Paks
A package manager that lets you create, install, publish, and share Agent Skills across coding agents like Claude Code and Codex.
- Source repo
- stakpak/paks
- Stars
- ★ 66
- Last updated
- 4mo ago
- License
- Apache-2.0
- Primary language
- TypeScript
- FA score
- 42/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude Code
- Cost
- Free, no paid service needed
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- A platform team wants its coding standards and deployment runbooks packaged as skills so Claude Code and Cursor users install the same version.
- Not a fit if
- Teams that need a polished GUI; the web registry is documented only as a local dev server
- Users who want zero-config, GUI-driven setup without a terminal
- Teams that expect skills to work without installing the Paks CLI
- Website
- paks.stakpak.dev ↗ set on GitHub
- Maintenance
- Commits in the last 6 monthslast commit 2026-07-04latest release v0.1.18 (2026-01-21)
- Source review
- 42/100 · Major gaps
What does this agent do, and when should you use it?
Paks is a CLI-first package manager for Agent Skills, built in Rust by Stakpak and aligned with the AgentSkills standard. A skill is a reusable instruction set packaged as a SKILL.md file with YAML frontmatter and Markdown body, optionally accompanied by scripts, references, and assets directories. The CLI exposes create, install, publish, validate, list, remove, search, info, plus agent management and login/logout, and can install a skill into the global or project directory of stakpak, claude-code, cursor, vscode, copilot, goose, or opencode. Install sources include the registry, GitHub/GitLab browser URLs, and local paths, with semantic versioning and dependency declarations. The repository also ships a TanStack Router-based web registry with a REST API for browsing and searching skills, though the README only documents a local development port. It addresses the problem of distributing versioned coding standards and domain knowledge consistently across multiple coding agents.
paks create <name> scaffolds a skill directory from a template (basic, devops, coding), optionally adding --with-scripts, --with-references, and --with-assets. paks validate checks structure and metadata against the Agent Skills spec, and --strict promotes warnings to errors. paks install <source> resolves a source—a registry name, a full GitHub/GitLab browser URL, or a local path—and writes it to the directory derived from --agent and --scope (global/project), such as ~/.claude/skills or ./.claude/skills, overridable with --dir; it also supports --version and --force. paks publish logs in, bumps the version via --bump patch|minor|major, and pushes to the registry, with --dry-run for preview. paks list prints installed skills as table/json/yaml, while paks search and paks info query the registry. paks agent add/remove/default/show maintains custom agent directories, and configuration lives in ~/.paks/config.toml and a project-level .paks/config.toml, with precedence CLI flags > project config > global config > defaults.
- A platform team wants its coding standards and deployment runbooks packaged as skills so Claude Code and Cursor users install the same version.
- A DevOps engineer packages Kubernetes deployment steps and distributes them with
paks install kubernetes-deploy --agent stakpak. - An individual developer shares a skill from a GitHub repo and the recipient installs it by pasting the browser URL.
- A skill author scaffolds with
paks createand gates releases withpaks validate --strictbefore bumping versions. - A repository commits
.paks/config.tomlto standardize the default agent and installation scope for everyone on the team. - Someone managing many agents audits everything installed with
paks list --all --format json.
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Install / run
brew install paks
- Config files
~/.paks/config.toml- Runtime
- Rust 2024
- Install notes cover
- macOS · Linux
How do you install or deploy this agent?
On macOS and Linux via Homebrew:
brew tap stakpak/stakpak
brew install paksOr download a prebuilt binary from the Releases page, for example Linux x86_64:
curl -L https://github.com/stakpak/paks/releases/latest/download/paks-linux-x86_64.tar.gz | tar xz
sudo mv paks /usr/local/bin/Building from source requires the Rust 2024 edition:
git clone https://github.com/stakpak/paks.git
cd paks
cargo build --release -p paks-cli
cp ./target/release/paks ~/.local/bin/Verify the install:
paks --version
paks --helpHow do you use this agent?
Create and validate a skill:
paks create my-awesome-skill --with-scripts --with-references
paks validate my-awesome-skill --strictInstall from the registry or a GitHub URL into a chosen agent:
paks install kubernetes-deploy --agent claude-code --scope project
paks install https://github.com/user/repo/tree/main/path/to/skillInspect what is installed and register a custom agent:
paks list --all --format json
paks agent add my-agent --dir ~/my-agent/skills --project-dir .my-agent/skillsDry-run a publish, then log in and push:
paks publish my-awesome-skill --dry-run
paks login
paks publish my-awesome-skill --bump patchGlobal configuration lives in ~/.paks/config.toml:
default_agent = "stakpak"
default_scope = "global"
[registries.default]
url = "https://registry.paks.dev"What are this agent's strengths and limitations?
- One CLI installs the same skill into stakpak, claude-code, cursor, vscode, copilot, goose, and opencode directories, replacing manual copying.
- Three install sources—registry, GitHub/GitLab browser URL, and local path—keep switching costs low.
- SKILL.md follows the AgentSkills spec and extends it with version, authors, keywords, and dependencies for package management.
