Developer

How Lyzr Built GitAgent to Turn a Git Repo into a Portable AI Agent

Built by Lyzr · · Source: Product Hunt

Quick answer

Lyzr built GitAgent (now also listed as OpenGAP), an open standard that keeps an agent's config, logic, tools and memory in a git repo. According to the team, you define the agent once and run it on Claude, OpenAI, CrewAI or OpenClaw.

Result: Ranked #5 of the day on Product Hunt for March 20, 2026, with 226 points

What they built

Lyzr launched GitAgent on Product Hunt. The listing now also calls it OpenGAP by Lyzr, with the tagline "Your repository becomes your agent". It is an open standard that pulls an AI agent's config, logic, tools and memory into a portable, version controlled definition that lives in a git repo.

Maker Shreyas Kapale says the team built it because of three problems they kept hitting: agents had no single source of truth, portability between frameworks was incomplete, and there was no real versioning. Their answer was to keep agents in git and apply the habits developers already have, like forking, branching, pull requests and tagging.

How it works

According to the listing, you define the agent once and run it on several runtimes, including Claude, OpenAI, CrewAI and OpenClaw. Because the definition sits in git, you can roll back prompts, branch, review and reproduce an agent.

Kapale shared the command to run an agent straight from a repo: npx @open-gitagent/gitagent@latest run -r https://github.com/your/agent -a claude. They say you switch runtimes by changing the last flag to -a openai, -a crewai or -a openclaw.

A few details came out in the comments:

  • Kapale describes GitAgent as a standard and a converter. Choosing which context the agent sees is left to the runtime you pick.
  • Repos have to follow the GitAgent format. Kapale says an architect agent in Claude Code can help convert existing agents.
  • Model providers can be overridden at the CLI or kept on separate git branches.
  • It works with self-hosted GitLab or Bitbucket, according to Kapale.
  • Right now it suits single agent setups. Monorepo support is on the roadmap.

The source does not go deeper on setup than that. One commenter asked how it handles drift between the agent's understanding of the code and its actual behavior over time, and that question has no answer in the thread.

The result

The team does not share usage numbers or customer results. The Product Hunt page shows the launch ranked #5 of the day for March 20, 2026, with 226 points. It has no reviews yet.

How to build something similar

This is our general advice, not something from Lyzr.

  • Put the whole agent in one repo. Keep the system prompt, tool definitions and any memory files as plain text in git so every change shows up as a diff.
  • Treat prompt changes like code changes. Work on a branch, open a pull request, and tag the versions that work so you can roll back.
  • Keep runtime specifics separate. Store the agent's instructions apart from anything tied to one framework, so moving to another runtime means writing a small adapter and leaving the rest alone.
  • Test on a second runtime early. Run the same definition in two tools before you rely on it, since each one handles context and tools a little differently.

The original post

Shreyas Kapale (Lyzr) · View on Product Hunt

Want to build something like this?

Join the Shipping Skool community free. Paid members get live bootcamp calls every week.

Join Shipping Skool — Free

More use cases

← All AI agent use cases