Skills, MCP servers and agent docs all take up your agent's context, in every session. Skill Tree gives each project only what it needs, in Claude Code, Codex, Pi and opencode.
Illustrative example. Every tier turns, so a long list still fits.
A skill is a folder with a SKILL.md. It lives in one place: ~/.skill-hub.
Bundles are named sets. android holds what every Android app needs.
storefront-android gets android, testing and ship-it, plus the global essentials. Usage shows what the agents actually call.
hub sync links each skill into every agent the project uses. Links, not copies.
Edit compose-screens once. Both Android apps have the new version in their next session.
--- name: compose-screens description: Build Android screens with Jetpack Compose. Keep state in the ViewModel, keep composables stateless, and give every screen a preview. --- # Steps 1. Read the feature module and find the closest screen. 2. Define a UiState data class and a sealed Action type. 3. Write the stateless composable, then the ViewModel. 4. Add previews for loading, empty, error and content.
Applied to storefront-android and courier-android.
13 skills unused this month. Each still costs context in every session.
$ hub sync Project skills: storefront-android [~/code/storefront-android] effective harnesses: Claude Code, Codex, opencode ✓ .claude/skills/compose-screens → ~/.skill-hub/skills/compose-screens ✓ .agents/skills/compose-screens → ~/.skill-hub/skills/compose-screens ✓ .claude/skills/room-database → ~/.skill-hub/skills/room-database … 39 more (21 skills × 2 folders) # Codex and opencode both read .agents/skills/ courier-android [~/code/courier-android] ✓ .claude/skills/compose-screens → ~/.skill-hub/skills/compose-screens … more ✓ sync complete
One project in the desktop app. Five answers you otherwise dig out of dotfiles and session logs.
Tokens, cost and sessions per agent, project and skill, read from your agents' own logs on your machine. See which skills earn their place in the context window, and which ones just sit there.
A skill brings the sub-agents, hooks and permission rules it needs. Equip the skill, and they come along.
Same model as skills: define once, keep secrets as ${VAR} references, equip per project. Each agent gets it in its own config.
Follow a Git repository of someone else's skills, take the ones you want, and put them in bundles and loadouts like your own. Skill Tree shows what changed upstream before you pull it in.
Every AGENTS.md and CLAUDE.md in a project, nested ones included, with the skills each mentions and what it costs in context.
Snapshots your library, sub-agents and agent docs to a private GitHub repo you own. A new machine restores from it in one command.
A Linux box follows the loadouts you pick. The Mac publishes signed revisions to a private Git feed, and a small receiver applies them, even while the Mac is offline.
Skill Tree ships its own MCP server. Ask Claude Code or Codex to equip a bundle or write a new skill, and it works on the same registry you see in the app.
hub command. Script it.› Set up payments-api like our other Go services, and add the security bundle.
⏺ skill-tree · project_list 7 projects
⏺ skill-tree · bundle_list go-service, security, testing …
⏺ skill-tree · equip bundle go-service → payments-api on
⏺ skill-tree · equip bundle security → payments-api on
payments-api now has go-service, security and testing, linked for Codex and opencode. Start a new session to load them.
Your library is plain files in ~/.skill-hub. No account, no cloud service in between.
Developers who build with coding agents across many projects, solo or in a small team. It is context engineering for your agent harnesses: you decide once what Claude Code, Codex, Pi or opencode should know in each project, and Skill Tree keeps it that way.
Yes. You keep one copy of each skill in ~/.skill-hub. Skill Tree links it into .claude/skills/ for Claude Code and into .agents/skills/ for Codex, Pi and opencode. Fix a skill once, and every agent sees the fix.
That is the whole idea. Put the skill in a bundle and apply the bundle where it belongs; every other project stays clean. Skills you want everywhere go into a global bundle.
No, it links them. Each project gets symlinks into your library, so there is nothing to keep in sync by hand and nothing drifts.
Yes. Add a Git repository as a source, choose which of its skills you want, and use them in bundles and loadouts like your own. Skill Tree checks the source for updates and shows the diff before you sync.
No. The same library holds MCP servers, with secrets kept as ${VAR} references. Skill Tree also manages AGENTS.md and CLAUDE.md, permission rules and hooks, per project and in each agent's own format.
Yes. Skill Tree ships an MCP server, so Claude Code or Codex can list your projects, equip bundles, create skills and sync on your behalf. Every action is also a hub command for scripts.
Your library lives in ~/.skill-hub. Usage numbers come from your agents' local logs, and no prompts are uploaded. Backups go to a private GitHub repo you own, and only if you turn them on.
The desktop app is built for macOS. Linux works from source on a best-effort basis. Windows is not supported.
First run reads the skills you already have and builds the registry out of them. Nothing to author before it's useful.
curl -fsSL https://skill-tree.tech/install.sh | shInstalls Skill Tree.app and the hub command on macOS. Checks the download's SHA-256; no sudo, no shell-profile edits.