LLMaps

A Python library for creating interactive web maps, optimized for LLM-assisted development.

Encapsulates best practices for interactive web map development behind a predictable, composable API — so both you and your LLM produce correct code on the first try. Outputs a single HTML file powered by MapLibre GL JS.

Table of Contents

• Features
• Installation
• Quick Start
• Getting Started with Compass
• Examples
• Built With
• Comparison with Alternatives
• Architecture
• Documentation
• Tile Providers
• License
• Contact
• Contributing

Features

• Declarative API — composable Map + layers + sources + components. Describe what the map should show; the library handles how.
• Single HTML output — standalone file with MapLibre GL JS; works via file:// when data is embedded.
• Embedded mode — inline GeoJSON (optionally Geobuf + Gzip) so you can share one file.
• Comparison mode — before/after slider using the MapLibre compare plugin.
• Feature-state expressions — GPU-efficient dynamic styling (fill_color, opacity) via setFeatureState.
• Extensible — custom JS/CSS/HTML injection, embed_data() for arbitrary JSON.
• Built-in attribution — automatic LLMaps branding with GitHub link (customizable or removable).
• LLM-friendly — predictable names, stable contracts; use get_llm_context() for a compact reference in chat.

Available building blocks:

Category	Classes
Layers	CircleLayer, FillLayer, H3Layer, VectorTileLayer, SymbolLayer
Sources	FileSource, ApiSource, VectorTileSource
Components	Legend, Popup, Sidebar, Search, FeatureSearch, Controls, BasemapSwitcher, Storytelling

Installation

pip install llmaps

Cursor / Claude Code — both support the Agent Skills standard. Copy the repo's cursor-skill folder so the AI has full context when generating map code:

Tool	Skills path
Cursor	~/.cursor/skills/llmaps/SKILL.md
Claude Code	~/.claude/skills/llmaps/SKILL.md

On Windows use %USERPROFILE%\ instead of ~/. Create the folders if needed. The skill is picked up automatically on next launch. If you use either tool with this skill, you can skip the step below.

Context for other LLM chats: Run once in your project directory, then attach llmaps_context.md to the chat (e.g. @llmaps_context.md in Cursor):

python -c "import llmaps; open('llmaps_context.md','w').write(llmaps.get_llm_context())"

Quick Start

from llmaps import Map
from llmaps.layers import CircleLayer
from llmaps.sources import FileSource
from llmaps.components import Legend, Popup, Controls

source = FileSource(id="points", path="data/points.geojson")
layer = CircleLayer(
    id="points-layer",
    source=source,
    radius=6,
    color="#3182bd",
    opacity=0.8,
)

m = Map(center=[10.0, 50.0], zoom=4, title="My Map", tiles="osm")
m.add_layer(layer)
m.add_component(Legend(layer_labels={"points-layer": "Points"}))
m.add_component(Popup(fields=["name", "value"], field_labels={"name": "Name", "value": "Value"}))
m.add_component(Controls(zoom=True, scale=True))

m.auto_extent()
m.save("my_map.html")

Getting Started with Compass

Compass is an AI-first workflow that helps you go from raw data to a working build_map.py with minimal prompting.

It is documentation-only and uses existing LLMaps APIs.

What Compass Does

1. Surveys your dataset (geometry, fields, stats, bbox).
2. Asks a few focused questions about intent and style.
3. Chooses a recipe and generates runnable map code.
4. Refines colors, components, and interactions on request.

Where It Lives

• compass/README.md — overview and workflow
• compass/decision-tree.md — recipe selection logic
• compass/question-bank.md — adaptive interview prompts
• compass/recipes/ — build_map.py templates with placeholders

Activation Modes

• Implicit: ask naturally, for example Create an interactive map from examples/real-world/cafes/data/paris_cafes.geojson.
• Explicit: use /llmaps.compass, /llmaps.compass-survey, or /llmaps.compass-refine in Cursor.

