Getting started

Install ipm, pick the repos to follow, search them and build your first plugin. It takes about ten minutes, most of it waiting for the first fetch.

  1. Requirements
  2. Install
    1. From a clone
    2. With go install
  3. Onboard
  4. Check your setup
  5. 1. Follow some repos from the marketplace
  6. 2. Search the catalog
  7. 3. Build a plugin
  8. 4. Use it in Claude Code
  9. Next

Requirements

You need For Notes
Go 1.26 or later Building ipm  
pnpm 12 or later, or Node.js ipm catalog sync, ipm pack build With only Node, ipm runs npx -y pnpm@12.10.1 for you.
Docker (or OrbStack) ipm catalog serve, ipm search Runs the local Typesense search server. Everything else works without it.
GitHub CLI gh, logged in ipm marketplace stats Optional: stars, topics and descriptions.
Claude Code claude Validating and using the plugins you build Needed for ipm install claude; optional for ipm pack build, which skips validation without it.

ipm is tested on macOS and Linux. Windows paths are defined but untested.

Install 

curl -fsSL https://raw.githubusercontent.com/thinkjones/intelligent-package-manager/main/install.sh | bash

The script runs go install for the latest release of ipm into ~/.local/bin (set IPM_BIN_DIR to choose another folder). It needs Go. If that folder isn’t on your PATH, it prints the export PATH=… line to add to your shell’s startup file; if another ipm comes first on your PATH, it says which. In a terminal it then starts ipm onboard; otherwise it prints next: ipm onboard.

For tab completion of commands and flags, install the script for your shell, for example zsh: ipm completion zsh > "${fpath[1]}/_ipm". ipm completion --help lists the shells.

From a clone

If you want to work on ipm, or track main closely:

git clone https://github.com/thinkjones/intelligent-package-manager.git
cd intelligent-package-manager
./install.sh        # builds this checkout into ~/.local/bin/ipm

From a checkout, install.sh builds cli/ instead of fetching the published module. If you plan to commit, install mise, then run mise run setup once to install the git hooks, and mise run verify before a pull request. To run your working copy without installing it, use bin/ipm <command>: it rebuilds the CLI whenever its source changes.

With go install 

go install github.com/thinkjones/intelligent-package-manager/cli/cmd/ipm@latest

This puts the latest release of ipm in $(go env GOPATH)/bin; make sure that folder is on your PATH. ipm version prints which release you have. Then run ipm onboard, or follow the manual steps below.

Onboard 

ipm onboard

ipm onboard takes you from nothing to a searchable library. It first prints the four concepts, one line each: marketplace (curated repos you can follow), catalog (the repos you follow, fetched, pinned and searchable), pack (the skills, agents and plugins you pick) and install (a pack, built and installed into Claude Code). Then it loads a curated list of repos, shows them grouped into lists, and lets you tick the ones to follow. Two lists, “AI · Agentic Engineering Methods” and “AI · Skills, Plugins & Registries”, are ticked for you on a fresh start.

Key Does
space Tick or untick a repo, or a whole list
→ / ← Expand or collapse a list
/ Filter by repo, description or topic (esc clears)
a / n Tick all or none of the current list
enter, then y Continue with the ticked repos
q Quit without adding any sources

Each repo shows its stars, language and how long ago it was last pushed (these two at 100 columns or wider), and its description; the line above the key hints shows the highlighted repo’s topics and homepage. Then ipm onboard adds the repos as sources and runs ipm catalog sync, ipm marketplace stats and ipm catalog serve with ipm catalog rebuild, printing ✓ or ✗ for each step. While it works, one line at the bottom shows a progress bar, the step, what it is doing and the time so far, for example [████████░░░░░░░░░░░░] 3/5 stats · 17/42 anthropics/skills · 0:48. Stats are skipped when gh is missing or logged out, and search when Docker isn’t running; it says so and what to run later. It ends with ✓ library ready — N sources, M artifacts and a few searches to try.

If ipm already has data, ipm onboard says what it found and asks whether to use it or back it up and start fresh. Using it keeps your sources: they show ✓ in the checklist and nothing is ticked for you. Backing up renames the home and cache folders to <folder>.backup-YYYYMMDD-HHMM and prints how to restore them. If a step fails, it prints the command to rerun; running ipm onboard again picks up where you left off.

To use your own curated list, pass --list <path or URL>; the format is in Curated lists. Every flag is in the commands reference.

Check your setup 

ipm doctor

ipm doctor prints one row each for its home and cache folders, pnpm, your sources, the catalog, Docker, the search server and claude. Rows marked ! are warnings (for example, no sources yet); ✗ is a problem to fix, and makes it exit 1.

ipm keeps its data in your user folders, never in the directory you run it from. See Files and paths to move them.

1. Follow some repos from the marketplace

ipm onboard does this step and the next for you. To do them by hand:

ipm marketplace                    # the curated repos, with ✓ on the ones you follow
ipm marketplace add obra/superpowers anthropics/skills
ipm catalog sync

ipm marketplace lists the curated repos and marks the ones you follow. ipm marketplace add records the repos; ipm catalog sync has pnpm fetch each one, pin it to a commit and build the catalog. The first sync downloads every repo; later ones only fetch what changed. ipm catalog alone prints a summary: sources followed, items by type, the last sync and the search server’s state.

2. Search the catalog 

ipm catalog serve   # start Typesense in Docker, once per boot
ipm catalog rebuild # rebuild the catalog and index it for search
ipm search tdd

ipm catalog sync and ipm catalog rebuild load the search index themselves when the server is up; with it off they print a hint to run ipm catalog serve.

Each result shows a SCORE, the artifact’s type, name, repo, stars, the plugins it belongs to and its description, with its ID on the line beneath, for example skill:obra/superpowers:skills/test-driven-development. IDs are how you refer to anything in ipm:

ipm show 'skill:obra/superpowers:skills/test-driven-development'
ipm show tree 'repo:obra/superpowers'

3. Build a plugin

In any project folder:

ipm pack new my-pack --description "My picks"
ipm pack add my-pack 'skill:obra/superpowers:skills/test-driven-development'
ipm pack build

ipm pack build writes .ipm/ in the project (ipm pack new already created ipm.toml) and ends with the command to install the result. It also writes .ipm/build.stamp, which lets ipm install skip a build that is already current.

4. Use it in Claude Code 

ipm install claude my-pack

ipm install claude builds the pack when .ipm/ is out of date, then runs claude plugin marketplace add and claude plugin install for this project (--scope user or local choose another scope; --dry-run prints the commands without running them). It records the install in ipm.toml, so a bare ipm install later reinstalls everything recorded. The skill is available as my-pack:test-driven-development.

To wire it yourself instead: claude plugin marketplace add ./.ipm, then in Claude Code /plugin install my-pack@<project folder>-ipm.

Adding a marketplace changes your Claude Code configuration. To try it without that, point CLAUDE_CONFIG_DIR at an empty folder first: CLAUDE_CONFIG_DIR=$(mktemp -d) claude plugin marketplace add ./.ipm.

Next