agentpack

module
v0.8.6 Latest Latest
Warning

This package is not in the latest version of its module.

Go to latest
Published: Oct 28, 2026 License: MIT

README

agentpack

Pin skills and plugins for coding agents in one project manifest. Agentpack resolves packages, records commits and content hashes in a lockfile, and stages supported artifacts for Claude Code, Cursor, OpenCode, Codex, Grok and Antigravity.

Agentpack initializing a project, locking a local skill and verifying staging

Actual CLI output from a local skill in a disposable, network-disabled container. This checks package staging, not an agent's response to the skill.

Install

Download an archive for your platform from GitHub releases, check it against the release's checksums.txt, extract it and put agentpack on your PATH. Release targets are macOS and Linux on ARM64/x86-64, plus Windows x86-64.

Installer scripts, Homebrew and source-build instructions describe the other options. The source build requires Go 1.26 or newer:

git clone https://github.com/olegpustovit/agentpack.git
cd agentpack
go build -o ./agentpack ./cmd/agentpack
./agentpack --version

This is pre-release software. CLI behavior, lockfiles and staging layouts can change between versions.

Try a local skill

With agentpack on your PATH, run this in a new directory:

mkdir agentpack-demo
cd agentpack-demo
agentpack init
mkdir local-skill
cat > local-skill/SKILL.md <<'SKILL'
---
name: local-skill
description: Summarize a change and explain how to verify it.
---
Describe the change in plain language. Include the verification command.
SKILL
agentpack add ./local-skill --no-sync
agentpack sync
agentpack sync --verify-only

You now have agentpack.toml, pack.lock and staged copies of the skill. The lock records a content hash. A local path dependency follows your directory's contents when you lock or sync; it does not freeze the source directory. Demo commands, output and assertions show the checked run.

For a GitHub package, use a repository or subdirectory spec such as agentpack add anthropics/skills/skills/canvas-design. Review the content and lockfile changes before launching it. Commit both the manifest and lockfile if the project should share those dependencies.

Launch an installed agent

Run one launcher from your project. Install and authenticate the underlying agent separately.

Command Agent
agentpack claude Claude Code
agentpack agent Cursor Agent
agentpack opencode OpenCode
agentpack codex Codex
agentpack grok Grok
agentpack agy Antigravity

Arguments after the launcher are forwarded to that agent. Staging an artifact does not establish that every agent discovers it or follows its instructions. Harness guides describe the adapters; the conversion table lists fallbacks.

Files and verification limits

Downloaded content and metadata live under AGENTPACK_HOME, normally ~/.local/share/agentpack on Unix, or an XDG/Windows data location. Staging is per project. Previous generations are retained for active sessions and can consume disk space.

The manifest and lockfile are project files by default. Cursor and Antigravity launchers can add workspace links; eject deliberately writes staged content into a project. Some adapters copy or link user configuration and synchronize session or authentication state. Read the staging guide before treating this as an isolated environment.

A full sync verifies hashed cache entries. An unchanged launch can take a metadata-based fast path; set AGENTPACK_FULL_VERIFY=1 to hash cached content on every launch. Legacy lock entries without hashes are allowed with warnings unless AGENTPACK_REQUIRE_VERIFIED=1 is set. sync --verify-only checks staged files. Hashes detect changes to locked content; they do not establish that its instructions or executable hooks are safe.

MCP pinning depends on the launcher and recorded status. Complete dependency-tree locking is limited to supported npx cases. Other runners may pin only a top-level package or remain unpinned. See MCP limits.

Documentation and development

Use agentpack --help for the command list. The documentation site, manifest guide, modes, external environments and installation/reference cover the detailed workflows, including the experimental proxy.

Contributors should read AGENTS.md. Run make test and make lint for Go checks. Preview the book with mdbook serve docs. Evaluation notes distinguish package integrity, native discovery and model behavior.

License

MIT. Experimental proxy work credits raine/claude-code-proxy.

Directories

Path Synopsis
cmd
agentpack command
internal
cli
eject
Package eject writes what Agentpack stages into the project itself, where each agent's own front ends (IDE extensions, desktop apps, the agents' CLIs started without Agentpack) read project configuration.
Package eject writes what Agentpack stages into the project itself, where each agent's own front ends (IDE extensions, desktop apps, the agents' CLIs started without Agentpack) read project configuration.
mcp

Jump to

Keyboard shortcuts

? : This menu
/ : Search site
f or F : Jump to
y or Y : Canonical URL