History Lab MCP

Give AI assistants like Claude direct access to ~5M declassified historical documents — CIA, State Department cables, FRUS diplomatic records, World Bank, NATO, UK Cabinet Papers, and more.

Data from the FOI Archive by History Lab at Columbia University.

This project ships three ways to use the same backend:

1. Remote MCP server (recommended) — hosted at mcp.declassification-engine.org, ready to plug into Claude or any MCP-compatible client.
2. Local MCP server — run it over stdio for development or self-hosting.
3. CLI — @history-lab/cli on npm, for shell scripting and ad-hoc lookups.

---

MCP Server

The MCP server exposes 9 tools for searching, retrieving, and analyzing declassified documents. It's a Model Context Protocol server using the Streamable HTTP transport.

Endpoint: https://mcp.declassification-engine.org/mcp Health check: https://mcp.declassification-engine.org/health Auth: none required — the server is public.

Connect from Claude Code

claude mcp add --transport http history-lab https://mcp.declassification-engine.org/mcp

Then in any Claude Code session, the history-lab tools become available. Try:

Search for CIA documents about the Cuban Missile Crisis and summarize the top three.

Connect from Claude Desktop

Edit your claude_desktop_config.json (Settings → Developer → Edit Config):

{
  "mcpServers": {
    "history-lab": {
      "url": "https://mcp.declassification-engine.org/mcp"
    }
  }
}

Restart Claude Desktop. The history-lab server should appear under the tools menu.

If your client doesn't support Streamable HTTP transport directly, use the mcp-remote bridge:

{
  "mcpServers": {
    "history-lab": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://mcp.declassification-engine.org/mcp"]
    }
  }
}

Verify the server is up

# Health check
curl https://mcp.declassification-engine.org/health
# -> ok

# List available tools
curl -s -X POST https://mcp.declassification-engine.org/mcp \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list","params":{}}'

MCP Tools

Tool	What it does
vector_search	Semantic search across all ~5M documents using natural language. Returns ranked chunks with relevance scores. Best for conceptual queries.
corpus_search	Full-text search (PostgreSQL FTS) across documents. Filter by corpus, classification, date range, title.
frus_search	Search the Foreign Relations of the United States collection (312K docs, 1620–1989) with sender/recipient/location filters.
get_document	Fetch full text + metadata for a specific document by doc_id or vector-search r2_key. Optional entity/topic enrichment.
list_corpora	List all collections with document counts, date ranges, word counts.
entity_lookup	Find named entities (people, places, orgs) by name. Returns entity IDs + Wikidata links.
entity_documents	Get all documents associated with a specific entity ID.
browse_topics	Explore topic models (keyword clusters) per corpus, or fetch top docs for a topic.
archive_stats	Aggregate stats: totals, breakdowns by decade or classification.

Each tool ships with detailed descriptions and performance hints embedded in its schema — your AI client will see warnings like "always provide a corpus filter for fast results" and adjust queries accordingly.

Example prompts

Once connected, try asking your AI assistant:

• "Find diplomatic cables about Vietnam between 1968 and 1975. Summarize themes."
• "What FRUS documents did Kissinger send to Nixon in 1973?"
• "Look up the entity 'Pinochet' and pull its top documents."
• "Browse topics in the CIA corpus and show me the top documents for the most relevant Cold War topic."
• "How many declassified documents are in the archive, broken down by decade?"

---

Self-hosting the MCP server

Run locally over stdio

For development or to integrate with stdio-based MCP clients:

git clone https://github.com/history-lab/history-lab-mcp
cd history-lab-mcp
npm install
npm run dev:mcp

Then point a stdio client at tsx bin/mcp.ts:

{
  "mcpServers": {
    "history-lab-local": {
      "command": "tsx",
      "args": ["/absolute/path/to/history-lab-mcp/bin/mcp.ts"]
    }
  }
}

Deploy your own Cloudflare Worker

The hosted server is a Cloudflare Worker (src/worker/index.ts). To deploy your own copy:

npx wrangler deploy

See wrangler.jsonc for the Worker config and docs/DEPLOYMENT.md for full deployment notes.

Configuration

The server works out of the box with the public corpus API key. Override via environment variables (or Cloudflare Worker secrets):

Variable	Description
HISTORYLAB_CORPUS_API_URL	Override the PostgREST corpus API base URL
HISTORYLAB_VECTOR_API_URL	Override the vector search API URL
HISTORYLAB_VECTOR_API_KEY	Override the vector API key
CLIENT_API_KEY	(Worker only) Require Authorization: Bearer <key> on incoming MCP requests

---

CLI

Same tools, exposed as a shell binary.

Install

npm install -g @history-lab/cli

Or run directly:

npx @history-lab/cli search "cuban missile crisis"

Quick examples

# Semantic vector search
history-lab search "nuclear weapons testing in the Pacific"

# Full-text search in CIA documents
history-lab corpus-search --query "cuban missile crisis" --corpus cia

# Kissinger to Nixon in 1973
history-lab frus-search --sender kissinger --recipient nixon --from 1973-01-01 --to 1974-01-01

# Entity lookup + drill-down
history-lab entities "Castro"

# Full document text
history-lab document CIA-RDP79T00429A001400010019-1

# Archive stats
history-lab stats --decades

Commands

Command	Purpose
search <query>	Semantic vector search
corpus-search	Full-text search with filters
frus-search	FRUS-specific search (sender/recipient/location)
document <id>	Fetch a document by ID
corpora	List collections
entities <query>	Find entities by name
entity-docs <id>	Documents for an entity
topics <corpus>	Topic model exploration
stats	Archive statistics

Every command supports --json (for piping into jq) and --markdown output. Run history-lab <command> --help for full options.

---

Available Corpora

ID	Collection	Documents
cia	CIA CREST	~936K
cables	State Department Cables	~718K
cfpf	Central Foreign Policy Files	~3.2M
frus	Foreign Relations of the US	~312K
clinton	Clinton Emails	~33K
kissinger	Kissinger Transcripts	~3K
briefing	Presidential Daily Briefs	~12K
nato	NATO Archives	~2K
un	UN Archives	~5K
worldbank	World Bank Archives	~9K
cabinet	UK Cabinet Papers	~3K
cpdoc	CPDOC (Brazil)	varies

---

Architecture

src/core/       All business logic (clients, types, formatters, config)
src/cli/        CLI adapter (commander.js, thin wrapper over core)
src/mcp/        MCP server adapter (tool definitions, thin wrapper over core)
src/worker/     Cloudflare Worker entrypoint (Streamable HTTP transport)
bin/cli.ts      CLI entrypoint (ships to npm)
bin/mcp.ts      Local MCP stdio entrypoint

Core is smart, adapters are dumb — see CLAUDE.md and docs/ARCHITECTURE.md for details.

License

MIT