superdesigndev/tools-registry

Treg, short for Tool Registry, is a shared vault for a team's API keys, command line tools, and reusable AI agent skills. The problem it solves is that people building with AI agents often end up pasting API keys directly into an agent's context or sharing them in chat messages, which risks leaking those keys. Treg gives a team one place to store that information, so agents and coworkers can use shared tools and secrets without ever seeing the actual keys. There are three kinds of things a team can register with Treg. An endpoint is a web API, like Stripe or Intercom, along with the credential needed to call it, requests get routed through Treg, which quietly attaches the right key before sending the request on to the real service. A CLI is a command line tool such as the GitHub or Vercel command line programs, which Treg can run on your behalf with the correct login already injected, so you never type or store the credential yourself. A skill is a bundled recipe, written as a file called SKILL.md, together with whatever secrets and tools it needs, so an entire capability can be shared and kept up to date in one place rather than copied around. Setup starts by installing a command line tool that talks to a hosted or self hosted Treg server, signing in, and then running a scan command that looks at a project's environment file and installed tools to suggest what could be shared. Everything in Treg belongs to an organization, and each person's access token identifies both who they are and which organization they belong to, with roles like owner, admin, member, and viewer controlling what they can do. Treg is aimed at teams building with AI agents who want those agents to call real services like Stripe or GitHub directly, with every call logged, instead of the agent ever holding a raw API key. The project can be used as a hosted service or self hosted, and it is written in Python, installed with pip, with a full server setup requiring extra dependencies such as a web framework and a database. The full README is longer than what was shown.