Open Archives MCP Server
Production-grade hybrid MCP + HTTP + SSE server generated from the Open Archives OpenAPI specification. Covers genealogical records (births, deaths, marriages, censuses), archive statistics, historical weather, and full-text page transcriptions of historical documents.
OpenAPI source used to generate tools:
../api/openapi.yaml (local)
https://api.openarchieven.nl/openapi.yaml (remote)
Overview
A schema-aware server that automatically converts the OpenAPI specification into callable tools and exposes them through multiple transports:
- MCP Remote (JSON-RPC over StreamableHTTP)
- HTTP JSON API
- SSE streaming with auto-pagination
- Chunked HTTP streaming with auto-pagination
- Redis caching (optional)
- Health checks
Use with Claude
Add as a custom connector
A hosted endpoint is available — no installation required.
In claude.ai or Claude Desktop:
- Open Settings → Connectors.
- Click Add custom connector.
- Enter the URL:
https://mcp.openarchieven.nl/
- Save and approve when prompted.
No authentication is required — Open Archives is a public dataset.
Example queries
Once the connector is added you can ask Claude, for example:
- "Who are the ancestors of Johannes Gregorius Marinus Coret? Give me an overview, including source citations in markdown format with the links to the original archives if possible, otherwise provide the Open Archieven links and provide a tree in SVG."
- "Did Johannes Coret and Antonia Uphus have descendants? Give me an overview, including source citations in markdown format with the links to the original archives if possible, otherwise provide the Open Archieven links, include thumbnails from scans from archival sources if available and provide a tree in SVG."
- "Provide me with a list of sourcetypes per archive(name) where I can find information about the Coret family. Show the result in a markdown document including links to the search pages on Open Archieven. Instead of the archive code use the ISIL if available."
- "What was the weather in Amsterdam on 1953-02-01?"
- "What does the 1850 census say about Utrecht?"
Claude will call the matching tool (search_records, show_record,
get_marriages, get_historical_weather, get_census_data, …) and
return links to the corresponding record pages on
https://www.openarchieven.nl.
Self-hosted
The server speaks MCP over Streamable HTTP; there is no stdio distribution. To run
your own instance:
git clone https://github.com/coret/openarchieven-mcp-server.git
cd openarchieven-mcp-server
npm install
npm run generate
npm run build:viewer
npm start
Point your MCP client at http://localhost:3001/ (or your own public URL) the same
way as the hosted endpoint above.
Core Features
OpenAPI Auto-Generation
Every API operation becomes a tool automatically via generate.ts.
All 21 operations:
| Tool Name | Description |
|---|
search_records | Search genealogical records |
show_record | Show a single genealogical record |
match_record | Match a person to birth and death records |
get_births_years_ago | List births from N years ago |
get_births | Find birth records |
get_deaths | Find death records |
get_marriages | Find marriage records |
get_archives | List all archives with statistics |
get_record_stats | Record count per archive |
get_source_type_stats | Record count per source type |
get_event_type_stats | Record count per event type |
get_comment_stats | Comment count statistics |
get_family_name_stats | Family name frequency |
get_first_name_stats | First name frequency |
get_profession_stats | Profession frequency |
get_breakdown | Cross-tabulation grouped by archive, source type, event type, place or year |
get_historical_weather | Historical weather from KNMI |
get_census_data | Dutch census data 1795–1899 |
search_transcriptions | Full-text search across page transcriptions of historical documents |
browse_transcriptions | Hierarchically browse transcriptions by source archive, archive number or inventory |
show_transcription | Retrieve a single page transcription by id |
Note: The callback (JSONP) parameter present in the upstream API is excluded from all tools — it is irrelevant in an MCP/JSON-RPC context.
