What problem it solves
Your coding agent re-derives the same solution every time it bumps into a familiar problem. You either re-explain it, or hand-maintain a prompt library that quietly rots when no one is looking.
With xskill running, that work goes away:
- Patterns that actually worked get distilled into Skill files your agent loads automatically.
- The library grows itself as you keep using your agent — no review queue, no one curating "best practices."
- When you edit a Skill by hand, xskill picks up your edit immediately and learns from it.
- A new Skill version only replaces the old one if it measurably serves users better (UX-driven evolution, not naive LLM self-grading).
Get started
pip install xskill # Python 3.9+
xskill serve # writes ~/.xskill/config.yaml, then exits
Open ~/.xskill/config.yaml and fill in two model endpoints:
skill_dir: ~/.xskill/skill
llm:
base_url: https://api.deepseek.com
model: deepseek-v4-flash
api_key: YOUR_KEY
embedding:
# DeepSeek does NOT provide embeddings. Alternatives:
# Alibaba DashScope: base_url=https://dashscope.aliyuncs.com/compatible-mode/v1 model=text-embedding-v4
# OpenAI: base_url=https://api.openai.com/v1 model=text-embedding-3-small
# Ollama (local): base_url=http://localhost:11434/v1 model=nomic-embed-text
base_url: https://dashscope.aliyuncs.com/compatible-mode/v1
model: text-embedding-v4
api_key: YOUR_KEY
dim: 0
Run xskill serve again — it auto-detects every supported agent on your machine (Claude Code, Codex, OpenCode, OpenClaw, Cursor) and starts watching. To also index an archive of older trajectories:
xskill registry add /path/to/trajectories
Rate limiting (cloud-plan users)
Default watcher.max_concurrent: 4 is conservative and OK for accounts without
a hard concurrency cap. If your provider has RPM/TPM quotas (OpenAI Tier-1,
Azure 60 RPM, OneAPI, etc.) add rate_limit under llm:
llm:
base_url: https://api.deepseek.com
model: deepseek-v4-flash
api_key: YOUR_KEY
rate_limit:
rpm: 60 # match your provider plan
tpm: 100000 # optional within rate_limit
burst: 10 # optional; default = ceil(rate/6)
Buckets are shared per base_url across utils/llm and agno_factory
paths, so the same key is never double-debited. Self-hosted vLLM / high-tier
accounts should leave rate_limit out and raise max_concurrent to 20-30.
Design rationale: docs/adr/0001-rate-limit-diy-not-litellm.md.
Team mode — the killer use case
The way xskill really wants to be deployed in an organization is team mode: one machine is the server, everyone else joins as a thin client, and the whole team works against the same evolving Skill library.
xskill serve --server # prints a join token
xskill connect <host:port> --token <token>
- Silently distill your top performers. When one person solves something in their own work, the rest of the team gets that solution automatically — nobody has to write it down. (Capability democratized.)
- Any coding workflow plugs in. Codex, Claude Code, Cursor IDE — pick whatever; everyone joins the same library, synced across tools.
- Trajectories stay private. Sessions are redacted before upload; agent privacy built in.
- A/B-driven evolution. A Skill change is measured per person before it spreads — the more people in the team, the faster and sharper the evolution.
- Experts can teach manually. When an expert edits a Skill locally, the change is pulled into the server as
user-staging/<client_id>and feeds the next round of evolution.
Architecture
How it works
A few narrow LLM agents do the work. One splits a trajectory into single-intent Atoms; one routes each Atom to a Skill; one rewrites the SKILL.md once a Skill has enough material; one A/B-tests new versions on live traffic and keeps the winner. Every Skill is its own git repository, so every change is versioned and reversible. Details: docs/agent.md.
Works with your agents
| Agent | Status | Trajectory ingest | Skill install |
|---|---|---|---|
| Claude Code | ✅ verified | auto-detects ~/.claude/projects/ |
symlink → ~/.claude/skills/<name>/ |
| Codex CLI | ✅ verified | auto-detects ~/.codex/sessions/ |
symlink → ~/.agents/skills/<name>/ |
| OpenCode | ✅ verified | SQLite ~/.local/share/opencode/opencode.db |
symlink → ~/.agents/skills/<name>/ |
| OpenClaw | 🟡 implemented, not well tested | auto-detects ~/.openclaw/agents/ |
copy → ~/.agents/skills/<name>/ |
| Cursor | 🟡 implemented, not well tested | auto-detects ~/.cursor/projects/*/agent-transcripts/ |
symlink → ~/.cursor/skills/<name>/ |
| Any other agent | manual | SDK: xskill.adapters.submit_trajectory |
copy or symlink the SKILL.md directory |
Concepts
| Term | Meaning |
|---|---|
| Trajectory | One agent run — the transcript of a session. Stored as traj_*.md. |
| Atom | The smallest single-intent slice of a trajectory. Routing happens at this level. |
| Skill | A SKILL.md plus optional scripts, in its own versioned git directory. |
| Canary | A live-traffic A/B test of the current Skill against a new candidate. |
| UX score | How well a Skill served the user on a given Atom, scored 1–10 from the interaction itself. The canary keeps whichever version scores higher. |
Roadmap
- More agent adapters — Trae, Goose, OpenHands, Aider
- More mature user profiling and recommendation
- Native MCP server interface (Skills exposed as tools)
- Web UI for browsing the library and viewing canary stats
- Skill marketplace — import / export portable bundles
- Multi-tenant libraries (per-team
skill_dir)
News
- 2026-05-23 — Officially open-source,
v0.5.0released: team mode (client-server), trajectory redaction, Python 3.9 support, nogitbinary needed at runtime. See the release notes. - 2026-05-20 — MIT-licensed open source; on PyPI:
pip install xskill. - 2026-05-12 — Claude Code, Codex, OpenCode supported; OpenClaw and Cursor connected.
License
1 maintainer
Metadata
1 maintainer
Release files for xskill 0.5.2
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | |
|---|---|---|---|
| xskill-0.5.2.tar.gz | 1.7 MB | Details |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Reset |
|---|---|---|---|---|
| xskill-0.5.2-py3-none-any.whl | Python 3 | none | any | Details |
Total release size: 2.0 MB
Release files / xskill-0.5.2.tar.gz
| Download URL | xskill-0.5.2.tar.gz |
|---|---|
| Size | 1.7 MB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
383c511101d682d8ed81281cbed149845933bf82946fffc43b20a396f2bc72c0
|
|
BLAKE2b-256 checksum How to use checksums |
60129dad7dd827e6424c2b82a7decc107370bb668342d901067920f9834c6dd6
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.2.0 CPython/3.11.13
|
Release files / xskill-0.5.2-py3-none-any.whl
| Download URL | xskill-0.5.2-py3-none-any.whl |
|---|---|
| Size | 254.7 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
14aa3ad1d08950f0ff9bfa78c3aa990eeae5e7e01a1cf8aec9f4c5d4e3ea7ae8
|
|
BLAKE2b-256 checksum How to use checksums |
7585a865139c11b3f9b12ae728a5a29762fe4e0c4982cc1b4236b185ee593f78
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.2.0 CPython/3.11.13
|
