open-source · MIT · v0.6.2 on PyPI

Trajectories in.
Reusable skills out.

One solves it. Everyone gets it.

Your coding agent re-derives the same solution every time it hits a familiar problem. xskill distills the patterns that actually worked into Skill files your agent loads automatically.

🏖️ Get started 🌐 Join public hub ★ GitHub

MIT-licensed · Python 3.9+ · no model weights touched.

💎 SKILL.md canary ✓
v3·UX 8.7/10·winning

example — a Skill is a versioned git dir, kept by its UX score

greetings from  ~/.xskill/skill 🌅
🏷️v0.6.2current release
🌴0agents supported
🐍3.9+Python
🌊MITopen source
latest release

v0.6.2 adds identity-aware recommendations

The current README highlights user profiling, SkillHub retrieval, UX score APIs, and Windows scheduled-task persistence.

ID

Stable user identity

80

Hybrid recommendations

UX

UX score queries

WIN

Windows daemon

how it works

A few narrow agents, one self-growing library

Every Skill is its own git repository — every change versioned and reversible.

  1. 👀

    1 Watch

  2. 🔪🥥

    2 Split into Atoms

  3. 🧭

    3 Route to a Skill

  4. ✍️

    4 Rewrite SKILL.md

  5. 🐤

    5 Canary A/B + UX score

works with your agents

Plug into the agent you already use

Trajectory ingest and Skill install are wired up per ecosystem.

Agent Status Trajectory ingest Skill install
🌴 Claude Codeverified~/.claude/projects/~/.claude/skills/<name>/
🛶 Codex CLIverified~/.codex/sessions/~/.agents/skills/<name>/
🐙 OpenCodeverifiedopencode.db (SQLite)~/.agents/skills/<name>/
🐚 OpenClawimplemented~/.openclaw/agents/~/.agents/skills/<name>/
💠 Cursorimplemented~/.cursor/.../agent-transcripts/~/.cursor/skills/<name>/
🚄 Traeimplementedstate.vscdb / trajectory_*.json~/.trae-cn/skills/, ~/.trae/skills/
Any other agentmanualxskill.adapters.submit_trajectorycopy / symlink SKILL.md

Status as of v0.6.2 · 🟡 implemented, with Claude Code / Codex / OpenCode verified.

vocabulary

Five words and you've got it

🧭

Trajectory

⚛️

Atom

💎

Skill

🐤

Canary

⭐

UX score

the killer use case

Team mode: one library, the whole team

One machine is the server; everyone else joins as a thin client.

team
# Start a server, then join from any machine:
$ xskill serve --server
$ xskill connect <host:port> --token <token> --name <user-id>
🪄

Silently distill your top performers

🔌

Any workflow plugs in

🔒

Trajectories stay private

🧪

A/B-driven evolution

🧑‍🏫

Experts can teach by hand

公网演示实例

一键粘贴，接入我们的公网实例

不用自己搭服务端。装好 xskill，粘贴下面这条连接命令，再填上你自己的用户名即可。

🌐

地址

https://hub.xskill.wiki

🔑

连接令牌

dd7f641c16ced6d1db43e754055fd2c8

📊
公网实例 
# 1）安装　2）把 YOUR_NAME 换成你的用户名
$ pip install -U xskill
$ xskill connect https://hub.xskill.wiki --token dd7f641c16ced6d1db43e754055fd2c8 --name YOUR_NAME
why xskill

Stop hand-maintaining a prompt library that rots

🥥

Distilled automatically

🌱

Grows itself

✋

Learns from your edits

⭐

UX-driven evolution

🌅

Distill your first Skill in two lines.

~/.xskill — bash
# install — Python 3.9+
$ pip install xskill
# writes ~/.xskill/config.yaml, then exits — add your model keys
$ xskill serve
# run again: auto-detects your agents and starts watching
$ xskill serve
Read the quickstart ★ Star on GitHub