Seeker

What are your agents missing?

Your agents already have skills installed — Claude, Codex, Hermes, OpenCode. Run one command and LoopSkill scans those directories on your own machine, then diffs what it finds against the public catalog. Read-only. Nothing is uploaded except the skill names and versions it needs to compare.

Run this now — no signup, no local clone

Needs one thing: a LoopSkill API key. Free signup gives you one instantly at /library — every MCP call, including this read-only scan, requires a key. Paste this, swap in your key, run it from a terminal on the machine whose skills you want scanned:

curl -sS https://app.loopskill.io/api/mcp/http/   -H "Content-Type: application/json"   -H "Accept: application/json, text/event-stream"   -H "x-api-key: $LOOPSKILL_API_KEY"   -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"seeker-cli","version":"1.0"}}}'   -D /tmp/lsk-headers.txt -o /dev/null

SID=$(grep -i mcp-session-id /tmp/lsk-headers.txt | cut -d' ' -f2 | tr -d '
')

curl -sS https://app.loopskill.io/api/mcp/http/   -H "Content-Type: application/json"   -H "Accept: application/json, text/event-stream"   -H "x-api-key: $LOOPSKILL_API_KEY" -H "mcp-session-id: $SID"   -d '{"jsonrpc":"2.0","method":"notifications/initialized"}' -o /dev/null

curl -sS https://app.loopskill.io/api/mcp/http/   -H "Content-Type: application/json"   -H "Accept: application/json, text/event-stream"   -H "x-api-key: $LOOPSKILL_API_KEY" -H "mcp-session-id: $SID"   -d '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"loopskill_seeker","arguments":{}}}'

Prefer an editor over raw curl? Any MCP-aware agent (Claude Desktop, Cursor, Cline, Hermes) wired per the MCP setup guide can just be asked: "call loopskill_seeker".

Real output

This is the actual response from a live call against https://app.loopskill.io/api/mcp/http/ on 2026-09-01 (verified while building this page — not a mockup):

{
  "vendors": {
    "claude": [],
    "codex": []
  },
  "recommendations": [],
  "unsupported_paths": ["hermes", "opencode"]
}

Read literally: this particular machine had no skills installed under ~/.claude/skills or ~/.codex/skills at scan time, and the Hermes/OpenCode vendor directories didn't exist on that box — unsupported_paths means "not found", not "unsupported vendor." On a machine with skills installed, vendors lists each one (name, version, path) and recommendations carries a diff entry per skill where the catalog has something newer or better-quality than what's on disk — reason is newer, better-quality, or missing.

What it scans

~/.claude/skills, ~/.codex/skills, ~/.hermes/skills, ~/.opencode/skills (honoring XDG_CONFIG_HOME on Linux, the equivalent on macOS/Windows).
Parses each installed SKILL.md's YAML frontmatter for name and version — no code execution, no network access to your machine beyond reading these files.
Diffs against the live public catalog server-side — READ-ONLY on vendor directories; it never writes, moves, or deletes anything on your machine.

Found gaps? Fix them from the same session.

Every recommended skill installs with loopskill_install — no extra key, no extra step.

Wire up MCP →