Tools Matter (Manganin)

Published 25 July 2026 on blog.manganin.dev. The title suggests a general argument about tooling; the post is a design manifesto for one tool. Manganin is a git forge written in Zig, named after the copper-manganese alloy whose electrical resistivity barely moves with temperature. Nothing in the post describes shipped behaviour, benchmarks, or a timeline, so what follows is stated intent.

The vouch tree

The one idea with real mechanism behind it. A prospective contributor has to be vouched for by an existing contributor, who is promising that their vouchees will follow the project's rules. Vouching grants the ability to open issues, submit pull requests, and otherwise participate. Because every account hangs off a chain that leads back to the top of the tree, vouching someone who then griefs the project has consequences for the voucher. The stated result: how much of the core team's time you are able to waste is tied to how much you are trusted. The author's evidence is comparative rather than measured — lobste.rs, where discussion of hard topics stays broadly civil and bans are rare, against Reddit.

The resource being rationed is the same one contributor-poker is about: maintainer review time is the bottleneck in any project that succeeds, and burnout makes it scarcer every year. The two answers point in opposite directions. Loris Cro's argument for Zig is that you should spend attention on unproven contributors, because carrying someone's imperfect first patch over the line is how you grow the person who reviews for the next one. The vouch tree instead makes access conditional on someone already trusted staking their own standing. It optimizes for not wasting attention; contributor poker optimizes for compounding contributors. A forge that builds the first model into its permission system makes the second one harder to run on top of it, which is the kind of thing worth knowing before adopting one.

Clone gating and self-hosting

The same vouching check can be extended to cloning. An instance can require that you be vouched before you can pull the repository at all, which the post frames as blocking unauthorized bots and model-training scrapers without JavaScript, redirects, or slowdowns — an alternative to the proof-of-work interstitial that has become the default answer to scraper load. Manganin is built to be self-hosted, with no proprietary data formats, no vendor lock-in, minimal hardware requirements, and everything stored natively in Git as plain text files.

Why another forge

The author's answer is that the existing ones mostly do the same things. Forgejo and Codeberg were deliberately designed as a drop-in replacement for GitHub, which he grants makes sense for that project's goals, but it means the design decisions being inherited are decades old. The alternative he wants is issue tracking that is Git-native and local-first: create and edit issues offline in your own text editor, sync them like any other commit, relying on tools that already exist rather than a web form. That is the keep the work next to the code instinct carried all the way into the repository, where the issues are literally files under version control instead of rows in the forge's database.

There is a normalization cost on the other side. forge, Andrew Nesbitt's unified CLI over GitHub, GitLab, Gitea/Forgejo and Bitbucket, works precisely because those forges converged on the same object model. A forge that reworks access control and moves issues into the repository is not something a compatibility layer can paper over.

The post closes on "Together, we serve the users." Taken as a whole it is positioning rather than a report, and the two claims worth returning to are whether the vouch tree survives contact with people who want to contribute and have nobody to vouch for them, and whether clone gating turns out to be a workable scraper defence or just a wall.