Docs · MCP
MCP server setup
The LoopSkill MCP server gives any MCP-aware agent 47 dedicated tools — search, install, sync, fleet ops, loops, bundles, personalities, and more. One hosted StreamableHTTP endpoint, same JSON shape. Cross-vendor: works in Claude Desktop, Cursor, Cline, Continue.dev, Zed, OpenClaw, Hermes, and Windsurf with the same config.
What MCP unlocks
What you can do with this
Once LoopSkill is wired in as an MCP server, your agent can search the skill catalog, fetch
full skill documentation, browse trending installs, and install a skill — all from inside
a conversation, without you touching a terminal. Ask your agent "find me an SEO audit
skill" and it calls loopskill_search,
reads the results, installs the winner with
loopskill_install, and reports back
— one turn of conversation, zero copy-paste. The same works for bundle syncs: a
a Pro agent can call
loopskill_skillify to
bootstrap a full bundle on a new machine in under 60 seconds.
Why ours
- Cross-vendor, one config. The same
hosted StreamableHTTP endpoint (
https://app.loopskill.io/api/mcp/http/) works in Claude Code, Cursor, Cline, Continue, Zed, OpenClaw, Hermes, and Windsurf. Single-vendor MCP marketplaces can't give you this — vendor lock-in is their business model. Ours is the opposite. - 47 dedicated tools, not a REST wrapper.
loopskill_search,loopskill_install,loopskill_sync,loopskill_skillify, plus fleet ops, loops, bundles, personalities, and sharing tools. The agent gets verbs it can reason about — not raw HTTP calls it has to construct and error-handle itself. - Signed tarballs with SHA256 verification. Every skill tarball is signed at publish time and verified on install. A tampered tarball is detected before it touches your skills directory — the agent rejects it and retries.
- Bundle-scoped keys (Pro). Give a client's agent a key scoped to a specific bundle — it installs exactly your chosen skills, nothing else in your private catalog. The agent PULLS the bundle — LoopSkill never pushes, and there is no client-deployment flow.
1. Get an API key
Sign in at app.loopskill.io/library and generate one. Every MCP call needs a key — including free-tier lookups (free signup gives a free key). The catalog installs the same on Free and Pro; Pro ($9.95/mo) raises limits — 50 private bundles and fleet sync.
2. Configure your client
Two ways to connect, depending on what your client supports.
Recommended · uvx loopskill-mcp (stdio bridge)
For clients that spawn a local stdio MCP process (Claude Desktop, Hermes, Codex CLI) —
loopskill-mcp is a published PyPI package: a
thin local bridge that proxies to the same hosted StreamableHTTP endpoint. No clone, no
manual venv — uvx fetches and runs it on demand.
{
"mcpServers": {
"loopskill": {
"command": "uvx",
"args": ["loopskill-mcp"],
"env": { "LOOPSKILL_API_KEY": "rec_live_your_key_here" }
}
}
}
MCP calls require a key for every operation (verified live: anonymous
tools/list
returns 401) — including free-tier lookups. For a truly keyless path, use the
meta-skill or raw REST install guide instead.
Manual · raw StreamableHTTP JSON
For clients that connect directly over HTTP (no local process) — point them at the hosted
endpoint with your key in the x-api-key header:
{
"mcpServers": {
"loopskill": {
"url": "https://app.loopskill.io/api/mcp/http/",
"headers": { "x-api-key": "rec_live_your_key_here" }
}
}
} Claude Desktop
File: ~/Library/Application Support/Claude/claude_desktop_config.json (macOS)
or %APPDATA%\Claude\claude_desktop_config.json (Windows). Use the config block above.
Cursor
Settings → MCP → Add new global MCP server. Paste the same JSON shape; Cursor reads it as
~/.cursor/mcp.json.
Cline (VS Code extension)
Open the Cline panel → Settings (gear icon) → MCP Servers → Edit Configuration. Same JSON shape as Claude Desktop.
Continue.dev
File: ~/.continue/config.json
{
"experimental": {
"modelContextProtocolServers": [
{
"transport": {
"type": "streamable-http",
"url": "https://app.loopskill.io/api/mcp/http/",
"headers": { "x-api-key": "rec_live_..." }
}
}
]
}
} Zed
File: ~/.config/zed/settings.json
{
"context_servers": {
"loopskill": {
"source": "custom",
"url": "https://app.loopskill.io/api/mcp/http/",
"headers": { "x-api-key": "rec_live_..." }
}
}
} 3. Restart your client
Then ask your agent: "Search the LoopSkill marketplace for an SEO audit skill." The agent will call loopskill_search and report results.
Tools exposed (47 total)
A few of the most-used ones — see llms.txt for the full list, grouped by category.
loopskill_search— full-text + tier + category filter, returns compact JSONloopskill_install— download, SHA256-verify, tar-slip-safe extractloopskill_sync— pull patch updates for all installed skillsloopskill_skillify— drop the master meta-skill into the agent's skills dirloopskill_recall— search across bundles, loops, personalities, and connectors in one call- Plus fleet ops (
loopskill_fleet_*), loops (loopskill_search_loops,loopskill_get_loop), sharing, and more.
Custom install dir
By default, loopskill_install writes to
~/.claude/skills/<slug>/.
Override with LOOPSKILL_INSTALL_DIR env var
(e.g. ~/.codex/skills for Codex,
./skills for project-local installs in Cursor or Cline).
Machine-readable manifest: app.loopskill.io/llms.txt · Self-host the whole platform: self-host guide (MPL-2.0)