ibcs-plotly

Interactive IBCS-compliant business charts on Plotly, with optional Streamlit components.

IBCS (International Business Communication Standards) fixes a small set of rules for business reports: current-period and comparison-period values get consistent colors, variance gets its own consistent colors, composition charts are sorted bars instead of pies, and comparable small multiples share one scale. This library gives you those chart types as plain functions that return a plotly.graph_objects.Figure, plus a thin Streamlit layer for KPI cards and the AC/PY/variance legend.

There's an existing IBCS-inspired Python package, clean_business_chart, built on matplotlib. This one targets interactive dashboards (Plotly + Streamlit) rather than static reports.

Install

pip install ibcs-plotly            # core: plotly + pandas + numpy
pip install ibcs-plotly[streamlit] # + Streamlit KPI cards / legend

Quick example

from ibcs_plotly import grouped_column, compare_periods

df_ac, df_py, py_year = compare_periods(df, period_col="year", current=2026)

fig = grouped_column(
    ac=df_ac.groupby("month")["revenue"].sum(),
    py=df_py.groupby("month")["revenue"].sum(),
    labels=month_names,
)
fig.show()

See examples/streamlit_demo.py for a full dashboard on synthetic data — run it with:

pip install -e ".[streamlit]"
streamlit run examples/streamlit_demo.py

Plan charts with AI (MCP)

ibcs_plotly.planner profiles a dataset, works out what can be calculated (variance AC-PY, variance %, share, rank, trend, rolling, YTD), and emits a plan of IBCS chart specs that a developer - or an agent - can build from. An MCP server exposes this to any MCP-capable assistant.

The operator decides how much of the data the AI may see; the AI can ask for less per call, never more:

pip install "ibcs-plotly[mcp]"
IBCS_MCP_DATA_DIR=./data IBCS_MCP_MAX_TIER=STATS ibcs-plotly-mcp

Tools: load_dataset, describe_dataset, plan_charts, llm_context_bundle, check_chart, render_chart. Charts are always rendered locally on the full data; the AI only gets back a file path. Without MCP, the same pieces work as a library:

from ibcs_plotly.planner import build_plan, render_spec
plan = build_plan(df, tier="STATS")
print(plan.to_markdown())
render_spec(df, plan.charts[0]).show()

Exposure preview

streamlit run examples/exposure_preview.py shows exactly what an assistant would receive from your data at each tier, before you connect anything. Upload a file (or use the built-in sample), pick a tier, mark extra sensitive columns, and read the JSON, the plan and the charts. The last tab prints the matching MCP server settings. Nothing leaves your machine.

At the SAMPLE tier the email column is masked in every row:

LLM-ready export and compliance check

from ibcs_plotly.planner import llm_context, estimate_tokens, to_json
ctx = llm_context(df, tier="STATS", goal="revenue by region")  # instructions + data view + plan
print(estimate_tokens(ctx))        # check it fits the prompt before pasting
prompt_payload = to_json(ctx)

from ibcs_plotly import check_figure
for f in check_figure(fig):        # works on any Plotly figure
    print(f.severity, f.rule, f.message)

check_figure flags pies, 3D, truncated bar axes, unshared small-multiple scales, wrong AC/PY/variance colors, too many categories and color noise. It checks only what is mechanically verifiable from the figure; it cannot judge titles, units or message quality.

Tested on real data

Every chart above and below is the planner's own output for a real public dataset (nothing hand-drawn); each image carries its source and licence.

What was tested

The test suites are kept out of this repository (unit tests, seed data, the downloaded datasets, the harnesses and their reports all live in git-ignored tests/ and testdata/), and were run locally before release. CI runs lint, an import check of every module, and an end-to-end smoke test of plan, render and lint.

Real data found problems synthetic data did not, all fixed and covered by unit tests: code columns such as zip, district and iucr read as measures or dates; invoice ids with cancellation prefixes turned into numbers and blanked; latitude and longitude summed; regional totals double counted; stray dates picked as the comparison year; a part-year compared with a full year; monthly lines for one month of data; street addresses missed as PII; ID-like columns charted as categories.

Not covered: charts are never judged for meaning (only for the rules listed above), PII detection is heuristic, and derived measures such as quantity times price are not built for you.

Charts built from CC BY datasets credit their sources in the image: Our World in Data, "CO2 and Greenhouse Gas Emissions" (CC BY 4.0); Chen, D. (2012), "Online Retail II", UCI Machine Learning Repository (CC BY 4.0).

What's included

• grouped_column, horizontal_bar, variance_waterfall, small_multiples, line_comparison — chart primitives, framework-agnostic

• compare_periods, safe_delta_pct — period-over-period comparison helpers

• IBCSTheme — colors/layout as a dataclass you can override

• ibcs_plotly.streamlit_ext — kpi_card, legend_html, inject_theme_css

• ibcs_plotly.planner — dataset profiling, exposure tiers, chart plans, spec renderer, LLM export

• ibcs_plotly.check_figure — IBCS compliance checker for any Plotly figure

• ibcs_plotly.mcp_server — MCP server (ibcs-plotly-mcp) for AI-assisted chart planning

Not in scope yet: matplotlib/other backends, PDF/PPT export.

License

MIT