root-ext-viz

Data-visualization MCP tool server for ROOT Workspace (and any MCP host). Five pillars:

1. Sessional kernel — data_load / py_run / session_state / session_reset: dataframes, imports, and variables survive across tool calls (pandas, numpy, scipy, matplotlib, plotly, altair). Resource-capped child interpreter; a wall-timeout names its state loss, never hides it.

2. Charts as source documents — chart_spec writes the Vega-Lite .vl.json AND renders its PNG beside it. The workspace renders the spec live (bundled vega, same v6 major as our vl-convert), the user keeps editing it; pixels are derived.

3. The vision loop — every render returns its path; the model reads the PNG back through the workspace's image-read path, looks, and refines.

4. Recipes, not 25 tools — viz_recipes: line/scatter/bar/ histogram/cdf/log_axes/error_bars/tolerance_band/control_chart/ heatmap (Vega-Lite) + bode/fft/contour (python).

5. Bounded data plumbing — data_profile first; inlined chart data capped; aggregate in the kernel, chart the summary.

Charter, binding and tested: compute and plot; outputs land only in derived or artifact directories, never on a source. Nothing reaches the network — rendering is local vl-convert (no node).

Develop

uv venv && uv pip install -e ".[dev]"
.venv/bin/pytest

Related MCP server: mcp-plots

Package (signed .rootx, ADR-0018)

uv build
.venv/bin/python tools/build_pack.py   # --unsigned for a test build