OpenInsurance MCP Server

openinsurance-wiki

A brain for a country's insurance market. A self-sufficient, open-source, country-agnostic framework that turns a nation's public insurance documents into a rich, interconnected, source-cited knowledge base that any AI agent can read.

Not a chatbot. Not a RAG black box. A transparent, reproducible knowledge graph: the repo contains the whole chain - it finds insurers' public general-conditions PDFs, downloads them, and turns each one into a faithful Markdown page that preserves the maximum of what the PDF actually says, with a citation back to the source - all cross-linked into a navigable graph of products, insurers, branches, regulations and concepts.

As far as we know, this is the only open-source, machine-readable, source-cited database of insurance products (the closest equivalents are commercial and closed). Point it at any country: the taxonomy is data, not code, and adding a country is a documented recipe. The first reference dataset already covers 17 insurers and 162 products across 12 branches in Belgium (auto, home, health, liability, travel, legal protection, ...), each page cited to its source document.

---

Use it in 2 minutes - no API key needed

The dataset ships in the repo, already built: 162 product pages, insurer pages, glossary, plus the structured JSON behind them. You only need an LLM key to re-extract from scratch, never to use it.

1. Browse it. Open the repo as an Obsidian vault (the [[wikilinks]] become a navigable graph). Note: github.com does not render [[wikilinks]] as links, so the vault or the website is the comfortable way to read.

2. Plug it into an agent (MCP). The MCP server is keyless and read-only:

git clone https://github.com/sluyasu/OpenInsurance.git
cd OpenInsurance
python3 -m venv .venv && .venv/bin/pip install "mcp[cli]" pyyaml

Then register it with any MCP client, e.g. Claude Code:

claude mcp add insurance-wiki -- "$(pwd)/.venv/bin/python" "$(pwd)/mcp/insurance_wiki_mcp.py"

You get search, get_product, compare_products, find_overlap (candidate duplicate cover when combining two policies), get_branch_overview, ... See mcp/README.md.

3. Take the raw data. data/be/extracted/ holds one structured JSON per source document, validated against schema/; data/be/index.json is the flat index. AGENTS.md is a generated manifest (note types, counts, per-page path / source_url / freshness) so a file-reading agent can navigate without guessing.

---

How it works

Why this exists

Insurance products are documented in dense PDFs scattered across dozens of insurer websites. There is no neutral, machine-readable, navigable map of what actually exists in a national market. This project builds one - as a public good, and in a form an AI agent can plug into.

It also plugs into a market that is standardizing around it: EU regulation 2017/1469 gives every non-life product a standardized summary (the IPID), EIOPA actively promotes product comparison and switching, and the open insurance agenda (OPIN, the EU FIDA proposal) pushes for machine-readable access to insurance data. This project is the missing public documents layer of that picture: what the products actually say, in the open.

Three things make it different:

1. Self-sufficient & reproducible. Clone it, add your own LLM key, run make all. It scrapes, downloads and extracts from scratch. No hidden datasets - every input is committed, every output is regenerable.
2. Transparent extraction. The exact prompt sent to the LLM is a file in this repo (extraction-agent/), not buried in code. You can read precisely what the model was asked, and run the identical extraction with your own model (Claude, Gemini, GPT, or a local model).
3. Grounded & cited. Every product page traces back to the source PDF and cites page numbers. Quotes are verified against the raw text. If it isn't in the document, it isn't on the page.

---

What's in the wiki

Per country (wiki/be/):

Generated and hand-authored pages live in separate folders and never collide: you fix a fact by editing the extraction data and rebuilding, never by editing a generated page.

Every page is Obsidian-compatible Markdown with YAML frontmatter and [[wikilinks]].

---

Reproduce it from scratch

This is the only path that needs an LLM key (the extraction step). Scraping and download use a free stack (httpx + Playwright).

git clone https://github.com/sluyasu/OpenInsurance.git
cd OpenInsurance
make setup                      # deps + playwright chromium (no paid scraping dependency)
cp .env.example .env            # set LLM_PROVIDER + your API key (any provider)

# Reproduce a slice end-to-end:
make download COUNTRY=be INSURER=<slug>   # fetch the public PDFs
make extract  COUNTRY=be INSURER=<slug>   # PDFs -> rich Markdown + JSON (uses YOUR model)
make build    COUNTRY=be                  # assemble the wiki
make validate COUNTRY=be                  # citation / wikilink / frontmatter gates

# ...or the whole chain:
make all COUNTRY=be

Extraction is resumable (skip-existing keyed by source checksum + prompt version), so large runs can stop and restart safely.

How the pipeline works

sources/be/<insurer>.yml     (committed: where the public PDFs live)
        │  discover.py   crawl listing pages (httpx, Playwright fallback)
        ▼
data/be/pdfs/…               (downloaded; gitignored - regenerable; manifest.json committed)
        │  extract.py    PyMuPDF text  ──►  LLM (extraction-agent/ prompts)  ──►  MD + JSON
        ▼
data/be/extracted/…          (rich Markdown + structured JSON, page-cited)
        │  build_wiki.py
        ▼
wiki/be/…                    (the browsable, agent-readable knowledge base)

Details: CONTRIBUTING.md (how to add a country / insurer / product) and extraction-agent/ (the exact prompts).

Every push runs the CI gates: wiki validation (frontmatter, wikilinks, citations) and build idempotence (rebuilding the committed wiki must produce a zero diff).

---

Add a country

1. sources/<cc>/_country.yml - regulator, languages, branch taxonomy.
2. sources/<cc>/<insurer>.yml - where each insurer's public PDFs live.
3. wiki/<cc>/ - hand-author branch/regulation/glossary overviews (or start them as stubs).
4. make all COUNTRY=<cc>.

Nothing in the schema is Belgium-specific - the taxonomy is data, not structure.

---

Licensing & provenance

Dual-licensed: code (pipeline/, mcp/, adapters, schema) under MIT; original content (wiki/, extracted data, prompts, sources) under CC-BY-4.0. Short verbatim excerpts quoted from insurers' public documents remain the property of their publishers and are not relicensed - see LICENSE, LICENSE-CONTENT and NOTICE.

Product pages are a factual extraction from insurers' publicly published documents, attributed to each source PDF. They are not the insurers' official documents and may contain extraction errors - always verify against the cited source_url. This project provides information, not personalized insurance advice.