A recipe is data.
One YAML file per tool, holding the phases erbina runs in order: detect, install, configure, verify. Install methods are listed in preference order, each guarded by the manager it needs, so the same recipe resolves to Homebrew on a Mac, winget on Windows and cargo or pipx on a bare Linux box. The server handles recipes generically: there is no per-tool code in it, and a recipe that breaks the schema is refused at load rather than half-run.
Verify means run it.
A check that a file exists proves a file exists. erbina’s registry policy rejects a verify that only inspects the filesystem: it has to execute the tool and get a zero exit, so a truncated download or a binary that cannot find its own libraries fails the install instead of passing it. The same honesty runs backwards through the other phases. detect gates installation, so bootstrapping something twice is a no-op rather than a reinstall, and uninstall re-runs detection afterwards to confirm the tool is actually gone.
Nothing executes before you have seen it. Every recipe has a plan surface that renders the exact commands without running any of them, and the install path is reachable only after that plan could have been shown. It installs software with real privileges, so consent is a structural property, not a prompt.
And what you already have.
Claude Code resolves MCP servers across three scopes, and the usual confusion is one name defined in two of them. erbina audits all three and names the shadowed ones, health-checks the servers that no longer connect so dead entries can be removed from the right scope, and re-runs detection and verification over the tools it installed to sort them into healthy, missing and broken. Version checks are read-only and comparison is asymmetric: an unparseable version is reported as unknown rather than guessed at, and a pinned tool is never offered an update.
Apache License 2.0, with the name and logos reserved in a NOTICE file: as free to fork as before, a fork just needs its own name.