Murmell launches AVP, an open source Git alternative for AI coding agents

Murmell has released AVP, an open source tool that its makers position as an alternative to Git for AI coding agents, distributed as a single binary. Agents open a task, write code and land it, and each landing is checked against the exact current head of main. If main has moved in the meantime, the task is replayed on the new head and checked again.

Every landing is still a regular Git commit, so GitHub and existing deploys keep working. Running `avp up` in any Git checkout connects Claude Code and Codex. Murmell says its own team no longer uses Git directly, works through AVP, and keeps GitHub only as a fallback for storing code.

The maker reports that pull request CI runs on Murmell's repository fell from 104 a day to 4. In an internal benchmark with six to eight Claude Code agents on one repository, main turned red under Git and Jujutsu and tasks went missing, while under AVP main stayed green and no work was lost in any run so far. The method and figures are published at avp.murmell.com.

Teams running several coding agents against one repository may want to test AVP to avoid broken main branches, conflicting edits and lost tasks. The benchmark and CI numbers are self-reported by the maker, so they are worth checking on your own workload before relying on them.

Because landings remain standard Git commits, existing GitHub workflows and deployments should not need to change, but any agent tooling that calls Git directly will need to be checked against AVP's `avp up` setup.