mbx wraps ordinary Cargo commands with a content-addressed rustc cache. Cargo
still resolves dependencies, plans builds, and links outputs; mbx restores
supported compilations it has seen before. There is nothing to configure and
nothing to install into Cargo: put mbx in front of the command you already
run.
The first build prints what it set up — where the cache lives, the budget it
is pruned to, and when a target/ directory becomes collectable.
Why mbx?
- Warm every worktree. Cache keys contain no checkout-specific absolute paths, so building one checkout warms its siblings automatically without sharing a Cargo target lock.
- Bounded disk, without a chore. The action store prunes itself to a
budget, and managed
target/directories are deleted when their checkout is gone, unused for 30 days, or over their share of the disk. Both budgets scale with the disk rather than assuming every machine is the same size. - Warm CI safely. GitHub Actions cache can warm fork pull requests from a
cache built on
main, while a self-hosted remote can serve trusted runners and teammates. Pull requests never publish remote objects. - See the whole result. mbx reports hits, misses, actions it could not look
up, and actions it deliberately bypassed. A high hit rate cannot hide work
that never entered the cache.
mbx tuishows the same outcomes as they happen, for every build on the machine at once.
Install
With mise:
With Cargo:
Or install the latest Linux x86-64 release archive:
archive=mbx-x86_64-unknown-linux-gnu.tar.gz
release=https://github.com/jdx/mr-boxington/releases/latest/download
|
Use the corresponding -musl archive for a static Linux binary. Release
archives cover both libc variants on x86-64 and ARM64, plus Apple Silicon and
Windows x86-64 and ARM64.
Every release includes SHA256SUMS.
See all installation options →
Automatic pruning
Collection runs after a build, at most once an hour, and needs no
configuration. Both size budgets default to a share of the disk holding the
cache — 5% for the action store and 10% for managed target/ directories,
each from its floor (5 GiB and 10 GiB) up to 100 GiB and rounded down to a
whole 5 GiB — and a managed directory is also collected once its checkout is
gone or it has sat unused for 30 days.
mbx keeps a running total of what that has been worth and reports one line of it after a build:
mbx[savings]: 41.7 GiB of target/ had outlived its checkouts. it has been dealt with.
The line is drawn from a pool, so it does not repeat itself. savings = "plain" states the same facts without the joke, and savings = "off" keeps
the totals without printing anything (MBX_SAVINGS from the environment). Inspect or collect the store explicitly with:
For a checkout without an existing target/, managed target directories are
enabled automatically:
mbx places the target directory under its cache root and leaves target as a
symlink, so familiar paths still work — and the outputs of a checkout that is
deleted get collected rather than stranded.
For an existing real target/, an interactive mbx asks before removing the
old outputs and replacing the directory with a managed link. The safe default
is to keep it. Non-interactive runs never remove it. Set MBX_TARGET_VIEWS=0
to disable managed placement and the prompt.
Worktree and CI warming
An equivalent rustc action keys the same across checkout paths. One worktree's dependency build can therefore warm another while every checkout keeps its own target directory.
For GitHub-hosted CI, jdx/mr-boxington-action
can install mbx and use GitHub Actions cache, saving only from the default
branch and restoring in pull requests. Trusted environments can switch the
same action to a compatible remote such as the self-hostable
cache server.
How it works
An mbx Cargo command starts an in-process cache agent, points Cargo at a rustc shim with
RUSTC_WRAPPER, and exits the agent with the build—there is no daemon. The shim
derives an action key, restores cached outputs when possible, or runs the real
compiler and publishes a successful result.
Anything mbx cannot model exactly bypasses the cache. Native linking is not cached; built-in WebAssembly targets with self-contained linkers are the narrow exception because their linker, CRT, and libc inputs ship in the Rust toolchain. Incremental compilations bypass the cache. Correctness comes before hit rate.
Read the architecture and limits →
Documentation
- Get started
- Configuration
- GitHub Actions
- Remote cache
- Protocol compatibility
- Cache results
- Watching builds
- CLI reference
- Current limits