Skills

A skill is a reusable instruction pack (a SKILL.md with optional support files) the agent loads for matching tasks — a team’s operating procedure made executable. Skills live in the org catalog and are shared to the org, groups or individual people.

Where skills come from

A Git repository (GitHub, GitLab, self-hosted)

Each skill points at a repo URL — either a whole repository or a subfolder:

https://github.com/org/skills
https://github.com/org/skills/tree/main/competitor-teardown
https://gitlab.company.com/platform/skills/tree/main/qbr-pack

Register one: Admin → Skills → Add skill → paste the repo URL, name and description. On install the repo (or subfolder) is pulled into the runner’s skills directory.

Private repos need a credential. Admin → Skills → Git credentials:

1. Host — github.com, gitlab.com, or your self-hosted domain.
2. Provider — GitHub / GitLab (defines the auth scheme).
3. Token — a PAT with read access to the repos (ghp_… for GitHub — fine-grained, Contents: Read on the skills repos; glpat-… for GitLab with read_repository).
4. Optional org skills repo URL — the repo where authored skills are pushed back (see below).

Credentials are stored encrypted and matched to skills by host: a skill on github.com uses the github.com credential. No matching credential → private skills simply stay uninstallable, public ones still work.

Authored in place

Admin → Skills → New skill opens an editor: write the SKILL.md right in the product. If an org skills repo is configured on a git credential, saving an authored skill also commits it to that repo — the UI shows a synced state, so the repo remains the source of truth your engineers can review.

Installing and sharing

• Share with — the whole org, selected groups, or individual people. From Members, expand any person’s row to see their effective skill set and to grant a skill directly.
• Install — members install a skill into their own workspace; admins can install org-wide (every runner loads it).
• Adoption — each skill shows who installed it, how many runs used it and the estimated minutes saved (from its value declaration).

How the agent uses them

At session start the runner mounts the installed skills; the agent reads the matching SKILL.md when the task fits its description. Keep descriptions specific (“when to use”) — that is the trigger.