dartwork-mpl

Publication-quality matplotlib — a thin utility layer, not a wrapper.

dartwork-mpl keeps Figure / Axes 100% native and adds the parts matplotlib makes tedious: a physical-width geometry API, curated style presets, an OKLCH-aware color system, deterministic content-aware layout, visual validation, and a first-class integration for AI coding assistants (an MCP server + a bundled prompt corpus). You never learn a new plotting API — you keep writing matplotlib, just without the friction.

import matplotlib.pyplot as plt
import dartwork_mpl as dm

dm.style.use("scientific")                                       # 1. curated preset
fig, ax = plt.subplots(figsize=dm.figsize("13cm", "standard"))   # 2. physical width × aspect
ax.plot(x, y, color="oc.blue5", lw=dm.lw(0))
ax.set_xlabel("Time [s]")
dm.simple_layout(fig)                                            # 3. content-aware margins
dm.save_formats(fig, "figure", formats=("svg", "png", "pdf"))    # 4. multi-format save

That four-step pattern — preset → figsize(width, aspect) → simple_layout → save_formats — is the whole workflow. No tight_layout(), no hand-tuned figsize=(w, h) arithmetic, no dpi= guesswork.

Installation

pip install dartwork-mpl        # or:  uv add dartwork-mpl

Requires Python 3.10+. The core install is intentionally lean; add an extra only when you need it:

pip install "dartwork-mpl[notebook]"   # or [mcp], [ui]

Highlights

• Geometry, decoupled from inches. dm.figsize("13cm", "standard") takes a physical width (cm / in / mm / pt, or dm.col1 = 9 cm / dm.col2 = 17 cm) and one of ten aspect tokens (square / portrait / tall / standard / golden / wide / a4 / slide / cinema / panoramic); the height follows. Bare numbers are rejected so the unit is always explicit.
• Deterministic layout. simple_layout(fig) measures every visible artist and places the GridSpec arithmetically — reproducible across machines, unlike tight_layout()'s heuristics. margin="2%" (or dm.mm(2)) adds a buffer.
• OKLCH-aware color. Named palettes (oc.* Open Color, tw.* Tailwind, md.*, ad.*, cu.*, pr.*) plus a Color class spanning OKLab / OKLCH / RGB / hex with perceptual interpolation (cspace) and gamut-correct mapping.
• Curated styling. Seven presets (scientific, report, presentation, …), each with a Korean -kr variant, and preset-relative scaling helpers fs / fw / lw so literals never drift when you switch themes.
• Validation & export. validate_figure(fig) flags overflow, text/legend overlap, tick crowding, and empty axes — invisible failures in headless agent pipelines. save_formats(fig, ...) writes SVG / PNG / PDF / EPS at once, with deterministic SVG/PDF/SVGZ output for unchanged figures.
• AI-native. A bundled MCP server exposes lint + auto-fix, figure validation, color lookup, and the live policy corpus to Claude Code / Cursor / Windsurf. No-MCP agents read the same corpus from disk.
• Batteries included. Material Design Icons + Font Awesome 6 fonts, ready-made plot templates (plot_diverging_bar, …), and a FastAPI viewer for live tuning.

Core API at a glance

import dartwork_mpl as dm

# ── Geometry ──────────────────────────────────────────────────────────
dm.figsize("13cm", "wide")          # width × aspect token  → inch tuple
dm.figsize("13cm", 0.6)             # ...or a numeric ratio / "8cm" / dm.cm(8)
dm.cm(13); dm.mm(170); dm.inch(4.6); dm.pt(24)   # Length values
dm.col1; dm.col2                    # 9 cm / 17 cm academic-column sugar

# ── Styling & scaling ─────────────────────────────────────────────────
dm.style.use("scientific")          # apply a preset
dm.style.stack(["base", "font-scientific", "lang-kr"])   # compose
dm.fs(2); dm.fw(1); dm.lw(-0.3)     # preset-relative font size / weight / line width

# ── Color ─────────────────────────────────────────────────────────────
ax.plot(x, y, color="oc.blue5")     # named palettes register with matplotlib
dm.color("oc.blue5")                # parse name / "#4285F4" / "rgb(...)" / "oklch(...)"
dm.oklch(0.7, 0.15, 150); dm.rgb(66, 133, 244); dm.hex("#4285F4")
dm.cspace("#FF0000", "#0000FF", n=5, space="oklch")      # perceptual interpolation
dm.mix_colors("oc.blue5", "white", alpha=0.35)

