My Most INSANE AI Agent Ever (OpenClaw Clone)

An AI agent built around a dedicated Mac mini can research on social platforms, schedule autonomous tasks, and even generate and publish videos—without relying on API keys. The workflow starts with a WhatsApp command that triggers Claude Code to navigate sites like X, gather context by opening posts and tweets, and send back findings in chat. In the demonstration, the agent searches for “Maltbook,” reads what’s being discussed, and returns a message summarizing what it found—positioning the system as a practical “do the browsing for me” layer rather than a purely conversational chatbot.

The core design choice is operational: the agent runs inside a browser session where accounts are already logged in (YouTube, X, LinkedIn, Gmail, and more). Instead of storing and using API keys, the system leverages authenticated web access, which the creator frames as both more secure and easier to keep running. Claude Code runs continuously on the Mac mini, tied to a $200 subscription, so scheduled and on-demand actions can happen without manual supervision.

Beyond research, the agent adds a scheduling layer using cron jobs. The system can automatically open sites such as Hacker News, check the top post, navigate into it, and produce summaries—then repeat on a daily cadence. A live test schedules a job for 15:46 (about a minute later), and the machine proceeds without direct control, demonstrating that the agent can operate like a background newsroom: collect, summarize, and potentially post or email results while the user is away. Jobs can also be removed from the queue to stop recurring behavior.

The most ambitious capability is video production. The agent is trained with a skill system that includes X skills, video research, video editing (using Remotion), thumbnail creation, and YouTube publishing. In an example workflow, the user provides a URL to a trending Google project (“Project Genie”), then instructs the agent to plan a story-driven video: use main clips as silent background, add voiceover, and insert additional “spots” from clips. The agent runs in an autonomous pipeline—researching, drafting a plan, downloading or reusing clips, assembling scenes, and stitching the final output—before producing a near-finished video.

The resulting video narration focuses on Project Genie’s “world sketching,” where users describe a world in plain text and then “walk through it,” including steps like choosing a character and fine-tuning the world. After generating content, the system can extend into distribution: the creator describes posting to YouTube and then sharing outward to X, with the broader goal of automating social media workflows.

Overall, the project’s pitch is less about a single flashy model and more about an end-to-end agent stack: authenticated browsing for context, cron-based automation for routine tasks, and a Remotion-backed skill pipeline for producing and publishing videos—then iterating by expanding the agent’s skills over time, including more complex actions like applying for jobs.