gtc-mcp

MCP server exposing insurance terms and conditions from the NN SOAP API.

What it provides

• list_gtc_documents

  • returns document metadata

  • supports arbitrary metadata filters

  • supports substring or exact matching

  • supports sorting and limiting

• get_gtc_document

  • fetches a document by id

  • extracts text from PDF content for agent use

  • truncates returned text by default and reports truncation metadata

• find_gtc_document_context

  • returns a compact context block with matching metadata and truncated bodies

• search_gtc_document_text

  • searches a document body for phrases

  • returns grep-like line context windows around matches

  • caps the total response size

• search_gtc_documents

  • filters by prodCode, componentCode, typeName, withdrawn

  • filters by dateFrom and modDate ranges

  • can restrict results to documents active on a given date

• diff_gtc_documents

  • compares two documents by extracted text

  • returns a unified diff and metadata for both files

• resources

  • gtc://documents/{doc_id}

  • gtc://documents/{doc_id}/full

The flexible filters are intentional because the exact metadata schema comes from the SOAP API response. The live API currently exposes fields such as idBodyDoc, idHeadDoc, docName, docTitle, typeName, componentCode, prodCode, dateFrom, dateTo, regDate, modDate, and withdrawn.

Related MCP server: docsearch

Install

python -m venv .venv
source .venv/bin/activate
pip install -e .

Run

gtc-mcp

Run with streamable HTTP:

gtc-mcp --transport streamable-http

Bind on all interfaces for container or homelab use:

gtc-mcp --transport streamable-http --host 0.0.0.0 --port 8000

Optional mount path for SSE:

gtc-mcp --transport sse --mount-path /mcp

Optional streamable HTTP path:

gtc-mcp --transport streamable-http --streamable-http-path /mcp

Optional override:

export GTC_WSDL="https://gtc.nn.pl/gtc/services/GtcServiceHttpPort?wsdl"

Document bodies are cached locally after the first fetch. Optional override:

export GTC_CACHE_DIR="/absolute/path/to/cache"

You can also set the transport defaults with environment variables:

export MCP_TRANSPORT="streamable-http"
export MCP_MOUNT_PATH="/mcp"
export MCP_HOST="0.0.0.0"
export MCP_PORT="8000"
export MCP_STREAMABLE_HTTP_PATH="/mcp"
export MCP_ALLOWED_HOSTS="gtc-mcp.gtc-mcp:8000"

If you run behind Kubernetes or another reverse proxy and see 421 Misdirected Request with Invalid Host header, set MCP_HOST=0.0.0.0 and provide the externally used hostnames in MCP_ALLOWED_HOSTS.

Container image

Build locally:

make docker-build

Publish to your Zot registry:

make docker-publish ZOT_REGISTRY=zot.example.lan:5000 IMAGE_TAG=latest

Build and push a multi-arch image for linux/amd64 and linux/arm64:

make docker-publishx ZOT_REGISTRY=zot.example.lan:5000 IMAGE_TAG=latest

You can override the target platforms if needed:

make docker-publishx ZOT_REGISTRY=zot.example.lan:5000 IMAGE_TAG=latest PLATFORMS=linux/amd64,linux/arm64

Run locally:

make docker-run

Example MCP configuration

{
  "mcpServers": {
    "gtc": {
      "command": "/absolute/path/to/gtc-mcp/.venv/bin/gtc-mcp"
    }
  }
}

Example tool calls

List documents where the product field contains travel:

{
  "filters": {
    "product": "travel"
  },
  "contains_match": true,
  "limit": 10
}

Get a document body:

{
  "doc_id": "12345",
  "max_text_chars": 6000
}

Get context for the agent:

{
  "filters": {
    "language": "pl",
    "product": "life"
  },
  "limit": 3,
  "max_text_chars_per_document": 2500,
  "max_result_chars": 7000
}

Search a document body for specific phrases with line context:

{
  "doc_id": "12345",
  "phrases": ["wyłączenia", "karencja"],
  "before_lines": 2,
  "after_lines": 3,
  "max_matches": 10,
  "max_result_chars": 6000
}

Compare two documents:

{
  "left_doc_id": "104352",
  "right_doc_id": "104356",
  "context_lines": 2,
  "max_diff_lines": 200
}

Find active documents for a product on March 11, 2026:

{
  "as_of_date": "2026-03-11",
  "prod_code": "FIR0",
  "withdrawn": 0,
  "limit": 10
}