MCP Server — Connect CaveauAI to Claude Code & Claude Desktop

2 views

MCP Server Integration

Blue Note Logic provides a Model Context Protocol (MCP) server that lets you search our legal and regulatory knowledge bases directly from Claude Code and Claude Desktop. This gives AI assistants real-time access to curated European legal knowledge without leaving your development environment.

What is MCP?

The Model Context Protocol (MCP) is an open standard by Anthropic that allows AI models to access external tools and data sources. Our MCP server exposes CaveauAI's knowledge corpus as a set of tools that Claude can call during conversations.

Instead of copy-pasting legal text into a chat window, you simply ask Claude a question and it automatically searches our knowledge bases, retrieves relevant passages, and cites its sources.

Available Tools

Our MCP server (legal-knowledge) provides 9 tools:

Search & Retrieval

Tool	Description
`search_corpus`	Semantic vector search across all shared corpuses. Accepts natural language queries and returns the most relevant document chunks with metadata.
`search_keywords`	Metadata-based search (no embedding). Filter by jurisdiction, authority type, year range, language, and category.
`search_private`	Search a client's private uploaded documents. Tenant-isolated — results are scoped to the specified client ID.
`get_document`	Retrieve the full text of a document by ID, with all chunks in order. Useful for reading a complete document found via search.

Analysis & Q&A

Tool	Description
`ask`	Full RAG pipeline: embeds your question → searches relevant chunks → builds context → generates an LLM answer with source citations. One-stop question answering.

Corpus Management

Tool	Description
`list_corpuses`	List all available knowledge corpuses with document and chunk counts.
`get_categories`	Get the hierarchical category taxonomy for a corpus (categories → subcategories).
`corpus_stats`	Detailed analytics: document/chunk counts broken down by category, jurisdiction, language, and publication year.
`ingest_text`	Ingest a text document into a corpus. Chunks the text, embeds each chunk, and stores in both MySQL and Qdrant.

Available Corpuses

Corpus Slug	Contents
`do-better-legal`	Norwegian law — Lovdata, KOFA decisions, court rulings, regulations, family law
`corpus_eu_ai`	EU AI Act — DPA decisions, EDPB guidance, enforcement actions, implementing acts
`corpus_co2`	CO2 & Climate — EU ETS, Paris Agreement, Nordic climate laws, ESG frameworks
`corpus_telecom`	Telecom regulation — BEREC guidelines, ETSI standards, Nkom decisions

Setup for Claude Code (CLI)

Add this to your .mcp.json file in your project root:

{
  "mcpServers": {
    "legal-knowledge": {
      "command": "ssh",
      "args": [
        "bluenotelogic@89.167.24.252",
        "/home/bluenotelogic/mcp-legal/venv/bin/python /home/bluenotelogic/mcp-legal/server.py --stdio"
      ]
    }
  }
}

Prerequisites:

SSH access configured (key-based authentication to S1)
Claude Code installed (npm install -g @anthropic-ai/claude-code)

Once configured, Claude Code will automatically discover the MCP server and make the tools available in conversations.

Setup for Claude Desktop

Claude Desktop connects via SSE (Server-Sent Events) transport on port 8001:

{
  "mcpServers": {
    "legal-knowledge": {
      "transport": "sse",
      "url": "http://89.167.24.252:8001/sse"
    }
  }
}

The SSE transport runs as a systemd service on S1, ensuring it's always available.

Search Filters

Most search tools support these optional filters:

Filter	Values	Description
`corpus`	`do-better-legal`, `corpus_eu_ai`, etc.	Limit search to specific corpus
`language`	`en`, `no`, `fr`, `de`, `sv`, etc.	Filter by document language
`jurisdiction`	`eu`, `no`, `se`, `dk`, `fi`, etc.	Filter by jurisdiction
`authority_type`	`law`, `regulation`, `guidance`, `case_law`, `commentary`, `report`, `treaty`	Filter by document type
`year_from` / `year_to`	e.g., 2020-2026	Publication year range
`category`	e.g., `gdpr`, `ai-act`, `nis2`	Filter by category slug

Example Usage in Claude Code

Once configured, you can ask questions naturally:

"Search the EU AI corpus for high-risk AI system requirements"

Claude will call search_corpus with your query, retrieve relevant chunks, and synthesize an answer with citations to specific articles and regulatory documents.

"What Norwegian laws apply to employee monitoring in the workplace?"

Claude calls search_corpus targeting do-better-legal, finds relevant provisions from arbeidsmiljøloven and datatilsynet guidance, and provides a sourced answer.

"Get the full text of document 18445"

Claude calls get_document and returns the complete document with all chunks.

Technical Details

Server: Python FastMCP server running on S1
Embedding model: nomic-embed-text (768-dimensional vectors)
Vector store: Qdrant on S2 (10.0.0.4:6333)
LLM for ask tool: Configurable (default: qwen2.5:14b)
Response format: JSON with chunk content, metadata, relevance scores, and document references

Go back Similar articles