Fig Foundry

A WYSIWYG matplotlib figure composer. Load your data (.npz / .pkl / .csv), drag signals onto a free canvas to build plots, arrange and style them, and export clean, reproducible Python.

The preview renders by executing the generated code, so what you see is byte-for-byte what the exported .py produces. There is no second renderer to drift.

Fig Foundry is a standalone local app with no dependency on any particular project's data. It ships with a small neutral sample.npz (1-D signals plus a 2-D field) so a fresh install has something to open; load your own dataset to work with real data.

Install

uv tool install ./path/to/fig_editor     # installs a `figfoundry` command
# or, from a git repo:
uv tool install figfoundry --from "git+https://github.com/<you>/<repo>.git"

Then run:

figfoundry             # starts the web app on http://127.0.0.1:8037 and opens the browser
figfoundry serve --port 9000 --no-browser

Command line

The same engines are available as subcommands for scripting:

figfoundry list                              # list saved figures
figfoundry new --name my-fig                 # create a blank figure, print its slug
figfoundry export SLUG [--out PATH]          # write a figure's generated .py
figfoundry render SLUG --format svg|pdf|png  # render a figure to a file
figfoundry datasets                          # list available datasets

Data

A dataset is a named collection of signals (arrays). Supported formats:

• .npz — numpy zip of named arrays. Safe by construction.
• .pkl — Python pickle, loaded through a whitelist unpickler that only permits numpy arrays and plain containers (blocks os, subprocess, socket, ctypes, …). A malicious pickle cannot execute code.
• .csv — a table with a header row; each numeric column becomes a signal.

1-D and 2-D arrays are both supported (2-D for image/contour/vector plots). Load a dataset with the Load data… button in the toolbar; each canvas is tied to one dataset, and its signals populate the library.

Your figures, datasets, and exports live in ~/.figfoundry/.

What you can do

• Drag a signal onto the canvas to make a plot (or onto an existing plot to overlay a line / σ-band / error bars / replace).
• Plot types (23): line, step, scatter, grouped scatter (clustering), bar, bar (horizontal), histogram, box plot, violin plot, 2-D histogram, hexbin, image, contour (filled / lines), pcolormesh, vector field, streamplot, decision boundary, pie, polar, 3-D scatter, surface (3-D), wireframe (3-D).
• Move, resize (with snapping), and nudge (arrow keys) plots and elements.
• Multi-select (Shift+click) to nudge / move / delete several at once.
• Per-plot: data, color, line/marker, σ-band, error bars, ticks, labels, limits, spines, title, colormap + colorbar, and artists (vline, hline, text, annotate, arrow, box, circle, ellipse, polygon, span, legend).
• Suggest accessible colors — one click applies the colorblind-safe Okabe-Ito palette to a plot's series.
• LaTeX mathtext in labels, titles, and text (e.g. $x_0$).
• Layout: grid, align, distribute, margins, sync axis ranges.
• Figure elements: axis labels, legend, suptitle, free text.
• Name / duplicate / delete figures. Export the .py (copy or download), and the figure itself as PNG / SVG / PDF (SVG and PDF are vector, for publication).

Everything is also available through a small JSON API (/api/canvases, /api/datasets, /api/palettes, /api/canvases/{slug}/render|code|figure|download).

Testing

The test suite lives in tests/ (outside the package, so it is never shipped). It proves the WYSIWYG invariant (for every plot type), the pickle-safety whitelist, and dataset I/O (npz/pkl/csv), each against a throwaway data dir:

uv run --extra dev pytest -q