# Skill Tree

> Give each project the agent skills it needs, and only those. One local library of skills, MCP servers, and agent docs for Claude Code, Codex, Pi, and opencode. Skill Tree links each project's set into the folders where each agent looks, so every session's context holds only what that project needs.

Skill scope in AI coding agents is binary. A skill either lives at the user level and loads
into every session, or you copy it into one project's `.claude/skills/`. The first wastes
context and steers the agent with advice that does not apply. The second loses the single
source: the same skill lives in six repos, and a fix to one copy fixes only that copy.

Skill Tree adds a layer between the two: bundles. A bundle is a named set of skills and MCP
servers that you apply to projects. Each project's loadout is global skills ∪ its bundles ∪
skills added directly. `hub sync` links those skills into each agent's folder, so an edit to
the library copy reaches every project that uses it at the next session.

Version 1.0.0 (September 2026): the registry format, the `hub` CLI, and the sync contract
are stable. License: Apache-2.0. Free.

## Who it is for

Solo developers and small teams doing agentic engineering across many projects. Skill Tree
is a context-engineering and harness-engineering tool: it decides what each agent harness
(Claude Code, Codex, Pi, opencode) loads into its context in each project, from one shared
library. It suits multi-project setups where the same skills, MCP servers, and instructions
would otherwise be copied from repo to repo.

Related terms: context engineering, agentic engineering, harness engineering, agent harnesses,
coding agents, agent skills, AGENTS.md, CLAUDE.md, MCP servers, multi-project, solo developer.

## Where skills go

| Agent | Folder Skill Tree links into |
|---|---|
| Claude Code | `.claude/skills/` |
| Codex | `.agents/skills/` |
| Pi | `.agents/skills/` |
| opencode | `.agents/skills/` |

## What it does

- One library in `~/.skill-hub/`. Skills are folders with a `SKILL.md` and any reference files.
- Bundles group skills and MCP servers by stack (`android`, `web`), by kind of work
  (`planning`, `testing`), or as a workflow (`ship-it`). Global bundles reach every project.
- Per-project loadouts, with the agents each project uses. Sync writes only for those agents
  and removes only links it wrote.
- MCP servers: defined once, secrets as `${VAR}` references, written into each agent's own
  config, with a delivery check per agent.
- Companions: a skill can declare the sub-agents, hooks, and permission rules it ships with.
  Equipping the skill offers to install them.
- Skill references: Skill Tree links skills that mention each other, warns about missing
  ones, and rewrites references on rename.
- Agent docs: every `AGENTS.md` and `CLAUDE.md` in a project, the skills each mentions, and
  its context cost.
- Permissions and hooks, globally or per project, in each agent's own format.
- Usage: local token and cost estimates per agent, project, session, and skill, read from the
  agents' own logs. No raw prompts are uploaded.
- Remotes and cloud: deliver loadouts to headless machines; export skills for claude.ai and
  ChatGPT; share a skill as a `.skillpack` file.
- Skill sources: follow a Git repository of other people's skills (`hub source add git <url>`),
  keep the ones you want (`hub source edit <id> --include a,b`), use them in bundles and
  loadouts like your own, and check, diff, and sync upstream updates. A linked bundle follows
  its source until you detach it.
- Git backup (optional): snapshots the library, sub-agents, and agent docs to a private
  GitHub repo you own. A new machine restores with `hub bootstrap --restore-from <repo>`.
- Three ways in, one registry: the desktop app (Tauri + React), the `hub` CLI, and a built-in
  MCP server that lets an agent manage its own setup.
- First run imports the skills you already have. Nothing to author before it is useful.

## Install

macOS, one line. It installs `Skill Tree.app` into `/Applications` and a `hub` command into
`~/.local/bin`, checks the download's SHA-256, and uses no sudo:

```
curl -fsSL https://skill-tree.tech/install.sh | sh
```

By hand: download `SkillTree-macos.zip` from https://github.com/Ramtoi/skill-tree/releases/latest.

Linux works from source on a best-effort basis (Python 3.9+; Node 20+ and Rust for the
desktop app). Windows is not supported.

```
git clone https://github.com/Ramtoi/skill-tree.git
cd skill-tree
python3 hub.py bootstrap
```

## Common tasks

```
hub project add storefront-android ~/code/storefront-android
hub bundle apply android --project storefront-android
hub enable review-diff --project storefront-android   # one skill, without a bundle
hub sync                                          # write links for every project
```

## Links

- [Landing page](https://skill-tree.tech)
- [Source and README](https://github.com/Ramtoi/skill-tree)
- [Releases](https://github.com/Ramtoi/skill-tree/releases)
- [Changelog](https://github.com/Ramtoi/skill-tree/blob/master/CHANGELOG.md)