AIPM — AI Plugin Manager

A production-grade package manager for AI plugin primitives (skills, agents, MCP servers, hooks). Think npm/Cargo, but purpose-built for the AI plugin ecosystem.

AIPM ships as two Rust binaries with zero runtime dependencies:

Binary	Role	Commands
aipm	Consumer CLI	init, migrate
aipm-pack	Author CLI	init

Both work across .NET, Python, Node.js, and Rust projects with no runtime dependency.

Install

Shell (Linux / macOS)

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/thelarkinn/aipm/releases/download/latest/aipm-installer.sh | sh

PowerShell (Windows)

powershell -ExecutionPolicy Bypass -c "irm https://github.com/thelarkinn/aipm/releases/download/latest/aipm-installer.ps1 | iex"

Installers are provided by cargo-dist. Run aipm-update to self-update.

Build from Source

cargo build --workspace          # build all crates
cargo test --workspace           # run all tests

---

aipm — Consumer CLI

Initializes workspaces and migrates existing AI tool configurations into portable marketplace plugins.

aipm init

Scaffold a workspace with a .ai/ local marketplace and tool-specific settings.

aipm init [OPTIONS] [DIR]

Flag	Description
-y, --yes	Skip interactive prompts, use defaults
--workspace	Generate a workspace-level aipm.toml with [workspace] section
--marketplace	Create .ai/ marketplace directory with tool settings
--no-starter	Skip the starter plugin (bare .ai/ directory only)
--manifest	Generate aipm.toml manifests for each plugin (opt-in)
--name <NAME>	Custom marketplace name (default: local-repo-plugins)

When run on a TTY without --yes, launches an interactive wizard.

What it creates:

• .ai/<marketplace-name>/ — local marketplace directory
• .ai/<marketplace-name>/starter-aipm-plugin/ — starter skill plugin (unless --no-starter)
• .ai/.claude/settings.json — Claude Code marketplace registration
• .ai/.copilot/ — Copilot agent settings (if detected)
• aipm.toml — workspace manifest (with --workspace)

aipm migrate

Migrate existing .claude/ configurations into marketplace plugins. Detects skills, agents, MCP servers, hooks, commands, and output styles.

aipm migrate [OPTIONS] [DIR]

Flag	Description
--dry-run	Preview migration without writing files (generates report)
--source <SRC>	Source folder to scan (e.g., .claude). Omit to discover recursively
--max-depth <N>	Maximum depth for recursive discovery
--manifest	Generate aipm.toml manifests for migrated plugins

Detected artifact types: skills (SKILL.md), agents (*.md in agents/), MCP servers (.mcp.json), hooks (hooks.json), commands (commands/*.md), output styles.

---

aipm-pack — Author CLI

Scaffolds new plugin packages for publishing.

aipm-pack init

Create a new AI plugin package with manifest and conventional directory layout.

aipm-pack init [OPTIONS] [DIR]

Flag	Description
-y, --yes	Skip interactive prompts
--name <NAME>	Package name (defaults to directory name)
--type <TYPE>	Plugin type: skill, agent, mcp, hook, lsp, composite

Generates an aipm.toml manifest and type-appropriate directory structure.

---

libaipm — Core Library

Shared library powering both CLIs. All logic lives here; the binaries are thin wrappers.

Modules

Module	Purpose
manifest	Parse, validate, and load aipm.toml manifests
manifest::types	Schema types: Manifest, Package, Workspace, Components, Environment, DependencySpec
manifest::validate	Name format, semver, dependency version, component path validation
init	Plugin package scaffolding (aipm-pack init)
workspace_init	Workspace + .ai/ marketplace scaffolding (aipm init)
workspace_init::adaptors	Tool-specific config writers (Claude Code, Copilot, Cursor)
migrate	Tool config migration with recursive discovery, dry-run, and all artifact types
fs	Trait-based filesystem abstraction (Real + test mocking)
version	Crate version constant

Manifest Format (aipm.toml)

[package]
name = "@company/ci-tools"
version = "1.2.3"
description = "CI automation skills"
type = "composite"
files = ["skills/", "hooks/", "README.md"]

[dependencies]
shared-lint = "^1.0"
core-hooks = { workspace = "^" }
heavy-analyzer = { version = "^1.0", optional = true }

[features]
default = ["basic"]
basic = []
deep = ["dep:heavy-analyzer"]

[components]
skills = ["skills/lint/SKILL.md"]
agents = ["agents/reviewer.md"]
hooks = ["hooks/hooks.json"]
mcp_servers = [".mcp.json"]
lsp_servers = [".lsp.json"]
scripts = ["scripts/format-code.sh"]
output_styles = ["styles/custom.css"]
settings = ["settings.json"]

[environment]
requires = ["git", "docker"]
aipm = ">=0.5.0"
platforms = ["linux-x64", "macos-arm64", "windows-x64"]
strict = true

[environment.runtime]
node = ">=18.0.0"

[install]
allowed_build_scripts = ["native-tool"]

Plugin types: skill · agent · mcp · hook · lsp · composite

Workspace Root Manifest

[workspace]
members = ["plugins/*"]
plugins_dir = "plugins"

[workspace.dependencies]
common-skill = "^2.0"

[overrides]
"vulnerable-lib" = "^2.0.0"

[catalog]
lint-skill = "^1.5.0"

[catalogs.stable]
framework = "^1.0.0"

[catalogs.next]
framework = "^2.0.0-beta.1"

---

Project Structure

crates/
  aipm/         Consumer CLI binary (init, migrate)
  aipm-pack/    Author CLI binary (init)
  libaipm/      Core library (manifest, validation, migration, scaffolding)
specs/          Technical design documents
tests/features/ Cucumber BDD feature files (220+ scenarios)
research/       Competitive analysis and design research

---

Roadmap

The following features are defined as BDD scenarios and tracked as open issues. They represent the full planned scope beyond what is currently implemented.

