docs · quickstart

Get up and running in five minutes

pakx is a static binary. No Node, no Python, no runtime. Pick your platform, paste one line, and you have a working install.

1.Install the CLI

macOS / Linux:

curl -fsSL https://pakx.dev/install.sh | sh

Windows (PowerShell):

irm https://pakx.dev/install.ps1 | iex

Each script downloads a prebuilt binary from GitHub Releases, verifies the sha256, and drops it on your PATH. Verify: pakx --version.

2.Initialise a project

pakx init

Creates an agents.yml in the current directory. This is the single manifest your agents share — commit it like a package.json or Cargo.toml.

3.Add a dependency

Add an MCP server from the federated index:

pakx search filesystem
pakx add mcp io.github.modelcontextprotocol/server-filesystem

The first call queries every source ( the official MCP Registry, Smithery, and the pakx registry) and prints them side-by-side. The second one writes the chosen dependency into your manifest.

4.Install across every agent on the machine

pakx install

Reads the manifest, resolves versions, and writes the right config files for every AI agent it detects (Claude Code, Cursor, Codex, Copilot, Windsurf — whichever are installed). The lockfile agents.lock pins what actually got installed so the next machine gets the same bits.

5.Publish your own package

One-time setup — log in with your GitHub identity:

pakx login

Then in a directory with a SKILL.md (or equivalent kind manifest):

pakx pack            # dry-run: builds <name>-<version>.tgz
pakx publish         # uploads to registry.pakx.dev

Your package is now visible at pakx.dev/p/pakx/<owner>/<name> and installable by anyone with pakx add skill <owner>/<name>.

6.Manage API tokens

Visit registry.pakx.dev/dashboard/tokens to issue, rotate, or revoke CLI tokens. Tokens are shown once at issue and stored as a one-way hash on the server.