MoonWalkers

LLMs such as Gemini are incredibly capable, but they are bottlenecked by the traditional chat interface. Moonclaw breaks autonomous agents out of the text box and drops them into a dynamic, spatial web workspace. Simply type a prompt, and Moonclaw provisions a secure cloud computer, deploys a team of specialized AI agents, and streams their execution live to your browser. This isn't an AI conversation—it's an AI experience. The Problem Current AI models act as advisors. You ask them how to write a script, book a flight, or scrape a website, and they give you a list of instructions to do it yourself. If they hit an error, the loop breaks. They are trapped behind a text input. The Solution Moonclaw transforms the AI from an advisor into an executor. By giving the AI a cloud-based operating system, real developer tools (terminals, browsers, file systems), and a spatial UI to display its work, Moonclaw actually performs the tasks you ask for. It writes the code, runs the tests, reads the errors, and fixes them autonomously. How we built it We split the architecture into three distinct layers: The Brain (Local): A local orchestrator built with CrewAI and LangGraph. It handles the planning, breaks large user prompts into smaller sub-tasks, and coordinates the agent swarm. The Security Proxy: A middleware gateway that acts as a bouncer. It intercepts and blocks dangerous shell commands, restricts outbound domain access, and enforces hard rate limits to stop runaway loops. The Execution Engine (Cloud): The actual work happens inside disposable Ubuntu containers running on GCP Spot instances. These are preloaded with dev utilities and headless browsers, and are destroyed immediately after the task finishes. Observability Dashboard: Instead of trusting the AI blindly, we built a real-time UI that streams WebRTC video and WebSocket terminal logs straight to the frontend. You can sit back and watch the agents click around and type in real-time.