Built by AgentSystem — the platform for shipping production-grade agentic systems. Visit agentsystem.dev to learn more.
The plugins in this repo are the free essentials. The real arsenal — the senior-engineer-grade skills I use to ship features end-to-end, audit auth, hunt regressions, harden types, write tests, and orchestrate multi-agent pipelines — is launching soon as AgentSystem Premium.
👉 Join the waitlist at agentsystem.dev
- 🧠 80+ production-tested skills covering the full software lifecycle (plan → ship → release)
- ⚡ Multi-agent orchestration — fan out parallel work, gate approvals, never lose context
- 🛡️ Built-in audits for authz, security regressions, data integrity, accessibility, performance
- 🎯 Stack-aware — TanStack Start, Electron, Next.js, and more, with idiomatic patterns baked in
- 🎁 Early-access pricing for waitlist members + first dibs on new skill drops
Stop wiring agents from scratch. Reserve your spot →
A plugin marketplace for agentic coding CLIs. Native support for the Claude Code marketplace feature, with copy-paste install paths for Cursor and Codex so the same plugins work everywhere.
| Name | Description |
|---|---|
commit-helper |
Drafts a conventional-commit message from staged changes. |
code-reviewer |
Reviews the current diff for bugs, security issues, and style. |
ship-pr |
One-shot ship: groups a dirty working tree into clean commits, pushes the branch, and opens a polished GitHub PR with auto-applied labels and "might address #X" issue links. |
In any Claude Code session:
/plugin marketplace add AgentSystemLabs/essentials
/plugin install commit-helper@essentials
/plugin install code-reviewer@essentials
/plugin install ship-pr@essentials
That's it. Slash commands and subagents become available in your next prompt. To remove:
/plugin uninstall commit-helper@essentials
Install the skills into the current project (or globally) with the bundled CLI:
npx agentsystem-essentials init # → ./.claude/skills
npx agentsystem-essentials init --global # → ~/.claude/skills
npx agentsystem-essentials init --harness codex # → ./.codex/skills
npx agentsystem-essentials list
npx agentsystem-essentials uninstallEach plugin ships a Cursor rule under plugins/<name>/cursor/rules.mdc. Copy the file you want into your project's .cursor/rules/ directory:
mkdir -p .cursor/rules
curl -o .cursor/rules/commit-helper.mdc \
https://raw.githubusercontent.com/AgentSystemLabs/essentials/main/plugins/commit-helper/cursor/rules.mdcRepeat per plugin. Cursor picks up new rules automatically.
Each plugin ships an AGENTS.md snippet under plugins/<name>/codex/AGENTS.md. Append the snippet to your project's AGENTS.md:
curl https://raw.githubusercontent.com/AgentSystemLabs/essentials/main/plugins/commit-helper/codex/AGENTS.md \
>> AGENTS.mdCodex (and any agent that reads AGENTS.md) will pick up the new instructions on the next session.
What it does. Takes a dirty working tree (often 20–50+ changed files after a feature) and produces a merge-ready GitHub pull request in one pass: logically grouped commits → branch created and pushed → PR opened with a clear title, a high-level description, labels matched against your repo's existing label set, and "might address #X" links to plausibly-related open issues.
Why use it.
- Replaces the one-mega-commit habit with a clean, layered commit chain a reviewer can read top-to-bottom.
- Writes the PR description the way a senior engineer would — what changed, why it was needed, what to look out for — without dumping diffs or code snippets.
- Auto-applies labels by matching against
gh label list(never invents labels, never auto-creates them). - Links related open issues with the literal phrase
Might address #Ninstead ofCloses #N, so heuristic matches don't auto-close the wrong issue on merge. - Flags risky areas (auth, payments, migrations) so reviewers know where to look closely.
When to use it. You finished work, your tree is dirty, and you want commits + branch + PR in one move. Trigger phrases: ship it, /ship-pr, wrap this up into a PR, commit and open a PR.
Skip it for: single-file copy tweaks, branches already pushed with an open PR (use gh pr edit), amend/fixup workflows, force-push requests, or repos without a GitHub remote.
How it runs. A phased Process workflow with two confirmation gates:
- Inventory + risk scan (parallel
git+ghreads). - Group files into commits — gate 1: you approve the grouping plan before anything is staged.
- Commit each group in dependency order (schema → backend → frontend → chore).
- Create branch if needed (no push yet).
- Synthesize PR title, body, labels, and related-issue links from the cumulative diff.
- Gate 2: shows you the full PR preview; you
(y)es / (e)dit / (c)ancel. - Push and
gh pr createonly ony.
The skill never force-pushes, never skips hooks, never auto-closes issues, never invents labels, and never edits files outside your working tree.
-
Create
plugins/<your-plugin>/with the following layout:plugins/<your-plugin>/ ├── .claude-plugin/plugin.json # required for Claude Code ├── commands/*.md # slash commands (optional) ├── agents/*.md # subagents (optional) ├── hooks/ # hooks (optional) ├── cursor/rules.mdc # Cursor rule equivalent └── codex/AGENTS.md # AGENTS.md snippet equivalent -
Register it in
.claude-plugin/marketplace.jsonunderplugins. -
Open a PR. Keep each plugin focused on one job — small and composable beats one mega-plugin.
The Claude Code side follows the official plugin spec (commands, agents, hooks, MCP servers). The cursor/ and codex/ siblings restate the same intent in those tools' native formats so the user-facing behavior stays consistent across CLIs.
MIT