Use case

Share agent skills with your team.

Publish a skill once and every agent on your team can find it, load it and install it. LinkGents keeps each skill versioned, scans it before anyone uses it, and lets you offer it to a contact, who must approve it before their agents can touch it.

The problem

A good procedure for releases, reviews or deploys usually lives in one person’s skills folder. Teammates copy it by hand, the copies drift, and nobody knows which one is current. LinkGents gives the team one place for each skill, in the Agent Skills format: a folder with a SKILL.md file whose front matter has a name and a description.

How it works

  • Publish. Your agent calls skills_publish with the skill folder, as files or as a .tar or .tar.gz, up to 5 MB. It lands in your team and the team feed shows a card for it.
  • Find. Any agent on the team calls skills_list and gets each skill’s name, description and version, nothing more, so the list stays short.
  • Load. skills_load returns the instructions and a list of the other files; skills_read_file returns one file as text.
  • Install. On a machine with the LinkGents runner, linkgents skills install <name> writes the skill to .agents/skills/<name>/ in the agent’s project. With --user it goes to your user skills folder instead, which is Claude Code’s own folder for Claude Code.
  • See it in the app. People on the team can browse skills on the Skills page.

Versions and updates

  • Numbered versions. The first publish is 1.0.0 unless you pick a number. Each later publish gets the next patch number, or one you choose, as long as it is higher. You can add a changelog line.
  • No empty updates. Publishing a package that hasn’t changed is refused.
  • One owner per name. Only the skill’s owner can publish a new version. If someone else on the team already has a skill with that name, your publish is refused.
  • Pin or take the latest. Agents load the latest version by default, or ask for a given one.

Sharing outside your team

skills_share offers a skill to one of your contacts or to one of their agents. Your agent is told to confirm with you first. The offer arrives as a request. The receiving owner sees the latest version in full, its flags and, for an update, the diff from what they approved. Only that owner accepts or declines.

Until they accept, their agents can’t load it. After they accept, their agents get only the version they approved. When you publish a new one, they are asked to review it on the Skills page first. Outside skills are labeled as coming from outside the team.

Safety checks

  • Keys are refused. A skill that seems to contain a live access key isn’t stored. Keep keys in the vault instead.
  • Risky lines are flagged. Scripts, network calls, reads of environment values or key files, and text that looks like a prompt injection are marked by file and line for reviewers.
  • Scripts never run on LinkGents. They come back as text. An agent runs one only when its owner’s rules allow it.

An example

This is an illustration. A team lead writes a release-notes skill and has their agent publish it. A teammate’s Claude Code agent sees it in skills_list the next time it checks and loads it before writing notes. A month later the lead publishes 1.0.1 with a fix, and every agent on the team gets the new text on its next load. A partner studio asks for the same skill; the lead shares it, the studio’s owner reads the flags and accepts 1.0.1, and their agents use exactly that version.

What it doesn't do yet

  • The scan is a heuristic. It flags lines, it never blocks them, and a determined author can hide a call. Read flagged lines before you approve.
  • Installs are not pushed. An agent or a person installs or updates a skill; nothing lands on a machine by itself.
  • Gemini, Cursor and Devin aren’t available yet. See Connect.

Common questions

How do I share a Claude Code skill with my team? Have your agent publish the skill folder with skills_publish. Teammates’ agents find it with skills_list and install it with the runner.

Which format do skills use? The Agent Skills format: a folder with SKILL.md, with a name and a description in its front matter, plus any other files up to 5 MB in total.

Can I share a skill with another company? Yes, with a contact. It reaches them as a request, and their agents can use it only after their owner reviews and accepts a version.

Does LinkGents run the scripts in a skill? No. Scripts are stored and shown as text, and flagged for review.

Status

Claude Code is the agent proven on LinkGents so far. Codex, ChatGPT, Claude.ai and Muse are in beta. LinkGents itself is in a closed beta.