Interactive Transcription Viewer (MCP App)
Beyond the 21 auto-generated tools, the server registers one hand-written tool — view_transcription — that opens transcribed pages in an interactive IIIF deep-zoom viewer (OpenSeadragon) with the transcription text alongside, using the MCP Apps extension (io.modelcontextprotocol/ui).
| Tool Name | Description |
|---|
view_transcription | Open one or more transcribed pages (ids: ["NL-SdmGA_1504889_11", …]) in a deep-zoom IIIF viewer with the transcript; optional highlight_term |
- Apps-capable hosts (Claude web/desktop, paid custom connectors) render the viewer (
ui://openarchieven/viewer.html) in a sandboxed iframe that loads images directly from the transcription image hosts (CSP-allowlisted as *.transkribus.eu, *.archief.nl, *.archieven.nl, *.memorix.nl). Transkribus and the archief.nl iipsrv server provide true IIIF deep zoom; the preserve*.archieven.nl thumbnail hosts render as flat images.
- Plain hosts receive a graceful fallback: a text summary with the IIIF/source URLs plus a few inline preview images.
The viewer is a single self-contained dist/viewer.html, bundled at build time with npm run build:viewer (vite + vite-plugin-singlefile); it is not listed by the REST GET /tools endpoint, only via MCP tools/list.
Schema-Perfect Validation
Uses actual OpenAPI parameter schemas. Validates:
- required parameters
- integer fields
- number fields
- enum values
- minimum / maximum constraints
Multiple Interfaces
MCP Remote (StreamableHTTP)
POST / ← canonical public endpoint (mcp.openarchieven.nl)
POST /mcp ← local / legacy alias
Stateless JSON-RPC transport — a new MCP server instance is created per request.
Origin validation: Browser requests must come from claude.ai,
claude.com, or any domain listed in ALLOWED_ORIGINS. Requests with no
Origin header (native MCP clients, curl, server-to-server) are
accepted. Unknown origins receive HTTP 403.
HTTP JSON
GET /tools
POST /tools/:name
SSE Streaming (auto-paginating)
Chunked HTTP Streaming (auto-paginating)
Static, hand-editable JSON files served verbatim from well-known/:
GET /.well-known/mcp/server-card.json ← SEP-1649 MCP Server Card
GET /.well-known/mcp.json ← alias of the server card
GET /.well-known/agent-card.json ← A2A v0.3 Agent Card
GET /.well-known/agent.json ← alias of the agent card
Edit well-known/mcp-server-card.json and well-known/agent-card.json directly — no restart required (files are read on each request). Responses are sent with Content-Type: application/json; charset=utf-8 and Cache-Control: public, max-age=3600.
Streaming endpoints (/events/:name, /stream/:name) automatically paginate through results for endpoints that support a start offset:
- Increments
start by number_show per page
- Stops when results are exhausted or after 20 pages (safety cap)
- SSE sends a
: heartbeat comment every 10 seconds to keep connections alive
Redis Cache
Optional Redis support.
If Redis is running, upstream responses are cached with a per-tool TTL tuned to
the data's volatility:
| Category | TTL | Tools |
|---|
| Immutable historical | never expires | get_historical_weather, get_census_data |
| Slowly-changing metadata | 7 days | get_archives |
| Stats aggregations | 1 day | get_record_stats, get_source_type_stats, get_event_type_stats, get_comment_stats, get_family_name_stats, get_first_name_stats, get_profession_stats, get_breakdown |
| Individual record lookups | 1 day | show_record, show_transcription |
| Search-style | 6 hours | search_records, match_record, get_births, get_deaths, get_marriages, search_transcriptions, browse_transcriptions |
| Date-bound | next UTC midnight | get_births_years_ago |
CACHE_TTL (default 3600) is the fallback for any tool not in the map above.
If Redis is unavailable:
- server still runs normally (degraded mode)
Rate Limiting
The upstream API enforces 4 requests per second per IP. The server queues all upstream calls through a token-bucket rate limiter (configurable via RATE_LIMIT_RPS).
