Skills Hub: AI Governance at InBot
My role
I'm AI Enablement Lead & Frontend Tech Lead at InBot. I built the hub's front end and back end and designed the publishing flow with the AI team.
An AI skill is code that runs on the machine of whoever receives it. So distributing skills across a whole company is, above all, a governance problem. The Skills Hub is the core of the AI Enablement function I founded at InBot: it is how every team receives, publishes and updates the skills they use in Claude Code and Codex.
Company login: everything goes through company SSO (Keycloak), and each person only sees and receives their own teams' skills.
Staging and production: every skill has a staging and a production version; people with access test it before it reaches everyone.
Mandatory review: no release goes straight to people's machines, not even from root. Each submission becomes a PR reviewed by the AI team.
Signed installer: the Linux, macOS and Windows installer has its signature checked before download and keeps skills up to date on its own.
Zip-based publishing: authors upload the skill folder, review a diff preview and choose the target; governance fields missing from SKILL.md are filled in automatically.
Internal InBot system: the code is proprietary and the screenshots use fictitious data.
Before
An open hub protected only by an admin password: anyone with the link could see everything, with no traceable author, review or version.
The problem
Before I took over skills governance, InBot's hub was open. Only the admin area had a password: anyone with the link could see everything. Skills were uploaded by hand as zip files, with no sync to GitHub, and machines pulled updates using a GitHub token. There was no record of who published what, and no way to guarantee that the installed file matched the reviewed one.
Since a skill runs on the recipient's machine, this was a security problem.
The solution
I rebuilt the hub on top of company login (Keycloak) with a "see = receive" rule: each person sees exactly the skills the installer puts on their machine, and only their own team's. Publishing became a flow with a diff preview, a staging (stg branch) or production (PR to main) target, and mandatory review by the AI team.
To make sure the installed file is the reviewed one, releases carry a human signature made with an SSH key. The installing machine checks that signature, and if a file isn't signed, it doesn't get installed. The front end is React + TypeScript; the backend is Convex, shared with the update CLI people use in the terminal.
Decisions and trade-offs
Corporate login (Keycloak) instead of a shared password, with a "see = receive" rule: each person sees only what the installer delivers to them. Signed releases, verified by the installer before anything is installed. Publishing split between staging (stg branch) and production (PR to main), always reviewed by the AI team.
The results
The hub became the single entry point for the company's skills, with around 50 skills and automations in production and 77 AI Enablement initiatives delivered in 4 months. Every release now has a traceable author, review and version.
It used to be just a place to upload files; now it has per-skill usage telemetry, logs and team permissions. The telemetry is already showing how people use the skills and where to focus.
Results by the numbers
- ~50Skills in production
- 3Operating systems
- Claude Code + CodexAgents supported
- 100%Releases reviewed before going live