Examples

Explore interactive maps showcasing key features and capabilities: LLMaps Examples Gallery

Real-World Examples

Example	Features	Data
Paris Cafes & Restaurants	CircleLayer, Search, FeatureSearch, BasemapSwitcher, Popup, Sidebar	OpenStreetMap Overpass
World Population	FillLayer, feature-state styling, Jenks classification, color ramps	Natural Earth
Earthquakes	CircleLayer, interpolate expressions, Jenks classification, temporal filtering	USGS Earthquake API
Pennsylvania Gerrymandering Story	FillLayer, Storytelling, comparison mode, custom overlays	PA election and district data

Technical Examples

Example	Features	Data
Dashboard Primitives	Dashboard component, filter bridge events, lightweight integration pattern	Small embedded GeoJSON

Built With

• MapLibre GL JS — frontend map rendering
• Jinja2 — HTML template engine
• Optional: H3, GeoPandas, Geobuf

Comparison with Alternatives

Criterion	Kepler.gl	Folium / ipyleaflet	Custom MapLibre/Leaflet	LLMaps
Ready-made components	⚠️ Limited by UI	⚠️ Few primitives	❌	✅ Full set: layers, legend, popup, sidebar, search, controls
LLM-friendly	❌	⚠️ Partial	⚠️ Depends on code	✅ Clear API, stable contracts
H3 / aggregation	✅	❌	⚠️ Manual	✅ H3Layer
Embedded (file://)	❌	⚠️ Often needs server	⚠️ Manual	✅ embedded=True
Comparison (before/after)	❌	❌	⚠️ Manual	✅ enable_comparison
Single HTML output	❌	✅ Possible	⚠️ Manual	✅ save / to_html
Customization	⚠️ Limited	✅ Good	✅ Full	✅ Full (config + templates + custom JS/CSS)
Extensibility	⚠️ Limited by UI	✅ Good (plugins, custom HTML/JS)	✅ Full	✅ Full (custom JS/CSS/HTML, embed_data(), templates)
No backend	❌	✅ Often	⚠️ Possible	✅ Embedded mode

When to use LLMaps: You want a single Python API to produce a standalone interactive map (especially with embedded data), with minimal boilerplate and good support for LLM-generated code.

When to choose something else: You need a full GIS UI (Kepler.gl), tight Jupyter-only integration (Folium/ipyleaflet), or a fully custom frontend stack (raw MapLibre/Leaflet with your own backend).

Architecture

Python API  →  Config (to_dict())  →  HTML (Jinja2)  →  Frontend (MapLibre GL JS + plugins)

Everything reduces to a serializable dict — no framework magic. The generator turns that config into one HTML file with inline or linked JS/CSS. See PHILOSOPHY.md for design principles and rationale.

Documentation

• LLM context — run python -c "import llmaps; open('llmaps_context.md','w').write(llmaps.get_llm_context())" and use @llmaps_context.md in chat for a compact API reference.
• PHILOSOPHY.md — Concept, design principles, comparison with alternatives.
• CONSTITUTION.md — Development principles for meaningful public API changes.
• specs/README.md — Lightweight spec-driven workflow and templates for planning public features.
• docs/api/ — Map, layers, sources, components (parameters and examples).
• docs/recipes/ — Heatmap, comparison, embedded map, feature-state highlighting.

Tile Providers

Use the tiles argument when creating the map:

Key	Provider	Attribution
"osm"	OpenStreetMap (default)	© OpenStreetMap contributors
"carto-light"	Carto Light	© OpenStreetMap contributors, © CARTO
"carto-dark"	Carto Dark	© OpenStreetMap contributors, © CARTO
"yandex"	Yandex Maps	© Yandex Maps
"2gis"	2GIS	© 2GIS

License

MIT. See LICENSE.

Contact

Sergey Abramov

Contributing

See CONTRIBUTING.md.

For meaningful public API changes, use the lightweight spec workflow in specs/README.md before implementation starts.