GitHub - mohsen1/git-lazy-mount: Lazily mount a git repo without cloning it. Works with ordinary git commands

Lazily mount a git repo without cloning it. Files materialize as they are read or edited.

git lazy-mount https://github.com/example/huge-repo ~/huge-repo

After it returns, your ordinary git and tools just work:

cd ~/huge-repo
vim src/main.rs
git commit -am 'Some edit'
git switch -c feature
git push

Why?

This is aimed at microVMs that spin up to run coding agents against a git repository. The idea is that the coding agent can start working immediately without having to wait for a full clone.

When the agent runs a test or build, only relevant files are downloaded on demand.

`Grep` tool in AI session

Tools like rg and git grep read every file, so they pull the whole repo and undo the point of lazy-mount.

To mititgate this, we can route search through sgrep instead. It queries a code-search index (Sourcegraph by default, and pluggable) and overlays your uncommitted edits, fetching nothing.

Performance in real world

Across 20 repositories — facebook/react to the 179k-file LLVM tree — git clone --depth 1 totals 7.3 GB vs 1.3 GB of lazy mounts (5.5× less). Ready time is 268.7 s for shallow clone vs 88.7 s for lazy-mount (3.0× faster), and each lazy mount is ready in 0.8–15.1 s (measured in Firecracker microVMs):

Lazy mounts keep full history even when the clone baseline is shallow, are ready in seconds, and materialize only the files you touch. In the full Firecracker agent benchmark, lazy-mount won 19 of 20 code-search/edit/commit tasks: full clone+agent totaled 3394.2 s vs 1322.7 s for lazy mount+agent (2.57× faster), with the only loss on the tiny vue repo. Full data, time/disk charts, and the session-time breakdown: benchmarks/.

Linux Only

Linux only: because almost all microVMs are Linux-based.

The whole stack (a transparent kernel-mounted working tree) is built on Linux FUSE (libfuse3, /dev/fuse).

Windows and macOS

Windows and macOS are not supported. The design notes and feasibility studies are kept under docs/future-platforms/ if we pick them up later.

Install / build

# Linux. Needs libfuse3 + the system git (>= 2.36).
cargo build --release -p glm-cli --features fuse   # produces `git-lazy-mount`

Docs

Everything is in docs/:

Using it: compatibility (which git commands work, and how lazily) and limitations (what's deferred, and why).
How it works: the architecture overview, then deep-dives into the worktree model, FUSE semantics, and object fetching.

License

MIT + Apache