Documentation · Getting started
From zero to skill-aware in 60 seconds.
The agent-first workflow
Tell your agent: "install the relevant skills from app.loopskill.io".
The agent handles the rest via MCP — it searches the catalog, picks the best-scored skill for your task, installs it, and reports back. You don't browse the website by hand.
Prerequisites
- A LoopSkill API key — free, no card required. Every catalog skill installs on the Free tier; Pro ($9.95/mo) raises limits (private bundles, fleet sync, key cap).
- One of the supported agent runtimes below.
- Your key exported:
export LOOPSKILL_API_KEY=rec_xx...
Get your key: Sign up → Dashboard → API Keys → New key.
Step 1 — Install the meta-skill
The meta-skill is a single SKILL.md file that teaches your agent the entire
marketplace. Drop it into your agent's skills directory once. Your agent will then be able
to search and install any skill on-demand.
Hermes (native)
Hermes auto-discovers SKILL.md files under ~/.hermes/skills/:
mkdir -p ~/.hermes/skills/loopskill
curl -sL https://app.loopskill.io/skill \
-o ~/.hermes/skills/loopskill/SKILL.md Claude Code (native)
Claude Code reads skills from ~/.claude/skills/:
mkdir -p ~/.claude/skills/loopskill
curl -sL https://app.loopskill.io/skill \
-o ~/.claude/skills/loopskill/SKILL.md Codex CLI (native)
OpenAI Codex CLI uses the same skill directory layout:
mkdir -p ~/.codex/skills/loopskill
curl -sL https://app.loopskill.io/skill \
-o ~/.codex/skills/loopskill/SKILL.md Claude Desktop (MCP)
Claude Desktop uses the Model Context Protocol. LoopSkill is a hosted MCP
server — no local install. Add it to your
claude_desktop_config.json:
{
"mcpServers": {
"loopskill": {
"url": "https://app.loopskill.io/api/mcp/http/",
"headers": { "x-api-key": "rec_xx..." }
}
}
} After saving, restart Claude Desktop. You'll see the LoopSkill tools (search, install, sync) available in the tool picker. See MCP setup for full details.
Cursor (MCP)
Open Cursor Settings → MCP → Add Server:
{
"name": "loopskill",
"url": "https://app.loopskill.io/api/mcp/http/",
"headers": { "x-api-key": "rec_xx..." }
} Step 2 — Search the catalog
Once the meta-skill is installed, your agent can search the catalog. Try asking:
- "Search LoopSkill for an SEO audit skill"
- "What LoopSkill skills exist for client reporting?"
- "Find a LoopSkill skill that generates content calendars"
Under the hood, the agent calls GET /api/skills/search?q=... and ranks results
by quality_score, last_verified, and install count.
Step 3 — Install your first skill
Tell your agent to install a skill. Example:
- "Install the
super-memoryskill from LoopSkill" - "Install the best SEO audit skill from LoopSkill"
The agent calls GET /api/skills/install?slug=super-memory&mode=files,
downloads the tarball via an expiring signed URL, verifies its checksum_sha256,
and writes the files locally.
A _meta.json records the install for future syncs.
The free skill super-memory (Cognee + Postgres + pgvector, MIT-licensed) is a great first install — it works immediately with no subscription.
Step 4 — Sync to stay current
Skills are versioned. Ask your agent — "sync my LoopSkill skills to latest" — to
pull patch updates. The sync loop runs via the loopskill_sync MCP tool
(or the LoopSkill meta-skill on a non-MCP agent); the standalone loopskill
CLI does not have a sync subcommand today (it ships
import/diff/pull/apply only).
What's in the catalog
Skills across marketing, code, web scraping, ops, and sales — see the live count on
the catalog page or GET /api/stats. Every skill carries
a quality_score, install count, and last-verified date.
Browse the full catalog — but the fastest path is always to ask your agent.
Pricing
- Free: API key, catalog access, all skills (including super-memory), 2 private bundles.
- Pro ($9.95/mo): 50 private bundles, fleet sync, priority support. The full public catalog is already unlimited on Free — Pro raises limits, it doesn't unlock exclusive skills. Public bundles are unlimited on every tier, including Free.
- Self-host: run the whole platform yourself (MPL-2.0) — no account, no caps, no card. See the self-host guide.
Next steps
- How it works — the 3-level progressive loading model
- MCP setup — full Claude Desktop / Cursor / Cline config
- New agent runtime — 10-minute guide for unlisted runtimes
- Security model — auth, signed tarballs, sandboxing