AtlasPI is the world's first MCP server for historical geography. 1,038 empires, 4,819 academic sources, real borders in GeoJSON — your agent stops hallucinating about history and starts citing sources.
The embed below is the same production instance running on atlaspi.cra-srl.com.
Move the timeline between 4500 BCE and 2024, click on an empire to see academic
sources, confidence scores, and ETHICS notes.
prefers-color-scheme.AtlasPI exposes 119 REST endpoints under OpenAPI 3.1. Open CORS, rate limit 60 req/min per IP, zero registration. Your agent sends an HTTP GET and receives structured JSON — entities active in a given year, events, trade routes, GeoJSON borders.
$ curl https://atlaspi.cra-srl.com/v1/snapshot/year/1250 { "year": 1250, "active_entities": 247, "highlights": [ { "name_native": "Imperium Mongolicum", "name_en": "Mongol Empire", "confidence": 0.92, "sources": 14 }, { "name_native": "Sacrum Imperium Romanum", "confidence": 0.88, "sources": 22 } ], "events": [/* 18 events */], "trade_routes": [/* Silk Road, Trans-Saharan ... */], "sources_total": 4819 }
→ https://atlaspi.cra-srl.com/docs # Try every endpoint from the browser, inspect the schemas
Anthropic's Model Context Protocol (MCP) is the emerging standard for giving
an LLM access to external tools and data. atlaspi-mcp is an MCP server
installable from PyPI: one line in claude_desktop_config.json
and Claude Desktop can query the dataset as if it were its own memory.
snapshot, compare, similar, on_this_day, routes_active...atlaspi-mcp v0.9.0: just pip install.$ pip install atlaspi-mcp Successfully installed atlaspi-mcp-0.9.0
{
"mcpServers": {
"atlaspi": {
"command": "atlaspi-mcp"
}
}
}
▸ 34 AtlasPI tools available ▸ "Which empires controlled the Balkans in 1400?" → call: snapshot_year(1400, bbox=Balkans) → 7 entities, 12 sources, avg confidence 0.83
Existing historical datasets for LLMs mirror the bias of English-language Wikipedia: "colonial era" instead of "genocide", European names instead of native ones, certainty where there should be uncertainty. AtlasPI codifies transparency in the data model, not in a footnote.
The name in the original language is the primary field. The English exonym is secondary. Same for Tawantinsuyu (not "Inca"), Bharat (not "India" pre-1947), Bizantion.
12 explicit ETHICS records covering Armenia, the Atlantic Maafa, Holocaust, Rwanda, Bangladesh 1971. Categories follow academic consensus, not the ones that sell textbooks.
confidence: 0.4For pre-1800 borders the confidence_score is low and visible. The AI agent receives "confidence 0.4" as a structured field and can decide whether to cite cautiously or look elsewhere.
known_silence flag on records where sources were destroyed (Indus Valley script, Mayan codices).confidence_score 0.0 → 1.0 explicit for every geographic and temporal claim.A unified relational model — entities, events, rulers, cities, routes, languages, archaeological sites — linked together and each anchored to academic sources. The backend is 24,609 lines of Python, the frontend 18,424, test coverage 1,206 functions.
The same dataset, exposed in three different ways depending on what you're building.
You're building an assistant with Claude, Cursor, AutoGen, LangChain. Install atlaspi-mcp and your agent stops inventing dates and borders. Free, no API key, open CORS.
You need to cite a stable dataset. AtlasPI has a Zenodo DOI (10.5281/zenodo.19581784), a public GitHub repo, and auditable methodology. Fork it, challenge it, contribute.
You want an interactive historical map for your edtech platform, but with your branding and datasets extended for the region you care about. White-label hosted instance + custom extensions.
Stack chosen to last: PostGIS for real geography, FastAPI for speed, Redis for cache, Docker compose for deployment. All open-source Apache 2.0.
Permissive open-source, commercial use included. Fork, extend, redistribute.
Zero friction for prototypes and research. Registration will only arrive for serious quotas.
Callable directly from the browser. Ideal for Jupyter notebooks and front-end prototypes.
Real geography, not bounding boxes. Spatial queries in SQL: ST_Intersects, ST_Area.
Async Python, type hints, auto-generated OpenAPI. Average latency < 60ms for simple queries.
Hot path cached, year snapshots pre-computed. The timeline slider stays smooth at 60 fps.
Self-host in 5 minutes: docker compose up and you have your own instance.
Complete schema, client SDKs generable in any language. Live Swagger UI.
Error tracking ready to enable via env var. Structured JSON logging by default.
If you run an educational publisher, an edtech platform, a cultural newsroom, or you've won a digital humanities grant — let's talk. AtlasPI is open-source, but CRA SRL offers managed hosting, custom extensions (specific regions, periods, languages), and tailored MCP integration.