- Configuration precedence (CLI flags > project config > global config) lets teams commit
.paks/config.tomlto standardize installs. - Apache-2.0 licensed, Rust-implemented, with Homebrew and prebuilt binaries for macOS, Linux, and Windows.
- The web registry is documented only as a local development server at
http://localhost:3001; no public hosted instance is described. - Publishing and login depend on the
registry.paks.devservice, and self-hosting the registry backend is not covered in the README. - Default agents are tied to concrete product directories (e.g.
~/.claude/skills,.config/opencode/skill), so directory changes require manualagent addmaintenance. - Building the full monorepo from source needs the Rust 2024 edition (nightly) as well as Node.js 18+ and pnpm 8+.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Paks This agent | 42 · Major gaps | CLIFree | ★ 66 | 4mo ago | TypeScript | Codex · Claude Code |
| Beautiful Rust | 71 · Some gaps | Agent plugin / skillFree + model costs | ★ 2.1k | 12d ago | Python | ChatGPT · Codex · Claude Code · Claude.ai |
| Cutaway | 47 · Major gaps | — | ★ 46 | 1mo ago | JavaScript | Codex · Claude Code |
| RTK Command Output Optimizer | 72 · Some gaps | CLIFree | ★ 87k | today | Rust | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
Install writes into user directories (e.g. ~/.claude/skills) and pulls skills from remote registry/git, an external effect; remove has --yes to skip confirmation and publish has --skip-validation, so confirmation exists but is bypassable. Login token via --token risks shell-history exposure; no sensitive-data handling guidance. Dependencies are common crates (reqwest/rustls) but no audit or lockfile evidence. Rollback is only --force reinstall, no version revert or backup. Source attribution is only a Stakpak mention, no signing or verification. Thin overall, deductions warranted.
README and main.rs command sets largely align, but README's install --version option is absent from the Install subcommand in main.rs, a doc/code inconsistency. Dependencies use loose ranges (e.g. clap=4) with no Cargo.lock evidence. Failure messages are not visible in the sampled source, so error quality cannot be confirmed.
Targets multiple AI coding agents (Claude Code, Cursor, Copilot, etc.) with create/install/publish/validate scenarios and a clear audience. Capability boundaries (whether installed skills execute scripts, the experimental allowed-tools field) are unexplained, and trigger precision depends on SKILL.md descriptions with no concrete skill sample. Environment fit is solid with Homebrew, binary, and source install paths.
README information architecture is clear with install, quick start, CLI reference, configuration. Install notes are thorough. However Cargo.toml workspace.package license says MIT while LICENSE and README badge say Apache-2.0, a license inconsistency. No CHANGELOG despite version 0.1.18. Maintenance responsibility points only to Stakpak and Discord; publisher unverified. No known-limitations section.
CLI supports table/json/yaml output, validate has strict mode, publish has dry-run, so output usability is good. Clear marginal value as a skill package manager. Cost-benefit is limited by the registry defaulting to localhost:3001 (development), raising production-readiness doubts.
README claims partially map to source (commands, agent directories), but registry URL and API endpoints cannot be verified from the sample. CI workflow shows test steps but test contents are unseen. Fact vs inference is not clearly separated in docs (allowed-tools marked experimental without detail).
- Cargo.toml declares MIT while LICENSE/README state Apache-2.0; license inconsistency must be clarified before use.
- Registry defaults to localhost:3001 (development), so production readiness is unproven.
- Login token can be passed via --token on the command line, risking shell-history exposure.
- Installing skills writes to user directories and pulls remote content with no visible signing or integrity verification.
- README's install --version option is missing from the Install subcommand in source, a doc/code mismatch.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
Release notes excerpt
0.1.18 Released on 2026-01-21. Bug Fixes - Use Path::ends_with for cross-platform test compatibility Maintenance - Bump version to 0.1.18View on GitHub ↗ -
Release notes excerpt
0.1.16 Released on 2026-01-16. Bug Fixes - Sync remote tags before listing for publish Maintenance - Bump version to 0.1.16View on GitHub ↗ -
Release notes excerpt
0.1.15 Released on 2026-01-13. Bug Fixes - Use cross-platform path assertions in config tests Maintenance - Bump version to 0.1.15View on GitHub ↗
FAQ
Does using Paks cost anything?
Where do installed skills end up?
--agent and --scope. Global writes to the user directory (e.g. ~/.claude/skills), project writes inside the repo (e.g. ./.claude/skills), and --dir overrides both since CLI flags have the highest precedence.Can I use it without publishing to the public registry?
install accepts GitHub/GitLab browser URLs and local paths. Publishing, searching, and info lookups rely on a registry service, and the README only documents the default https://registry.paks.dev.What does a valid SKILL.md require?
name (1-64 characters, lowercase and hyphens) and description (1-1024 characters) are required; fields such as license, metadata, and allowed-tools are optional per the spec, and Paks adds version, authors, keywords, and dependencies.How do I check a skill before publishing?
paks validate <path>; --strict turns warnings into errors. Before pushing, paks publish <path> --dry-run shows what would be published.Is Paks open source?
Is Paks safe to use?
Is Paks still maintained?
Who makes Paks?
What is Paks's official website?
Thanks — we use this to improve the page.