# ── Layout & annotation ───────────────────────────────────────────────
dm.simple_layout(fig)               # deterministic content-aware margins
dm.simple_layout(fig, margin="2%", gs=gs)   # buffer + target a GridSpec
dm.label_axes(axes)                 # (a) (b) (c) panel labels
dm.arrow_axis(ax, "x", "Cost")      # Low ◄── Cost ──► High

# ── Validate, export, icons ───────────────────────────────────────────
dm.validate_figure(fig)             # overflow / overlap / tick-crowding / empty
dm.save_formats(fig, "fig", formats=("png", "svg", "pdf"), dpi=300)
mdi = dm.icon_font("mdi")           # also "fa-solid" / "fa-regular" / "fa-brands"

# ── Plot templates ────────────────────────────────────────────────────
from dartwork_mpl.templates import plot_diverging_bar
fig, ax = plot_diverging_bar(labels=["A", "B"], neg_values=[-30, -15], pos_values=[40, 55])

See the usage guide and API reference for the full surface.

Common pitfalls

Three patterns trip up new users (and AI assistants) more than any others. The built-in lint engine flags all three; dm.migrate_legacy_code rewrites them in place.

Full catalog: 02-anti-patterns.yaml. Reachable at runtime via dm.get_prompt("02-anti-patterns") or lint_dartwork_mpl_code(code) over MCP.

Upgrading from v4? See the Migration Guide for the palette codemod and removed-name map.

Style presets

Each has a Korean -kr variant (scientific-kr, report-kr, …) with Korean-aware fonts. List them with dm.list_styles().

AI-assisted development (MCP)

dartwork-mpl ships a built-in Model Context Protocol server so AI coding assistants pull the current policy guides, color palettes, lint catalog, and helper tools straight into the chat — no copy-pasting docs. It exposes 16 tools (lint + auto-fix, figure validation, render, color lookup, info, chart-type recommender, layered-plot composer, advanced-tier render), 10 resources + 4 resource templates (the prompt corpus + 18 basic + 18 tier-2 advanced plot templates), and 2 prompts.

pip install "dartwork-mpl[mcp]"     # installs fastmcp + httpx; adds the dartwork-mpl-mcp script

Point your client at the dartwork-mpl-mcp console script — e.g. Claude Code (~/.claude.json) or Cursor (~/.cursor/mcp.json):

{
  "mcpServers": {
    "dartwork-mpl": { "command": "dartwork-mpl-mcp" }
  }
}

Restart the client and ask it to list its MCP resources to confirm. Windsurf, Antigravity, generic stdio setups, the full tool/resource catalog, and the local-clone variant are covered in docs/integrations/mcp_server.md.

No MCP? The same corpus is bundled in the wheel and reachable from Python — dm.get_agent_doc("llms-full") (also "AGENTS", "CLAUDE", "llms") returns the text, dm.agent_doc_path(name) its path. The repo-root CLAUDE.md / AGENTS.md / llms.txt (per the llmstxt.org spec) are the 30-second onboarding.

Documentation

📚 Full documentation

• Quickstart — install, first figure, save
• Usage guide — width/aspect, layout, color, patterns
• Color system — palettes, OKLCH, colormaps
• Example gallery — rendered, copy-pasteable
• API reference — every public function and class
• Design philosophy — why thin utilities, not a wrapper

Project layout

src/dartwork_mpl/
├── units.py / scale.py        # figsize, cm/mm/inch/pt, col1/col2 · fs/fw/lw
├── style.py                   # Style class + preset management
├── colors/                    # Color (OKLab/OKLCH/RGB/hex) + named palettes
├── layout.py / annotation.py  # simple_layout, label_axes, arrow_axis
├── validate.py / lint.py      # validate_figure · lint + migrate_legacy_code
├── io.py / formatting.py      # save_formats, show · format_axis_*
├── icon.py / font.py / cmap.py / diagnostics/   # fonts, colormaps, viz helpers
├── templates/ / helpers/      # plot templates · high-level composition helpers
├── agent.py / prompt.py       # bundled LLM corpus · prompt guides
├── mcp/                       # MCP server (server / resources / tools / prompts)
├── ui/                        # interactive FastAPI viewer
└── asset/                     # bundled styles, colors, fonts, icons, prompts

Contributing & issues

Bug reports and feature requests go to the GitHub issue tracker. Released under the MIT License.