An AI agent is trying to win 5,000 GitHub stars on its own, and starts at zero

The experiment in one paragraph

As rogermas05.github.io reported on its own project page, an autonomous AI agent has been given a repository, its owner's GitHub account and a single target: 5,000 real GitHub stars by November 30, 2026. A star is the platform's version of a bookmark or a thumbs-up. People click it to show interest in a project.

The agent decides what to build, writes the code and runs the launch. The page says there are no purchased stars and no bots, and that it will show the real count next to the agent's own log, failures included. When the page was captured it was labeled Day 0, and the count stood at zero.

What the agent built first

The first product is a tool called agent-recap, based on the commands shown on the page. It is meant for people who use AI coding assistants and want to find a fix they already worked out in an earlier conversation. It searches across past Claude Code and Codex conversations and shows the text around each match. For supported local sessions, it also gives a command to resume the session.

It comes in two forms. A command-line tool (a program you run by typing commands) copies your sessions into a local archive. You can then search those copies even after the original files are gone. A browser version opens transcript files in a tab.

The page says the browser explorer runs locally. It states that it does no transcript uploads, no analytics and no browser storage. For anyone wary of pasting private coding conversations into a website, that is the claim to check first.

A demo that avoids your real files

The page offers a recovery demo that does not touch real data. It archives a synthetic (made-up) transcript, deletes the original and then searches the archive. The page shows the result as one recovered session, for a sample about fixing a refresh token race condition.

There is also an export command for a transcript file that the original app no longer lists. It writes readable text to a new file, and the page offers a flag for Codex files. It warns that this does not restore live session state, so you get the words back but not a running session.

• Archive: copy sessions to a local folder.
• Search: find words across the copies.
• Export: turn an intact transcript file into plain text.

Why zero is the interesting number

Writing a working tool is the part agents are increasingly good at. Getting strangers to notice it, try it and care enough to click a star is a different job. It involves explaining the project, choosing where to share it and answering people. The experiment puts that gap in public, with a number anyone can check.

The rules also make the test cleaner. By ruling out bought stars and bots, the page says the only route is real interest from real people. Whether the agent can find that interest, and how, is still open. The page we have describes the goal and the first tool, not the method or the outcome.

It is also a small lesson in reading star counts. Attention can be engineered, so a high number does not prove a project is well made. Maintenance and open issues tell you more about whether a project is looked after.

What to take from it

This is a modest story. It is a single experiment on its owner's own page, and the result is not known. The page also does not say how much a human is involved beyond lending the account, so it would be wrong to read it as a verdict on what agents can do alone.

It is still a useful thing to watch. If you ever launch a project yourself, the agent's log may show which steps matter after the code works. And if you use AI coding tools every day, a local archive of your own conversations is a handy idea whether or not this particular tool takes off.