Dependencies

• Resolution — semver solver with backtracking, version unification, conflict reporting, overrides (#1)
• Lockfile — deterministic aipm.lock creation, drift detection, --locked CI mode (#2)
• Features — default features, opt-out, additive feature unification across the graph (#3)
• Patching — aipm patch workflow for editing transitive deps without forking (#4)

Registry

• Install — aipm install with semver resolution, content-addressable store, integrity verification, strict isolation (#5)
• Publish — aipm-pack pack / publish with .aipm archives, dry-run, file allowlist, size limits (#6)
• Security — checksums, tamper detection, aipm audit, auth, scoped org permissions (#7)
• Yank — aipm-pack yank / un-yank, deprecation messages (#8)
• Link — aipm link / unlink for local dev overrides (#9)
• Local + Registry Coexistence — directory links, gitignore management, vendoring (#10)

Monorepo

• Orchestration — workspace protocol, catalogs, filtering by name/path/changed/dependents, Rush/Turborepo integration (#11)

Environment

• Dependencies — declare required tools, env vars, platforms, MCP runtimes; aipm doctor (#12)
• Host Versioning — [environment.hosts] section for Claude/Copilot/Cursor version constraints (#54)

Quality & Portability

• Guardrails — aipm lint, auto-fix, quality scoring on publish (#13)
• Compositional Reuse — publish/consume standalone skills, agents, MCP configs, hooks as packages (#14)
• Cross-Stack — verified portability across Node.js, .NET, Python, Rust, CMake; offline resolution (#15)

---

Why not apm?

microsoft/apm (apm-cli on PyPI) validates that AI plugin package management is a real problem. However, its architecture falls short for production use across several dimensions:

1. Runtime dependency. apm requires Python 3.9+ and 13 pip packages. This creates friction for .NET, Rust, Go, and Java teams, and adds version-management overhead. AIPM is a self-contained Rust binary — drop it in any repo regardless of tech stack.

2. YAML manifest. apm.yml uses YAML, which has the Norway problem (3.10 → 3.1), implicit type coercion (NO → false), indentation sensitivity, and active security CVEs in parsers. AIPM uses TOML — no coercion, no indentation traps, safe for AI-generated manifests.

3. No registry. Packages are git repos. There is no publish lifecycle, no immutable versions, no scoped namespaces, no centralized search, and no way to yank a broken release without deleting git tags. AIPM has a dedicated API registry with publish, yank, scoped packages, and multi-registry routing.

4. No semver resolution. apm pins by git ref — no ^1.0 ranges, no backtracking, no version unification. Two packages depending on different versions of the same dep each get a full clone. AIPM uses Cargo-model semver with caret/tilde ranges, backtracking, and major-version coexistence.

5. No integrity verification. The lockfile records commit SHAs but no file-level hashes. Force-pushes or host compromises silently change what a "version" resolves to. AIPM stores SHA-512 checksums per file and verifies on install.

6. Full git clones per project. Every project downloads full copies of every dependency — no deduplication, no global cache. AIPM uses a content-addressable global store (pnpm model) with hard links for 70%+ disk savings.

7. No dependency isolation. Everything in apm_modules/ is accessible to everything else. Phantom dependencies go undetected. AIPM enforces strict isolation — only declared dependencies are accessible.

8. Minimal security. No lifecycle script blocking (any package runs arbitrary code), no audit command, no principle-of-least-privilege binary split. AIPM blocks scripts by default, ships separate consumer/author binaries, and plans advisory-based auditing.

9. No transfer format. Packages are raw git repos — no archive format, no file allowlist, no secrets exclusion. AIPM uses deterministic .aipm archives (gzip tar) with sorted files, zeroed timestamps, and default secrets exclusion.

10. No offline support. Every install requires network access. AIPM supports --offline installation from the global cache.

11. No CI lockfile mode. apm install uses the lockfile if present, but there is no --locked mode that fails on drift. AIPM follows the Cargo model: install never upgrades, update explicitly resolves latest, --locked fails on any mismatch.

12. No workspace features. No workspace protocol, no catalogs, no dependency inheritance, no filtering. AIPM supports all of these for monorepo-scale plugin management.

13. Compilation coupling. apm tightly couples package management with AGENTS.md / CLAUDE.md generation. AIPM decouples packaging from host discovery — AI tools discover plugins naturally via directory scanning.

In short: apm is a useful prototype that proves the problem space. AIPM is designed to be the production-grade infrastructure for it.

---

Contributing

Contributions and suggestions are welcome! Please open an issue or pull request on GitHub.

License

This project is licensed under the MIT License.