Short answer
To share Claude skills with a team, commit them to a repo's .claude/skills folder, ship them in a plugin, or have an Owner provision them on a Claude Team or Enterprise plan. Those reach Claude only. Baalda Agent Skills publishes one versioned skill and installs it on every teammate's machine for Claude Code, Codex, Cursor and Gemini CLI.
To share Claude skills with your team, you need one place that holds the skill and a way to get the same version into everyone's skills folder. Claude Code gives you three: commit the skill to a repo, wrap it in a plugin, or have an organisation Owner provision it on a Claude Team or Enterprise plan. Baalda adds a fourth that also reaches Codex, Cursor and Gemini CLI.
A skill is a folder with a SKILL.md in it. Copying a folder is easy. Keeping ten copies of it the same, on ten laptops, across three AI tools, after the fourth edit, is where teams lose the thread. Somebody fixes the skill on their own machine, nobody else gets the fix, and a month later the team is running four versions of the same instructions without knowing it.
Where do Claude skills live on your machine?
Claude Code reads skills from two places on disk. Personal skills sit in ~/.claude/skills/<skill-name>/SKILL.md and load in every project on that machine. Project skills sit in .claude/skills/<skill-name>/SKILL.md inside a repository, and the Claude Code docs say to "Commit it so your team gets it too".
Codex reads a different folder. Its skills page lists $HOME/.agents/skills for a user's own skills and .agents/skills inside a repository, scanned from the working directory up to the repo root. Both tools load a skill the same way: the name and description first, then the full SKILL.md only when the model decides to use it.
| Tool | Personal skills | In a repository |
|---|---|---|
| Claude Code | ~/.claude/skills/<name>/ | .claude/skills/<name>/ |
| Codex | ~/.agents/skills/<name>/ | .agents/skills/<name>/ |
So a team that uses both Claude Code and Codex already has two folders to keep in step on every machine, before anyone has written a second version.
How do you share Claude skills with your team today?
There are three routes inside Anthropic's own tools, and each one fits a different team.
- Commit it to the repo. Put the skill in
.claude/skills/and everyone who clones the repository gets it. This is the right answer when the skill belongs to one codebase. It does nothing for the skill your sales lead wrote, which has no repo. - Ship it in a plugin. A plugin can carry a
skills/directory and be installed from a marketplace. Good for a set of skills you maintain like software, with someone owning the releases. - Provision it on a Team or Enterprise plan. As of 28 October 2026, Anthropic's help centre says "Owners of Team and Enterprise organizations can provision skills for all users", and that those skills "also load in Claude Code in your terminal" when people sign in with the same Claude account on Claude Code v2.1.273 or later. It also says skills do not sync when Claude Code runs on an API key or through Amazon Bedrock.
The third route is the closest thing to a team skills library, and if your whole team is on one Claude plan and only uses Claude, it is a good one. It stops at Claude. A provisioned skill does not reach the Codex user on your team, or Cursor, or Gemini CLI.
How does Baalda share agent skills with a team?
Baalda is the team second brain we build, and as of v0.1.89 (released 27 October 2026) every vault has an Agent Skills library. You add a skill once, and Baalda installs it on every teammate's machine, in the folders their AI tools already read.
Here is what happens, in order:
- An owner or admin opens Agent Skills in the vault, creates a skill or uploads a
.zip, and edits it in the built-in editor: a file tree on the left,SKILL.mdpinned at the top. - They press Publish. That freezes the draft into a numbered version (1, 2, 3) with release notes. Only published versions go anywhere. Half-finished drafts stay on the server.
- Each skill is set to Required, which installs it for every member, or Optional, which each person switches on for their own device.
- The Baalda desktop app on each machine writes the skill into
~/.claude/skills/<name>for Claude Code, and into~/.agents/skills/<name>, the folder Codex reads, which Baalda also uses for Cursor, Gemini CLI, Copilot and OpenCode. - When an admin publishes version 2, every machine picks it up within seconds of the change, and on a two minute check otherwise.
Two details matter more than they look. First, Baalda only ever touches a skill folder it installed. Each one carries a small .baalda-managed.json marker, and a folder without it, like a skill you wrote yourself, is never written to or removed. Second, if a teammate edits their installed copy, the next sync does not overwrite it. Baalda marks it as edited and lets them keep theirs or take the team version, and taking the team version backs up the edited copy first.
An admin can also pin a version. If version 4 of a skill breaks something, pin version 3 and every machine moves back to it, without anyone deleting files by hand.
Can you import skills from GitHub?
Yes. Agent Skills has a marketplace browser with three catalogues built in: anthropics/skills, openai/skills and BenAIOS/skills. You can add any other GitHub repository by pasting owner/repo or its URL, and Baalda reads Claude plugin marketplaces, Codex plugins, or simply every folder with a SKILL.md in it.
An imported skill remembers where it came from. When the upstream repository changes, the skill shows Update available, and pulling the update puts the new files into the draft for an admin to review before they publish it. Nothing from GitHub reaches your team's machines without someone pressing Publish.
Can an AI agent read the team's skills?
Yes, read only. Baalda's MCP server has three skill tools: list_skills, get_skill and read_skill_file. A cloud agent connected to the vault can find out which skills the team has and read their files, which is useful for an agent that is not running on a machine with the desktop app. Agents only ever see published versions, never drafts. The MCP setup guide covers connecting a client.
Which way should your team share skills?
| Your situation | The honest pick |
|---|---|
| The skill belongs to one codebase | Commit it to .claude/skills/ in that repo |
| You maintain a set of skills like a product | A plugin with its own marketplace |
| Everyone is on one Claude Team or Enterprise plan, Claude only | Owner provisioning on claude.ai |
| The team mixes Claude Code, Codex, Cursor or Gemini CLI | Baalda Agent Skills |
| You want versions, rollback and a review step before a change goes out | Baalda Agent Skills |
What Baalda Agent Skills does not do
It does not publish to claude.ai. A skill in your Baalda vault reaches Claude Code and the other tools on each person's machine, not the Claude web app or Claude Cowork's cloud sessions. If your team mostly works in the browser, owner provisioning on claude.ai is still the route.
Each machine needs the Baalda desktop app installed and signed in to the vault, because the app is what writes the folders. Skills live with a synced vault, so a purely local vault has no team library. Editing is for owners and admins. Members can read every skill and switch Optional ones on, but they cannot change the team copy.
There are limits, set out in the app: 500 skills per vault, 5 MB per file and 25 MB per skill. The feature is free on every plan, and the app is open source under Apache-2.0, so you can read exactly what it writes to your disk.
If you are setting up a shared place for the rest of your team's knowledge too, the team second brain guide covers the notes side, and Claude Code as a second brain for a team covers what happens when several agents write to it at once.