Model Context Protocol

Setup

{
  "mcpServers": {
    "noumenon": {
      "command": "noum",
      "args": ["mcp"]
    }
  }
}

{
  "mcpServers": {
    "noumenon": {
      "command": "noum",
      "args": ["mcp"]
    }
  }
}

{
  "mcpServers": {
    "noumenon": {
      "command": "noum",
      "args": ["mcp"]
    }
  }
}

Tool Catalog

Twenty-five tools across discovery, querying, pipeline control, introspection, and benchmarks. All return JSON; most accept the repo name or path as the first argument.

Cheap, Medium, Expensive

The three tools that surface knowledge from a digested repo sit on a clear cost gradient. Reach for the cheapest one that does the job.

• noumenon_search — milliseconds, zero LLM calls. TF-IDF cosine similarity over per-file and per-component summaries. Returns ranked file paths with scores. The cheapest way to find "which files are about X." Use this when you want discovery without the price of a full agent loop.
• noumenon_query — milliseconds, zero LLM calls. Run a named or raw Datalog query. The right tool when you know the question maps to existing structure ("hotspots," "files-by-layer," "top-contributors") and want the structured result back.
• noumenon_ask — seconds, multiple LLM calls. Iterative agent. TF-IDF warm start, routing-model hint, then a loop over Datalog queries until it has enough to answer. Use when the question is open-ended or requires composing multiple facts. See Ask for how it works.