Why it mattersSkills copied per project drift silently, so an agent behaves differently in two repositories for reasons nobody can see in the diff.
Kitter, a manager for agent skills, has reached 185 stars on GitHub. The repository was created on 2 September and last received a push on 4 September, so the star count is three days of growth. It is Apache-2.0 licensed.
The problem it names is one that only appears after a few months of use. The README puts it plainly: "Skills are easy to install and surprisingly hard to live with. Once you work across several projects, the same Skills end up copied into different directories, drift out of sync, and become difficult to update with confidence."
One library, managed links
Kitter's answer is to give every skill "one maintained home" and link it only into the projects that need it. Skills live in a single library directory on the machine, at ~/Library/Application Support/Kitter/skills on macOS, %LOCALAPPDATA%\Kitter\skills on Windows, and under $XDG_DATA_HOME on Linux. Projects receive managed links into .agents/skills or an agent-specific directory, rather than independent copies.
That is the whole idea, and the value is in what it removes. Fix a skill once and every project that links it has the fix. There is no sweep across repositories afterwards, and no question about which copy is current, because there is only one copy.
Skills can be imported from a local folder, a GitHub repository, a skills.sh-compatible source, or a Claude plugin source. The desktop app and the command line tool share the same local-first Rust core. Local-first here means what the README says it means: no account, no server, and no background indexing process.
Why drift is the expensive part
The copy-per-project habit fails quietly, which is what makes it worth a tool. Two repositories hold the same skill at two different versions, an agent produces different work in each, and nothing in either diff explains why. The cause sits in a file nobody opened. Teams tend to find it late, after they have spent time treating the symptom as model behaviour.
Linking turns that into a visible, single-file change. It also makes the per-project selection deliberate: a project gets the skills someone chose to link, rather than whatever accumulated in the directory over six months.
What it costs to try
The desktop app is macOS only today, with Windows and Linux listed as coming soon. The standalone command line tool is available on all platforms, so a Linux or Windows user is not shut out, just working without the graphical part. The app is unsigned, which on macOS means the usual manual approval step on first launch. That is a real friction point for anyone whose machine is managed by a security policy, and worth checking before you plan around it.
At three days old this is early software, and 185 stars is interest rather than a track record. The idea it is built on, though, costs nothing to evaluate against your own setup: count how many copies of your most-used skill are on your disk right now, and how many of them match.
Source
Source: GitHub
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually shipping with them. Short, and only when there is something worth reading.