⚙️ shipit-skill

Take an existing developer tool from "it works" to "published + listed + marketable" — one pass.

shipit-skill is an Agent Skill that runs the full launch pipeline for AI/developer tools (MCP servers, CLIs, libraries): engineering baseline → publish → directory listings → promo material. It was distilled from shipping wheel-hub, zspace-cli and media-manager-skill — every gotcha below is one that actually bit during those releases.

Install

Two ways to use it — as a CLI tool or as an Agent Skill:

A. CLI (via pip) — for running the pipeline commands yourself:

pip install shipit-skill
shipit-skill init ./my-tool --server my-tool --pkg my-tool   # scaffold CI/Dockerfile/promo
shipit-skill check-promo --dir promo --version 0.1.1 --prs 13600=open

B. Agent Skill — copy this folder into your project (or your agent's skills dir):

cp -r shipit-skill/ ~/your-project/.opencode/skills/shipit-skill   # opencode
# cp -r shipit-skill/ ~/your-project/skills/shipit-skill            # Claude Code, Cursor, etc.

What it does

See SKILL.md for the full agent instructions and the automation boundary (what the agent runs vs. what needs a human/credential).

Example

examples/mcp-server-launch.md walks all four phases against a toy hello-mcp server — reproduce it locally to learn the pipeline before applying it to a real project.

Automate releases (GitHub Action)

shipit-skill release --execute bumps, builds, tags, pushes, opens a GitHub Release, and publishes to PyPI in one shot (token from PYPI_TOKEN env). Hook it up as a manual GitHub Actions job:

# .github/workflows/release.yml
name: Release
on:
  workflow_dispatch:
    inputs:
      how: {type: choice, options: [patch, minor, major], default: patch}
jobs:
  release:
    runs-on: ubuntu-latest
    permissions: {contents: write}
    steps:
      - uses: actions/checkout@v4
        with: {fetch-depth: 0}
      - uses: actions/setup-python@v5
        with: {python-version: "3.12"}
      - run: pip install -e ".[dev]" build twine
      - run: shipit-skill release --lang python --pkg your-tool \
             --how ${{ inputs.how }} --repo owner/your-tool --execute
        env:
          PYPI_TOKEN: ${{ secrets.PYPI_TOKEN }}

Other one-shot commands with --execute: publish (real PyPI/NPM upload), bump --commit (bump + git commit), and awesome-pr --execute (opens the PR via gh). Everything else stays print-only until you pass --execute.

Development

pip install -e ".[dev]"
pytest -q          # 80 tests, coverage ≥ 85%
ruff check .       # lint
pyright            # type check (strict)

The scars (why this exists)

• mcp>=1.0 + a new SDK → fresh installs crash. Set dependency floors.
• Python 3.9 can't install mcp>=2.0 — split the CI matrix.
• YAML block scalar + inline heredoc = broken workflow. Use a script file.
• Glama introspection needs stdin held open (sleep 1) in smoke tests.
• PyPI tokens start pypi-; UUID-style strings are wrong and 403.
• npm typosquatting blocks similar names — scoped names dodge it.
• Glama builds take minutes to hours — poll, don't panic.

License

MIT