mbx-cache-core 0.10.1

Unstable CAS, transport, and cache-agent primitives for mbx embedders
Documentation

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. Run mbx setup once and keep using the Cargo commands you already run.

cargo build                # cached by mbx
cargo test --all-features  # cached by mbx
cargo clippy --workspace   # cached by mbx
mbx tui                    # watch every build's cache activity live
mbx gc --dry-run           # preview what cleanup would reclaim

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.
  • Run multiple Cargo builds at the same time. They share one machine-wide CPU and memory budget. Identical cold compilations already running are compiled once and restored into every other build. With a compatible cache server, the same in-flight deduplication extends across CI runners.
  • 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 tui shows the same outcomes as they happen, for every build on the machine at once.

Install

With mise:

mise use --global --postinstall "mbx setup --yes" mr-boxington

[!NOTE] The --postinstall setup hook requires mise 2026.8.15 or newer.

Open a new shell and verify that plain Cargo resolves to the stable mbx shim. On Unix:

command -v cargo
# ~/.local/share/mbx/bin/cargo on Linux

On Windows, use PowerShell or where.exe:

(Get-Command cargo).Path
where.exe cargo

mise activation handles interactive shells. SSH commands, coding agents, and other non-interactive tools may not load that activation; prepend ~/.local/share/mbx/bin to PATH in a startup file they read (for example, ~/.zshenv for zsh). Afterward, agents and people can use the ordinary cargo commands they already know; explicitly prefixing a command with mbx remains supported.

With Cargo:

cargo install mbx --locked
mbx setup

Or install the latest Linux x86-64 release archive:

mkdir -p ~/.local/bin
archive=mbx-x86_64-unknown-linux-gnu.tar.gz
release=https://github.com/jdx/mr-boxington/releases/latest/download
curl -fsSLO "$release/$archive"
curl -fsSLO "$release/SHA256SUMS"
grep "  $archive$" SHA256SUMS | sha256sum --check --strict -
tar -xzf "$archive" -C ~/.local/bin

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.

After installing from an archive, run $HOME/.local/bin/mbx setup. Without mise it prints the appropriate PATH command and does not edit your shell startup files.

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:

mbx cache stats
mbx cache projects
mbx cache largest --limit 10
mbx cache verify
mbx cache remove /path/to/workspace
mbx gc
mbx gc --max-size 3GB
mbx gc --dry-run

For a checkout without an existing target/, managed target directories are enabled automatically:

cargo build

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.

Learn about managed targets →

Worktrees and parallel CI

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.

GitHub Actions' parallel steps can start independent Clippy or test configurations together. Give each one a separate CARGO_TARGET_DIR and run it through mbx: the commands share one CPU and memory budget instead of each trying to fill the runner on its own. We saw mise's lint job finish up to 45% sooner this way.

Copy the parallel workflow →

How it works

An mbx Cargo command starts an in-process cache agent, points Cargo at rustc and rustdoc shims, and exits the agent with the build — there is no daemon. The shims derive action keys, restore cached outputs when possible, or run the real tool and publish a successful result. That includes cargo doc: rendered crate pages are cached independently, then rustdoc cheaply rebuilds their shared search indexes.

Anything mbx cannot model exactly bypasses the cache. A native link is cached only when the linker itself can enter the key: host binaries and tests on Linux and macOS, where mbx resolves the linker, startup objects, and libc, and self-contained WebAssembly targets everywhere, whose linker and libc ship in the Rust toolchain. Incremental compilations bypass the cache. Correctness comes before hit rate.

Read the architecture and limits →

Documentation

Acknowledgements

mbx relies on Cargo and follows earlier compiler-cache work in sccache and kache, which directly inspired its design. Read the acknowledgements.

License

MIT