Health Checks
Project Files
generate.ts
server.ts
tsconfig.json
package.json
.env.example
generated/
tools.json
spec.json
Requirements
- Node.js 18+
- npm
- optional Redis server
Configuration
Copy .env.example to .env and adjust:
| Variable | Default | Description |
|---|
PORT | 3001 | HTTP port |
OPENAPI_PATH | ../api/openapi.yaml | Path or URL to OpenAPI spec |
UPSTREAM_BASE | https://api.openarchieven.nl/1.1 | Upstream API base URL |
RATE_LIMIT_RPS | 4 | Upstream requests per second |
REDIS_URL | redis://localhost:6379/5 | Redis connection URL (db 5) |
CACHE_TTL | 3600 | Fallback cache TTL in seconds (used for tools not in the per-tool map; see Redis Cache) |
LOG_LEVEL | info | trace debug info warn error fatal |
NODE_ENV | (unset) | Set to production for JSON logs (default: pretty-printed) |
ALLOWED_ORIGINS | (empty) | Extra Origin headers allowed on the MCP endpoint (comma-separated). Claude domains and requests without an Origin header are always allowed. |
Install
Run from local spec:
Or from remote URL:
npx tsx generate.ts https://api.openarchieven.nl/openapi.yaml
Expected result:
Generated 21 tools
Output: generated/tools.json, generated/spec.json
Creates:
generated/tools.json
generated/spec.json
Start Server
Expected startup (development — pretty-printed):
[12:00:00] INFO: Open Archieven MCP server started
port: 3001
tools: 21
upstream: "https://api.openarchieven.nl/1.1"
rateLimit: "4 req/s"
redis: "redis://localhost:6379/5"
env: "development"
In production (NODE_ENV=production) each log line is a single JSON object.
Server binds to:
Test All Features
1. Health Check
curl http://localhost:3001/health
Expected:
{
"ok": true,
"tools": 21,
"redis": false,
"uptime": 1.23
}
curl http://localhost:3001/tools
Expected:
[
"search_records",
"show_record",
"match_record",
"get_births_years_ago",
"get_births",
"get_deaths",
"get_marriages",
"get_archives",
"get_record_stats",
"get_source_type_stats",
"get_event_type_stats",
"get_comment_stats",
"get_family_name_stats",
"get_first_name_stats",
"get_profession_stats",
"get_historical_weather",
"get_census_data",
"search_transcriptions",
"browse_transcriptions",
"show_transcription"
]
curl -X POST http://localhost:3001/tools/search_records \
-H "Content-Type: application/json" \
-d '{"name":"Coret"}'
4. Show a Single Record
curl -X POST http://localhost:3001/tools/show_record \
-H "Content-Type: application/json" \
-d '{"archive":"hua","identifier":"E13B9821-C0B0-4AED-B20B-8DE627ED99BD"}'
5. SSE Streaming
curl -N "http://localhost:3001/events/search_records?name=Coret"
Expected stream:
event: page
data: {...}
event: page
data: {...}
event: done
data: {}
6. Heartbeat Test
Leave SSE open for 15+ seconds — expect periodic keep-alive lines:
7. Chunked HTTP Streaming
curl -N -X POST http://localhost:3001/stream/search_records \
-H "Content-Type: application/json" \
-d '{"name":"Coret"}'
Expected (newline-delimited JSON):
{"query":{...},"response":{"number_found":...,"docs":[...]}}
{"query":{...},"response":{"number_found":...,"docs":[...]}}
8. MCP Initialize
curl -X POST http://localhost:3001/ \
-H "Content-Type: application/json" \
-H "Accept: application/json, text/event-stream" \
-d '{
"jsonrpc": "2.0",
"id": 1,
"method": "initialize",
"params": {
"protocolVersion": "2025-03-26",
"capabilities": {},
"clientInfo": { "name": "test", "version": "1.0" }
}
}'
curl -X POST http://localhost:3001/ \
-H "Content-Type: application/json" \
-H "Accept: application/json, text/event-stream" \
-d '{
"jsonrpc": "2.0",
"id": 2,
"method": "tools/list"
}'
curl -X POST http://localhost:3001/ \
-H "Content-Type: application/json" \
-H "Accept: application/json, text/event-stream" \
-d '{
"jsonrpc": "2.0",
"id": 3,
"method": "tools/call",
"params": {
"name": "search_records",
"arguments": { "name": "Coret" }
}
}'
Redis Testing
Start Redis
Restart the MCP server. Expected in /health:
Without Redis
Stop Redis and restart. Expected:
Common Commands
Regenerate after API changes
Restart server
Run tests
Tests cover the per-tool TTL strategy (cache-ttl.ts) using Node's built-in test runner — no extra dependencies. The coverage test asserts every tool emitted by generate.ts has an explicit TTL entry, so re-running npm run generate after an upstream OpenAPI change will surface any new tool that needs a TTL decision.
