---
title: "AI Integration via MCP"
output: rmarkdown::html_vignette
vignette: >
  %\VignetteIndexEntry{AI Integration via MCP}
  %\VignetteEngine{knitr::rmarkdown}
  %\VignetteEncoding{UTF-8}
---

```{r, include = FALSE}
knitr::opts_chunk$set(collapse = TRUE, comment = "#>")
```

## Overview

ReliaGrowR can expose its core analysis functions as **Model Context Protocol (MCP)
tools**, allowing AI assistants such as Claude to call them directly during a
conversation. This is powered by the
[mcptools](https://posit-dev.github.io/mcptools/) package from Posit.

Once configured, an AI assistant can:

- Fit reliability growth models on data you describe
- Forecast future failures or MTBF
- Run goodness-of-fit tests
- Plan reliability demonstration tests
- Analyze repairable systems with MCF and NHPP models

## Installation

Install the required packages:

```r
install.packages("mcptools")   # MCP server framework
install.packages("ellmer")     # Tool definition helpers (already in ReliaGrowR Suggests)
```

## Starting the MCP Server

The server is started with a single call:

```r
ReliaGrowR::rga_mcp_server()
```

By default this uses stdio transport (suitable for Claude Code and Claude
Desktop). To use HTTP transport instead:

```r
ReliaGrowR::rga_mcp_server(type = "http", port = 8080)
```

## Configuring Claude Code

Add the server to Claude Code from your terminal:

```bash
claude mcp add -s user reliagrowR -- Rscript -e "ReliaGrowR::rga_mcp_server()"
```

The `-s user` flag stores the configuration in your user-level settings so it is
available in every project.

## Configuring Claude Desktop

Add the following block to `claude_desktop_config.json`
(`~/Library/Application Support/Claude/claude_desktop_config.json` on macOS):

```json
{
  "mcpServers": {
    "reliagrowR": {
      "command": "Rscript",
      "args": ["-e", "ReliaGrowR::rga_mcp_server()"]
    }
  }
}
```

Restart Claude Desktop after saving.

## Available Tools

| Tool | Function | Description |
|---|---|---|
| `rga` | `rga()` | Crow-AMSAA reliability growth model |
| `nhpp` | `nhpp()` | NHPP Power Law / Log-Linear for repairable systems |
| `duane` | `duane()` | Duane log-log regression |
| `mcf` | `mcf()` | Mean Cumulative Function (Nelson-Aalen) |
| `predict_rga` | `predict_rga()` | Forecast cumulative failures from RGA model |
| `predict_duane` | `predict_duane()` | Forecast MTBF from Duane model |
| `rdt` | `rdt()` | Reliability Demonstration Test planning |
| `gof_rga` | `gof()` | Goodness-of-fit statistics (CvM, K-S) |

## Example Session

With the MCP server running, you can ask Claude questions like:

> "I have failure data with times [100, 200, 300, 400, 500] and failure counts
> [1, 2, 1, 3, 2]. Fit a Crow-AMSAA reliability growth model and forecast the
> cumulative failures at 1000 and 2000 hours."

Claude will call `rga` and `predict_rga` on your behalf and return the results
in plain language.

> "Plan a reliability demonstration test for 90% reliability at 500 hours with
> 90% confidence, using a Weibull model with beta = 1.5 and 10 test units."

Claude will call `rdt` and explain the required test duration.

## Security Considerations

The MCP server runs R code in your local R session. Only share the server
endpoint with trusted clients. For multi-user deployments, consider running the
server in a sandboxed environment.