# shewhart-mcp

An MCP server that exposes [shewhart](https://github.com/bertanucar/shewhart),
the statistical process control library for Python, as tools for AI agents.

The reasoning: agents asked about process data tend to generate their own
statistics, and generated statistics fail silently. Calculation belongs in a
deterministic, versioned, validated tool that the agent calls. The agent
interprets; validated code calculates.

## Status

In development. This release reserves the package name; the first working
server ships shortly, with these tools:

| Tool | Purpose |
|------|---------|
| `review` | one call: choose the right chart, check assumptions, return a structured verdict |
| `control_chart` | I-MR, Xbar-R/S, p/np/c/u, Laney, EWMA, CUSUM by alias |
| `capability` | Cp/Cpk/Pp/Ppk with confidence intervals, non-normal methods |
| `tolerance_interval` | normal (Howe k2) and nonparametric (Wilks) |
| `fit_baseline` / `evaluate_baseline` | freeze control limits, judge new data against them |

Every result carries provenance (library version, input hash, timestamp) and
is validated against published reference values, including NIST-certified
datasets, in CI.

## License

MIT. Built by [Bertan Ucar](https://github.com/bertanucar).