Troubleshooting
Generated files missing
Port already in use
Linux / macOS:
lsof -i :3001
kill -9 <PID>
Windows:
netstat -ano | findstr :3001
taskkill /PID <PID> /F
Redis not connecting
Server runs normally without Redis. Check REDIS_URL in .env.
Rate limit errors (429)
The upstream API allows 4 req/s per IP. The built-in rate limiter queues requests automatically. If you are running multiple server instances, reduce RATE_LIMIT_RPS or use a shared queue.
Privacy Policy
This server is a thin proxy over the public
Open Archives API. It does not require user
authentication and does not collect personal data of its own.
The full privacy policies of the operators apply in addition to this section:
Data collection
- Tool arguments (e.g. a search name, an archive code, a record
identifier) are received from the MCP client.
- HTTP request metadata — method, path, status code, latency, and the
source IP — is observed by the reverse proxy in front of the hosted
endpoint at
mcp.openarchieven.nl.
- No accounts, cookies, tokens, or session identifiers are collected.
The server is anonymous-by-design.
Use and storage
- Tool arguments are forwarded verbatim over HTTPS to
https://api.openarchieven.nl/1.1 to fulfill the request, and the
upstream response is returned to the caller.
- Application logs (tool name, arguments, status, latency) are written
to
stdout via pino. On the hosted
endpoint these logs are ephemeral: they are not written to disk and
are lost on process restart. Set LOG_LEVEL=warn to suppress argument
logging.
- Cache (optional): when Redis is configured, upstream responses are
cached under keys of the form
mcp:<tool>:<sorted-params-json>. The
cache contains response bodies only; no user identifiers are stored.
Third-party sharing
No data is sent to any service other than the upstream Open Archives API
listed above. There are no analytics, telemetry, advertising, or
observability third parties involved.
Data retention
| Data | Retention |
|---|
| Tool arguments and responses | Not persisted by the application |
| Application logs | Ephemeral (stdout, lost on restart) |
| Redis cache entries | Per-tool TTL (6 hours – 7 days; immutable historical lookups never expire). See Redis Cache. |
| Reverse-proxy access logs | Per the hosting provider's standard retention policy |
Security
The MCP endpoint validates the Origin header on every request and
rejects unknown browser origins (DNS-rebinding defense). All transport is
over HTTPS.
External links surfaced to clients
Tool responses include URLs that point to record pages on
https://www.openarchieven.nl. The submission declares the following
allowed link URI so users are not prompted to confirm each link:
https://www.openarchieven.nl
For privacy questions or requests, contact:
Recommended Production Upgrades
- HTTPS reverse proxy (nginx / caddy)
- PM2 or systemd process manager
- Structured JSON logging (pino / winston)
- Request tracing (OpenTelemetry)
- Auth middleware if server is public-facing
- Shared Redis for multi-instance deployments
Version
Schema-perfect OpenAPI-generated MCP server for Open Archives.