agents need to navigate websites, fill forms, or verify content, Lightpanda's CDP-compatible headless browser provides the most efficient execution possible. The built-in MCP server allows direct agent-to-browser communication.
Crawl4AI adds the data extraction layer, converting web pages into clean Markdown optimized for RAG pipelines and LLM consumption. Deep crawling with link discovery, LLM-based content extraction, and proxy rotation create a complete pipeline from URL to structured data.
Secure Infrastructure Access
Teleport Beams secures the infrastructure access layer. When agents need to interact with production servers, Kubernetes clusters, or databases, Beams provides cryptographically verified sessions with policy enforcement and audit trails. This defense-in-depth approach ensures agents cannot exceed defined access boundaries even if compromised.
The Bottom Line
Together these tools create a layered architecture where each component handles its specialty: CUA for desktop automation, Lightpanda for web browsing, Crawl4AI for data extraction, and Teleport Beams for infrastructure security. The MCP protocol serves as the integration layer connecting agent orchestrators to each capability.