mcpField
R
ragdrift-mcp

ragdrift-mcp

Local
github.com/MukundaKattaMemory & Knowledge0Updated 20w ago
Repository

About ragdrift-mcp

ragdrift-mcp is an MCP server that helps LLMs interpret RAG drift scores and reason about alerts. It provides severity guidance, threshold recommendations, and a reference to drift dimensions; detection itself is handled by companion libraries or telemetry.

Key features

  1. 01Interpret drift scores and severity
  2. 02Recommend thresholds by sample and FP budget
  3. 03Explain five drift dimensions

Use cases

  1. 01Assess the severity of a RAG drift alert
  2. 02Choose thresholds for a sample size and false-positive budget

MCP support

Exposes tools to interpret drift scores, recommend thresholds, and explain dimensions to MCP clients.

Configuration

Local server config is the same for every client. Copy it into your client’s MCP config file.

mcp.json · Universal config
{
  "mcpServers": {
    "mukundakatta-ragdrift": {
      "command": "npx",
      "args": [
        "-y",
        "@mukundakatta/ragdrift-mcp"
      ]
    }
  }
}
Run commandnpx -y @mukundakatta/ragdrift-mcp

Local servers need a runtime on your machine (e.g. Node.js 18+). Config file locations differ by client; see your client’s docs.

Packages
npm@mukundakatta/ragdrift-mcp0.1.1

Tools

3

Saved tool list · updated 2w ago

ParameterTypeRequiredDescription
scorenumberYesThe drift score from a detector.
dimensionenumYesdata | embedding | response | confidence | query
thresholdnumberNoOptional. If provided, returns whether the score exceeds it.

Details

From the project README

@mukundakatta/ragdrift-mcp

MCP server for diagnosing RAG drift. Lets an LLM reason about drift scores without needing to install Rust or PyO3.

Sibling to the ragdrift crate + ragdrift-py PyPI wheel — this MCP server is the interpretation layer; the actual drift detection happens in those libraries (or your existing telemetry).

Tools

  • interpret_drift_score(score, dimension, threshold?) — plain-English interpretation, severity classification, suggested next steps.
  • recommend_thresholds(dimension, sample_size, false_positive_budget) — threshold values tuned to your sample size and FP budget.
  • explain_drift_dimensions() — structured reference for all 5 dimensions (data, embedding, response, confidence, query).

When to install

Install when you want Claude Desktop / Cursor / Cline / Windsurf / Zed to help diagnose RAG drift alerts. Example session:

User: "Got an embedding-drift alert with score 0.27 on a sample of 500. Should I be worried?"

LLM (calling interpret_drift_score): "Score 0.27 on the embedding dimension lands in the 'significant shift, investigate' band. The underlying methods are MMD² (RBF kernel) and sliced Wasserstein-1. Likely next steps: did the embedding model change? Was the corpus re-indexed? Compare baseline and current document samples."

Install — Claude Desktop

{
  "mcpServers": {
    "ragdrift": {
      "command": "npx",
      "args": ["-y", "@mukundakatta/ragdrift-mcp"]
    }
  }
}

Install — Cursor / Cline / Windsurf / Zed

Same shape per tool. npx -y @mukundakatta/ragdrift-mcp over stdio.

License

MIT OR Apache-2.0.

Comments

Markdown supported

New to MCP?

From what the protocol is to installing your first server, these answers cover most first questions.

An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.