Fewer, better, checked.

Agent Loadout is a small library of setups for AI coding agents — the rules, skills and MCP configs that turn a general-purpose agent into one that knows your stack.

Why

There are directories listing millions of agent skills. Almost none have been run. Many are near-duplicates, some are tied to one tool, and a few are unsafe. If you are vibe-coding a game or a SaaS, you do not want to wade through that; you want the one setup that works with the agent you already use.

So we keep the library small and opinionated, aimed at vibe coders and indie game devs, and we package everything as a loadout: one bundle per job, installable into Claude Code, Codex, Cursor, Gemini CLI or Copilot.

Verified install (Level 1, live)

Every loadout is checked automatically on every change and once a week, for free, with no AI involved:

• every file parses, and skill and Cursor rule front matter has the fields each agent needs;
• each file installs where Claude Code, Codex, Cursor, Gemini CLI and Copilot document they look, and the laid-out project is re-discovered correctly;
• MCP configs are well formed, convert to each agent's format, hold no literal secrets, and any npm or PyPI package they name really exists;
• the same safety scanner as the submit form finds no credentials, risky shell, hidden text or prompt-injection wording;
• references between files resolve for every agent, sizes are within limits, and the licence is clear;
• where a free, repeatable check exists, a small starter project builds with the loadout applied and the loadout's own code examples compile. Where it can't (Unity needs a paid-for editor install), the check says not applicable and why.

Each detail page shows every check, its result and the date. A pass here means the files are sound; it does not mean an agent has used them.

Agent tests (Level 2, not live)

Later, a harness will run real coding agents against each fixture with and without the loadout and record pass or fail per agent, with the date and agent version. That is not running yet. Until it is, every loadout says not agent-tested. We will never show a tick for a test that did not run.

Built on open formats

• AGENTS.md for project instructions
• Agent Skills (SKILL.md) for reusable skills
• Model Context Protocol for tool servers
• llms.txt so agents can read this site directly

Who runs it

Agent Loadout is an experiment by JNW IT, run largely by AI agents with a human in charge of what gets published. Content is MIT-licensed unless a listing says otherwise.