Give Your AI Agent Access to Every Taxonomy

TL;DR: The WorldOfTaxonomy MCP server gives AI agents structured tool access to 1,000+ classification systems. 25 tools. Works with Claude, Cursor, VS Code, Windsurf, and any MCP client. Setup takes under 5 minutes.

The problem with LLMs and codes

Ask an LLM to translate NAICS 4841 to NACE and it will try. Sometimes it gets close. Often it hallucinates a plausible-looking code that does not exist.

LLMs lack three things for classification work:

Verified data - they guess from training data, not from authoritative sources
Graph structure - they cannot traverse hierarchies or follow crosswalk edges
Completeness - they know major systems but miss domain taxonomies and national adaptations

The fix is not better prompting. It is giving the AI structured tool access to the actual database.

How MCP works

sequenceDiagram
    participant AI as AI Agent
    participant MCP as MCP Server
    participant DB as PostgreSQL

    AI->>MCP: Initialize session
    MCP-->>AI: 25 tools + system knowledge
    Note over AI: Agent now knows all 1,000 systems
    AI->>MCP: search_classifications("hospital")
    MCP->>DB: Full-text search across 1.2M codes
    DB-->>MCP: Verified results
    MCP-->>AI: Structured JSON response
    AI->>MCP: translate_code("naics_2022", "5415", "nace_rev2")
    MCP->>DB: Query equivalence edges
    DB-->>MCP: NACE 62.01 (exact match)
    MCP-->>AI: Verified translation

Instead of guessing, the AI calls tools. The tools query the database. The AI gets verified, structured data.

Setup

Claude Desktop

{
  "mcpServers": {
    "world-of-taxonomy": {
      "command": "python3",
      "args": ["-m", "world_of_taxonomy", "mcp"],
      "env": { "DATABASE_URL": "your-database-url" }
    }
  }
}

Cursor / VS Code

Add to your MCP settings (.cursor/mcp.json or VS Code MCP config):

{
  "world-of-taxonomy": {
    "command": "python3",
    "args": ["-m", "world_of_taxonomy", "mcp"],
    "env": { "DATABASE_URL": "your-database-url" }
  }
}

Claude Code

claude mcp add world-of-taxonomy python3 -m world_of_taxonomy mcp

The 25 tools

Category	Tools	What they do
Search	`search_classifications`, `find_by_keyword_all_systems`	Full-text search across all systems
Translate	`translate_code`, `translate_across_all_systems`	Code-to-code translation
Navigate	`browse_system`, `explore_industry_tree`, `get_children`, `get_ancestors`, `get_siblings`	Hierarchy traversal
Detail	`get_node_detail`, `get_system_detail`, `list_all_systems`	Metadata and context
Compare	`compare_sector`, `get_system_diff`, `get_crosswalk_stats`, `get_equivalences`	Cross-system analysis
Country	`get_country_taxonomy_profile`, `get_country_systems`, `get_world_coverage_stats`	Geographic context
Knowledge	`read_wiki_page`, `list_wiki_pages`	Curated guide content

Example prompts

Classify a business:

"I have a company that manufactures electric vehicle batteries in Germany. What industry codes apply?"

The AI calls search_classifications, then get_country_taxonomy_profile for Germany, then translate_code - returning verified WZ 2008, NACE, and ISIC codes.

Cross-border job mapping:

"Our US job posting uses SOC 15-1252. What is the EU equivalent?"

The AI calls translate_across_all_systems and returns ESCO, ISCO-08, and other mapped systems with match types.

Gap analysis:

"Which of our NAICS codes have no NACE equivalent?"

The AI calls get_system_diff and returns the exact gap list.

Why MCP instead of RAG

Approach	Strengths	Weaknesses
RAG	Good for fuzzy semantic search	Approximate matches, no graph traversal
MCP	Exact lookups, hierarchy navigation, crosswalk edges	Requires structured queries

Classification codes are not fuzzy semantic content. They are structured hierarchical data with precise relationships. RAG returns approximate matches. MCP returns exact matches from a relational database.

The MCP server also enables operations that vector search cannot perform: hierarchy traversal, multi-hop translation, gap analysis, and system comparison.