@luckystack/mcp

Read-only MCP server that exposes a LuckyStack project's committed AI context to Claude Code as queryable tools — so an agent can ask precise questions instead of loading whole index files into context.

Install

Nothing to install in your app — it runs via npx. Add it to your project's .mcp.json:

{
  "mcpServers": {
    "luckystack": { "type": "stdio", "command": "npx", "args": ["@luckystack/mcp@latest"] }
  }
}

create-luckystack-app adds this automatically when AI instructions are enabled. Claude Code launches the server with your project root as the working directory.

Tools

| Tool | Answers | | --- | --- | | blast_radius(file) | "What breaks if I change this file?" (transitive importers) | | who_imports(file) | "What directly imports this?" | | god_nodes(limit?) | "Which files are risky hubs everything depends on?" | | list_decisions(tag?) | "What did we decide, and about what?" | | get_decision(id) | "Why did we decide X?" (full ADR) | | find_route(query) | "Which API/sync routes match this?" | | get_runbook(task?) | "How do I add an API/page/sync/helper here?" | | get_capability(name) | "Does a helper/component for this already exist?" |

All tools are read-only and read committed artifacts (docs/ai-graph.json, docs/decisions/, docs/AI_*.md). If an artifact is missing, the tool tells you which npm run ai:* command generates it.

It is independent of the playwright / chrome-devtools browser MCP servers — they coexist as separate entries in the same .mcp.json.

License

MIT