Seshat — Echoes of History

Ask counterfactual questions of the historical record. Seshat turns the Seshat Global History Databank into something you can interrogate — both programmatically through an MCP server and visually through a web app — and then project what might have happened if history had gone differently.

"What if the Roman Empire had developed gunpowder weapons in 200 BCE? How might its social complexity have evolved?"

This monorepo has two front doors:

• 🔌 MCP server — exposes 9 Seshat tools to Claude and other MCP-enabled assistants, so researchers can query 10,000 years of social-complexity data in natural language. See packages/mcp-server/README.md.

• 🌍 Echoes of History web app — a Next.js app for picking a civilisation, injecting a hypothetical change, and watching a forward-projected complexity trajectory with confidence bands.

Both are powered by the same estimation engine, which lives in the shared package.

About the Seshat Global History Databank

This project is built on top of Seshat — it does not generate historical data, it makes Seshat's data queryable and projectable.

Seshat (named after the ancient Egyptian goddess of wisdom and record-keeping) is a research databank founded in 2011 that systematically documents the social and political organisation of human societies across history. It calls itself "the most current and comprehensive body of knowledge about human history," structured so that the past can be studied with established scientific techniques. The databank is maintained by an international team of historians, social scientists, and data scientists, and underpins peer-reviewed research into the long-run evolution of social complexity.

Its scope is what makes counterfactual projection possible:

• 864 polities (political societies) across 47 regions grouped into 10 macro-regions spanning Africa, Europe, Asia, the Americas, and Oceania

• Coverage from roughly 3550 BCE to the modern era

• 77 social-complexity variables (territory, population, settlement hierarchies, infrastructure, information systems, money) — the basis for the PC1 composite complexity score

• 49 warfare variables (military technology, fortifications, conflict patterns)

• 26 general variables (capitals, languages, religions, degree of centralisation)

Learn more or explore the source data at seshat-db.com. See Data provenance below for the specific datasets and licensing this project relies on.

Repository structure

This is a pnpm workspace monorepo:

Quickstart

git clone https://github.com/mosse/seshat-mcp
cd seshat-mcp
pnpm install

# The shared package MUST be built before anything else
pnpm --filter @seshat/shared build

You'll need a Supabase project loaded with Seshat data (run the migrations in data/schemas and the ETL in data/etl). Copy the .env.example files in packages/mcp-server/ and packages/web/ and fill in:

Run the MCP server

pnpm --filter @seshat/shared build
pnpm --filter @seshat/mcp-server build
pnpm --filter @seshat/mcp-server dev

Then point Claude Desktop at it — see the MCP server README for config.

Run the web app

pnpm --filter @seshat/shared build
pnpm --filter web dev

The 9 MCP tools

Full parameter tables are in the MCP server README.

Testing

# MCP server (vitest) — 56 tests
pnpm --filter @seshat/mcp-server test

# Web app end-to-end (Playwright) — 34 tests, Desktop Chrome + mobile
pnpm --filter web exec playwright test

# ETL (pytest) — 75 tests, requires venv
cd data/etl && source .venv/bin/activate && pytest tests/

How estimates work — and their limits

The counterfactual engine uses an illustrative approximation of the dynamic regression model from Turchin et al. (2022), Disentangling the evolutionary drivers of social complexity (Science Advances), with Monte Carlo sampling over the model's residual noise to produce confidence bands.

⚠️ The current coefficients are approximate placeholders, not the published fit. Sourcing/fitting the real model (and other honesty fixes) is tracked in IMPROVEMENT_PLAN.md. Until then, read projections as directional illustrations, not authoritative estimates.

A few things to keep in mind when reading any projection:

• Uncertainty is a feature. Every projection ships with confidence bands and an explicit confidence_limits section. The bands capture statistical noise within the model — not the full range of historical contingency. Treat the central line as one illustrative path, not a prediction.

• Eurasian validation. The regression model was validated on Eurasian data. Projections for the Americas and Oceania-Pacific carry a prominent caveat and should be read with extra caution.

• Year convention. BCE years are negative integers (year 0 = 1 BCE, astronomical). They're displayed as "500 BCE", never "-500".

Data provenance

• Equinox-2020 — Seshat Global History Databank (Turchin et al., 2015). CC BY-NC-SA 4.0. Zenodo

• Cliopatria — Seshat geospatial boundaries. CC BY 4.0. GitHub

• Counterfactual model — Turchin et al. (2022). Disentangling the evolutionary drivers of social complexity. Science Advances.

License

Code is MIT. The underlying Seshat data is CC BY-NC-SA 4.0 — any use of the data must be non-commercial and attributed.