MeshXFS: AI Turns XFS Into a Clustered File-System

A new project called MeshXFS flaunts a principle that may unsettle a cautious administrator: according to its authors, "no human wrote any of its own code." The task of turning the XFS file-system driver into a clustered solution was handed to the language model Claude Opus.

The initiative was started by Steve Shoecraft, who flagged it to Phoronix. As he tells it, MeshXFS began as an experiment around a single question: can AI write not merely a kernel file-system driver, but a clustered file-system specifically — one with distributed locking, concurrent access from many nodes, fencing of misbehaving nodes, and crash recovery. It is now a working, released, GPL-2.0 POSIX clustered file-system, and the answer so far, in his words, is yes.

What the AI actually did

The foundation is upstream XFS code, written by the XFS developers. Everything that turns XFS into a system many machines can mount read-write at once was written by AI. That means two distributed lock managers: one running over the network, the other keeping lock state on the shared disk itself via the SCSI COMPARE AND WRITE command, with no lock network at all. It also covers cache coherency across nodes, fencing through SCSI persistent reservations, survivor-led journal recovery, an on-disk envelope around XFS, and the userspace tools (mkfs, fsck, resize).

The author stresses the scale: since the fork, the xfs/ tree has grown from roughly 226,000 to about 314,000 lines. The test harnesses, the test-rig automation, and the release validation were likewise prepared by the AI.

Claude Opus ran as an autonomous coding agent inside Claude Code and did a large share of the work, consulting Claude Fable, GPT, Gemini, and Astra on the hardest spots: lock-protocol design, recovery races, and fencing correctness. The AI wrote and ran the tests itself, diagnosed the failures, and fixed them.

The human's role

Shoecraft describes his own contribution modestly: he set the goal and the bar — data integrity first, never crash the host, and stay within twice the time native XFS takes on the same storage. The code, he says, he did not write, edit, or review.

For those uneasy about trusting their data to an AI-modified file-system, the project reports that each MeshXFS release goes through a suite of roughly thirty checks on real multi-node clusters across a range of real-world scenarios. Details and sources are posted on the project's GitHub.

Source: Phoronix