setup

skill-concierge setup

First-time bootstrap (and post-update refresh) for the vendored skill-search engine. setup.sh is idempotent — safe to re-run any time.

Prerequisites

• Python 3.10–3.12 on PATH (or set SKILL_PYTHON=/path/to/python3.12).
• Docker / OrbStack running (hosts the Qdrant vector store).

If either is missing, tell the user and stop — setup.sh cannot proceed without them.

Steps

1. Run the bootstrap from the plugin root:

   bash "$CLAUDE_PLUGIN_ROOT/setup.sh"
   

   (When working from a git clone instead, cd into the repo and run ./setup.sh.) It performs four idempotent steps: stable venv + deps → Qdrant container → build/refresh the index → apply curated overrides to ~/.claude/settings.json (backed up first).

2. De-duplicate any prior user-scope MCP so only the bundled one runs:

   claude mcp remove skill-search -s user
   

3. Restart Claude Code so the MCP server and the settings overrides take effect.

4. Verify the install:

   python3 "$CLAUDE_PLUGIN_ROOT/scripts/doctor.py"
   

   Expect status: OK. In Claude Code, /mcp should list skill-concierge:skill-search as connected. If anything is degraded, hand off to the skill-concierge:doctor skill.

When to re-run

• After any plugin update (/plugin marketplace update + reinstall) — refreshes the engine copy in the stable venv and rebuilds the index.
• When the Qdrant data volume or the venv was wiped.

Notes

• The venv lives at ~/.local/share/skill-concierge/venv (outside the plugin cache, so it survives reinstalls — ADR-0004). Override with SKILL_CONCIERGE_VENV.
• setup.sh reads the embedder + Qdrant URL from .mcp.json so the built index can never diverge from the model the live MCP uses.