Content-addressed filesystem for agent workspaces

An instant, disposable copy of your repo for every agent.

Fork a workspace in milliseconds, mount it instantly, hydrate files lazily, and throw it away when the agent is done.

Forking the Linux kernel: git worktree 7.4s vs lunar 13ms (548x)

How it works

Three primitives, one lean stack

Content-addressed store (CAS)

A BLAKE3 content-addressed store deduplicates every blob across all workspaces so a fork copies zero bytes upfront. Data exists once; every workspace references, never copies.

Copy-on-write overlay

Forks share underlying blobs. Diverged pages materialize only when written, so a fork is ~13ms regardless of repo size. Storage grows only with your actual changes.

Lazy FUSE mount

Workspaces mount via FUSE and bytes are fetched from the CAS only on first read, so mount is instant regardless of repo size. Agents see a full directory tree immediately.

Performance

548x faster than a git worktree

Linux kernel repository: git worktree 7.4 s vs lunar fork 13 ms.

git worktree
7.4 s
lunar fork
13 ms
548xfaster fork

Agent fleet

Dozens of isolated workspaces, zero contention

Every agent gets its own copy-on-write workspace forked in milliseconds. Per-path ACLs keep each agent in its lane. When the job is done, tear it down instantly: no cleanup scripts, no lingering state.

548xfaster fork
0 bytescopied per fork
ms-scalemount time

Stay in the loop

Join the waitlist

LunarFS is in private beta. Get notified when the cloud opens.

Ship faster with disposable, instant workspaces

Read the docs to get started, or join the community to follow development and ask questions.

Discord