Launched this week
Skillbox
One library for your AI skills, synced to every agent
279 followers
One library for your AI skills, synced to every agent
279 followers
Skillbox keeps your AI skills in one library and syncs them to Claude Code, Cursor and Codex. Follow GitHub sources, review diffs, and scan each version for prompt injection. Trim shortens noisy command output so sessions last longer.





Free Options
Launch Team / Built With

AWS Activate CreditsCredits that fuel your startup’s next milestone
Promoted

Hi Product Hunt, I'm Adenekan Wonderful, the maker of Skillbox.
I built this because my AI skills were everywhere. A few lived in GitHub repos, some sat inside project folders, and some were installed for Claude Code but missing from Cursor. A better version of one would turn up and I'd have no idea. Copying files was easy. Knowing what I had, where it came from, whether it had changed, and whether it was safe was the hard part.
So Skillbox does four things:
One library, and a smaller active set. Keep everything you've collected, and deliver only what each tool or project needs. Every skill your agent loads costs context, so we don't treat "saved" as "installed".
Sources stay attached. Skills added from GitHub or GitLab keep following their source. When one changes, you see the diff and decide whether to take it.
Trust, per version. Each version is scanned for prompt injection, data exfiltration, command injection, hardcoded secrets and obfuscated code, including by Cisco's Skill Scanner. Approval is tied to the exact content that was reviewed, so it doesn't carry over silently when a skill changes. A passing check is evidence, not a guarantee.
Trim. It shortens noisy test, git and build output before your agent reads it. The full output stays one command away, and your sessions last longer.
It syncs across Claude Code, Cursor and Codex. Trim currently covers more than that, but session insights are deepest on Claude Code.
There's a permanent free plan for the library, sync, updates and Trust. Trim and AI features are on Individual, with a 7-day trial.
I'd love to hear how your skills are scattered today, and what would make you trust a skill you didn't write. I'm here all day and will reply to everything.
Congrats on the launch!!🚀🚀
Our skills live in the repo, so they follow the branch. I run a few worktrees at once and the rules arent the same in each one.
If the library is the source of truth, does every checkout get the same Ver?👀
@seenew Thanks! 🙏 Short answer: yes, if a skill is managed by the library, every checkout that syncs it gets the same version. That’s the point of it, but it’s the wrong fit for rules that are meant to differ by branch.
For skills that should follow the branch, keep them committed in the repo. Skillbox only writes skills it manages, and it doesn’t touch the ones you leave alone. A good split is shared skills in the library and branch-specific rules in git.
Worktrees are the part I’d test. Skillbox links projects by folder path, and a worktree is a separate folder. It isn’t branch-aware, so each worktree you want synced would need linking on its own. Tell me how you set yours up
Does Trim ever cut the one line of a failing test you needed?
@leoheo Trim can drop lines, since that’s how it saves tokens but a failing command is the case we protect.
When a command fails, the full original output is saved locally. A hint in the trimmed output points to it, and the agent can pull the complete output back with one command.
Anything you don’t want trimmed can be switched off per command, and if a filter errors it falls back to the raw output instead of guessing.
@adenekan_wonderful pulling the full output back with one command works for me
Really nice one! Will definitely try it out. qq: Everything runs on user's computer or anything leaves computer?
@bhrugu
Thanks! 🙏 Not everything stays local, and here’s what does and doesn’t leave your machine.
Leaves your machine:
Your skill library lives in the cloud. The sync agent pulls skills from it and writes them into each tool’s folders. A skill you import or save from a chat is uploaded to your library, and import is accept-or-skip for each skill.
Basic device and sync status, and skill usage counts: skill name, content hash, project, which tool, and a timestamp.
Session-efficiency numbers, if you turn that feature on. These are consolidated numbers only.
Doesn’t leave:
Prompts, code, chat transcripts, tool output and command text. Usage logging and the Learn feature don’t read or send any of it. They record counts only, and there are tests for that.
Everything is open to inspection in the agent. Happy to point you at the exact fields.
The scattered-skills problem is real for me — mine live across repos, project folders, and random local copies, and I never quite know which version Claude Code is actually using. What I like most is that Trust is checked per version, not just per skill, and updates follow the source on GitHub. That's the part that would get me to centralize. One thing I'm wondering about: most of my useful skills are ones I wrote myself locally, not ones I pulled from GitHub — do those go through the same library and sync to Cursor and Codex, or is the flow mainly built around following external sources?
@jcbuilds Yes, local skills are first-class. Import them from disk (you accept or skip each one) or save them from a chat session. They sync to Cursor, Codex and Claude Code like anything else, and you can scope them per project. They show as “Yours”. Since they have no upstream, there’s nothing to follow on GitHub. Updates from the source only apply to skills you pulled in.
scanning every skill version for prompt injection is a really thoughtful touch. congrats on the launch!
@mohitvishwakarma Thank you! 🙏 One honest note so nobody over-trusts it. Every skill version is scanned by its exact content hash, and delivery to your machine waits for the checks to pass.
It does include prompt-injection checks, like phrases that try to override the agent’s instructions or hide things from you. But those checks are pattern matching, so they catch the obvious attempts and not clever ones.
The same pass also flags secret exfiltration and risky code. A second engine from Cisco reads the code structurally, and flagged skills go to an admin for review. Treat it as a strong first filter, not a guarantee.
I keep a few dozen skills across my coding agents, many pulled from GitHub repos, and keeping them in sync is exactly the mess you describe. The prompt-injection scan on each version is the part I care about most: a skill update is instructions that run with my permissions.
Does the scan run locally, and can I pin a source to a version I've reviewed, so new versions wait for my approval?
@georgeturbo Great question. The scan runs in our cloud, not on your machine, on the exact version of each skill before it can reach your agents. It checks for prompt injection, exfiltration, secrets and obfuscated code, and it's one layer of evidence, not a guarantee.
You do get approval control: every new version is scanned separately, and anything flagged is held back from syncing. Turn off auto-update on a skill and updates wait for your OK before they apply. What we don't have yet is pinning a whole source to a reviewed tag or commit, because pinning is per skill today.