Live now · v6.89.0 · 1,034 entities + 643 events + 55 periods + 110 cities + 41 trade routes + 94 dynasty chains

Reduce LLM hallucinations 3× on historical queries

Tool-augmented retrieval for AI agents working with historical geography. 1,034 entities with capital timelines, rulers, events, and 2,400+ academic sources. Empirically validated: hallucinations from 0.21 → 0.07 per query (benchmark v7.0).

✅ Free public API + MCP server — no login, no API key, no registration. Apache 2.0. CORS-enabled. Try it: curl https://atlaspi.cra-srl.com/v1/snapshot/year/1250

Try the live map View API docs About FAQ

GET /v1/snapshot/1500 · JSON · GeoJSON · OpenAPI 3.1 · 34 MCP tools · Apache-2.0

Empirical validation — v7.0 benchmark, 2026-04-23

Tool calls reduce hallucinations by 67% on hard historical queries

−67%

Hallucinations
(baseline 0.21 → AtlasPI 0.07 per query)

+4.6pp

Tool effect
(accuracy on trap-designed queries)

8 MCP tools

search, snapshots,
rulers, events, languages

Methodology: A/B test on Claude Sonnet 4.5, judged by Claude Opus 4.5 (cross-vendor with GPT-4o pending).
Bank v2: 30 questions designed to trap LLM baseline (capital anachronisms, dynasty conflations, ETHICS-aware atrocity questions).
Full report · ADR-007 · Reproducible code

Why AtlasPI

Built for machines first.
Honest with humans.

Structured for agents

Predictable JSON responses, OpenAPI 3.1 spec, and a native MCP server. Your agent stops scraping and starts reasoning.

Academically sourced

Over 2,400 academic citations. Every record carries a confidence score and a status: confirmed, uncertain, or disputed.

Ethically designed

Contested territories show every version. Colonial names coexist with original ones. We document history — we don't arbitrate it.

See it live

Drag the year. Watch civilizations breathe.

A 5,000-year timeline rendered from real boundary data. No mockup.

Model Context Protocol

Native MCP server.
Plug it into Claude.

AtlasPI ships as an MCP server. Add it to your Claude config and your AI agent can query historical geographic data natively — no custom integration, no glue code, no scraping.

Works with Claude Desktop, Claude Code, and any MCP client
34 tools: snapshots, events, cities, routes, chains, comparisons, similarity, date coverage, fuzzy search
Returns structured JSON with sources and confidence scores intact
Open source under Apache-2.0 — fork, audit, self-host

View MCP source REST API docs

claude_desktop_config.json

{
  "mcpServers": {
    "atlaspi": {
      "command": "python",
      "args": ["-m", "atlaspi_mcp"]
    }
  }
}

Restart Claude. Ask: "What empires existed in Anatolia in 1200 BCE?"

REST API

Three lines to query 5,000 years.

50+ endpoints. JSON in, JSON out. No auth required for the open tier.

snapshot.sh

# Get every entity that existed in the year 1500
curl https://atlaspi.cra-srl.com/v1/snapshot/1500 | jq '.summary'

# Compare two empires side by side
curl https://atlaspi.cra-srl.com/v1/compare/1/2

# Find what existed near a coordinate at a given year
curl "https://atlaspi.cra-srl.com/v1/nearby?lat=41.9&lon=12.5&year=100"

atlaspi_demo.py

import requests

# What existed near Rome in the year 100 CE?
r = requests.get(
    "https://atlaspi.cra-srl.com/v1/nearby",
    params={"lat": 41.9, "lon": 12.5, "year": 100},
)

for entity in r.json()["results"]:
    print(entity["name"], entity["confidence_score"])

atlaspi-demo.js

// Compare two historical entities by ID
const r = await fetch(
  'https://atlaspi.cra-srl.com/v1/compare/1/2'
);
const data = await r.json();

console.log(data.entity_a.name, 'vs', data.entity_b.name);
console.log('Overlap years:', data.temporal_overlap);

Full reference at /docs (Swagger UI) and /redoc.

Who is it for

Built for builders who care about facts.

AI Agents

Historical reasoning without hallucination. Cite sources, surface confidence, refuse to invent borders that never existed.

Digital Humanities

Timeline analysis, cross-cultural research, reproducible queries. Export GeoJSON or CSV with full citation chains intact.

EdTech

Drop the interactive map into your learning platform with a single iframe. Time slider included. No license fees.

Game Design

Realistic historical map generation, faction boundaries by era, period-accurate place names in their original languages.

1,034

historical entities

643

historical events

105

historical rulers

1,249

archaeological sites

6,500+

years covered

50+

REST endpoints

MCP tools

951

tests passing

For AI agent developers

Why your agent can't just scrape Wikipedia.

An agent reasoning about history is an agent reasoning under deep uncertainty. Sources contradict. Borders shift mid-sentence. The same city has four names depending on who you ask and when. Web scraping flattens all of this into prose that looks authoritative — and that is exactly how hallucinations enter pipelines that should know better.

AtlasPI returns structured records. Each one carries an explicit confidence_score, a status field, and a sources[] array of academic citations. When the data is contested, the API says so. When a boundary is approximated, the metadata flags it. Your agent can finally hedge correctly, cite properly, and refuse to fabricate the borders of an empire that ended 800 years before the year you asked about.

The same information is also available via the MCP server, so an agent running in Claude Desktop or Claude Code can call AtlasPI tools directly — no custom function definitions required.

Read the API spec Star on GitHub

Open core

Free forever. Premium when you scale.

Open source

Core

$0/month

Full REST API access
MCP server included
Self-host from GitHub
Apache-2.0 licensed
1,034 entities + 643 events

Get on GitHub

Coming soon

Premium

Custom

High-volume API access
Curated regional datasets
Enterprise SLAs
Hosted MCP infrastructure
Priority support

Get notified