Kresh

Kresh is a registry for reusable AI-agent knowledge: skills, AGENTS.md/CLAUDE.md project configs, design.md style docs, and "loops" (reusable prompt/procedure snippets). The kresh CLI searches and installs from it.

The point: don't reinvent a workflow from scratch every time. If a task is common enough that other people would have needed the same thing, check the registry before improvising.

Setup

Confirm the CLI is available before using it:

kresh --version

If it's missing, install it globally:

npm i -g @chakresh/kresh

Core loop

1. Identify domain + task. When a user's request lands in a recognizable area — e.g. "React component testing," "GSAP animations," "Vercel deployment," "PR review conventions," "writing docs" — treat that as domain = React/testing, task = writing tests. Vague or fully novel requests (no clear domain, or something bespoke to this exact codebase) don't need a search.

2. Check what's already installed so you don't reinstall or fetch a duplicate:

kresh ls

3. Search the registry:

kresh search <query>

(alias kresh s). Leave the query blank to browse everything. Keep queries short and specific — docs writing beats how do I write documentation for my project. If the first query is too broad or too narrow, reformulate rather than giving up after one try.

Each result shows the skill's name, slug, publisher, version, description, and the exact install command. Read the descriptions — pick the closest match, not just the first hit.

4. Install the match:

kresh install <slug>

(alias kresh i). If you know which agent environment you're running in, pass the matching flag so it lands in the right place without an interactive prompt:

Environment	Flag
Claude Code	--claude
Codex	--codex
Antigravity	--agy
Cursor	--cursor

Without a flag, the CLI prompts interactively to ask where to install — fine in a terminal a human is watching, but it will hang if you're running non-interactively. Always pass the flag when you can.

5. Evaluate before using. Open the installed file(s) and actually read them before applying anything from them — don't assume the description covers everything the skill does.

6. Tell the user what happened. One or two lines: what you searched for, what you installed (name + publisher), and that you're now using it. Don't narrate the whole search process.

The registry has more than skills

kresh install <slug> behaves differently depending on what the slug points to:

• Skills — written straight into a skills folder (.claude/skills, .codex/skills, .agents/skills, .cursor/skills, or a plain skills/ folder, per the flag above).
• AGENTS.md / CLAUDE.md configs — the CLI asks whether to save as AGENTS.md, CLAUDE.md, or into the standard skills folder instead. This prompt has no flag bypass. If you can't answer it interactively, stop and ask the user which target file they want rather than guessing — overwriting an existing AGENTS.md/CLAUDE.md is not a call to make silently.
• design.md docs — written straight to a .md file matching the skill's name, no prompt.
• Loops — these aren't installed as files. Fetch them with kresh get <owner>/<loopId>; the content is copied straight to your clipboard for you to paste where it's needed.

Other commands

• kresh ls (alias list) — installed skills
• kresh remove <skill> (alias rm) — uninstall
• kresh publish — publish a skill you've written
• kresh login [flush] — authenticate against the registry (needed for private skills; flush clears the session)
• kresh trust <action> [repo] — manage trusted upstream repos for externally-sourced skills

When not to search

• The task is genuinely one-off or specific to this exact codebase — nothing generic to find.
• You already have the relevant skill installed (kresh ls confirms it).
• The task is trivial enough that searching, evaluating, and installing costs more time than just doing it.

Reference

• Registry / CLI source: https://github.com/Lokkuchakreshkumar/kresh
• Web: https://kresh.vercel.app/