MCP Server

scitex-ml ships a Model Context Protocol (MCP) server so AI agents can compute metrics, generate reports, and project feature matrices autonomously — the same stateless analysis surface as the CLI.

Installation

pip install scitex-ml[mcp]

Starting the server

scitex-ml mcp start            # stdio transport
scitex-ml mcp doctor           # health check (deps, tool count)

The scitex umbrella also mounts this server automatically under the ml namespace, so the tools surface to agents as ml_*.

MCP host configuration

{
  "mcpServers": {
    "scitex-ml": {
      "command": "scitex-ml",
      "args": ["mcp", "start"]
    }
  }
}

Generate this snippet with scitex-ml mcp install.

Available tools

Tool (umbrella-mounted)	Description
`ml_compute_metrics`	Balanced accuracy, MCC, confusion matrix, ROC/PR-AUC from a predictions table
`ml_generate_report`	Full ClassificationReporter report (summary.json + plots) into a directory
`ml_reduce_dimensions`	PCA / UMAP 2-D projection figure from a feature matrix
`ml_skills_list` / `ml_skills_get`	Discover and fetch the skill pages shipped with scitex-ml

All tools take/return JSON. The analysis tools report input errors as {"success": false, "error": "..."} rather than raising.

Parity

Each tool maps 1:1 to a scitex-ml CLI verb with the same name and JSON shape (compute-metrics ↔ ml_compute_metrics). scitex-ml deliberately exposes only this stateless slice via MCP — training, optimizers and the deep submodule API stay Python-only — so it declares mcp_parity_exempt = true.

Diagnostics

scitex-ml mcp doctor
scitex-ml mcp list-tools -vv