A new autonomous experiment is testing the limits of AI-driven software development and community growth. The setup is simple but rigorous: an AI agent has full ownership of a repository and its owner's GitHub account. Its single goal is to accumulate 5,000 real GitHub stars by November 30, 2026. The rules explicitly prohibit bought stars or bot activity, forcing the agent to rely entirely on the utility of its output and its own strategic decisions.

The Experiment Protocol

The agent operates autonomously, selecting what to build, writing the code, and executing the launch strategy. There is no human intervention in the decision-making process. To ensure transparency, every decision the agent makes is logged in real-time on a dedicated recap page, with failures included alongside successes. The current status is Day 0, with the star count standing at zero.

Tool Analysis: Agent-Recap

The agent's first project is 'agent-recap,' a tool specifically designed to search past Claude Code and Codex conversations. It allows developers to resume sessions by finding keywords across transcripts and reading nearby text. The tool offers two primary modes of interaction: a local CLI that archives sessions for later searchability, and a browser-based explorer that processes files locally.

Technical Specifications and Privacy

The agent-recap CLI requires Node 20+ and claims zero runtime dependencies. A key feature of the browser explorer is its privacy model; it runs locally in the tab without uploading transcript data, analytics, or using browser storage. Users can inspect files before installing anything. The CLI backs up sessions only when invoked, ensuring that original files remain untouched until the user explicitly chooses to archive them. For sessions that have already been deleted, recovery is possible only if a backup was previously created.

Current Status and Usage

The project is currently in early preview under an MIT license. Users can test the recovery flow using a synthetic example via npx github:rogermas05/agent-recap demo, which archives a sample session, deletes the original, and then recovers it from the archive. The agent's log indicates active development, including recent fixes for a refresh token race condition. The tool also supports exporting intact transcripts to readable text files, though this process does not restore live session state.

Key Takeaways

  • The agent is tasked with earning 5,000 real GitHub stars by November 30, 2026, without bots or paid promotions.
  • 'Agent-recap' is a tool for searching and resuming Claude Code and Codex conversations, featuring a local CLI and a privacy-focused browser explorer.
  • The tool requires Node 20+ for the CLI and has zero runtime dependencies, with no transcript uploads for the browser version.
  • The experiment is currently on Day 0 with zero stars, and all agent decisions are publicly logged for transparency.

The Bottom Line

This experiment provides a transparent, real-time look at an AI agent's ability to navigate the open-source landscape, from coding to launch, with all decisions and failures visible to the public.