Northdata MCP

NorthData MCP Server

Connect the NorthData company-data API to Claude Desktop, Claude Code, Cursor, and other AI assistants.

The Model Context Protocol (MCP) standardizes how Large Language Models (LLMs) talk to external services like NorthData. This server exposes NorthData's suggest, search, company, person, publications, reference, and billing endpoints as MCP tools, with a built-in credit guard that prevents accidental high-cost calls. See the full list of tools.

Built on [FastMCP v3](https://gofastmcp.com). Uses stdio transport — the MCP client launches northdata-mcp as a subprocess.

Setup

1. Read the security notes

Before running the server, skim the security considerations. NorthData costs real money per returned company, and MCP lets LLMs spend those credits autonomously. The credit guard mitigates this, but defence in depth matters.

2. Install the server

pipx install northdata-mcp

This pulls in the sibling northdata-cli package automatically and exposes the northdata-mcp console script globally.

Alternative via uv:

uvx northdata-mcp --help

3. Get a NorthData API key

Obtain a key at https://www.northdata.com. You'll put it into the MCP client config in the next step.

Optional credit-guard overrides (set as env vars in the MCP client config):

4. Configure your MCP client

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) — adjust the path on Windows / Linux:

{

"mcpServers": { "northdata": { "command": "northdata-mcp", "env": { "NORTHDATA_API_KEY": "XXXX-XXXX" } } } }

Restart Claude Desktop. The NorthData tools will appear in the tool picker.

Claude Code

claude mcp add northdata --env NORTHDATA_API_KEY=XXXX-XXXX -- northdata-mcp

Cursor

Add to ~/.cursor/mcp.json:

{

"mcpServers": { "northdata": { "command": "northdata-mcp", "env": { "NORTHDATA_API_KEY": "XXXX-XXXX" } } } }

Any MCP client (generic)

Run northdata-mcp as a subprocess and speak MCP over stdio. The environment must contain NORTHDATA_API_KEY.

Tools

_Note: this server is pre-1.0, expect minor breaking changes between versions._

Every tool is a thin wrapper around the equivalent northdata-cli command. Tools are tagged free (no credits spent) or billed (consumes NorthData credits). MCP annotations (readOnlyHint, openWorldHint) are set on every tool so well-behaved clients can surface them in the UI.

Free tools

No NorthData credits are consumed by these tools.

• `suggest` — Autocomplete company or person names. Cheap first step before spending a credit on company or person.
• `billing` — Current remote credit usage for the billing period (numberOfRequests, periodStart, periodEnd).
• `reference_overview` — API reference data: standards, countries, legal forms, events.
• `reference_segment_codes` — Complete segment-code list for a given standard (default NACE2025).
• `local_credit_log` — Credit log maintained locally by this server: total, per-month, per-entry breakdown.

Billed tools

Every one of these spends NorthData credits. NorthData charges per returned company, not per request. The credit guard enforces a two-tier ceiling on search (see Credit guard). All billed tools accept dry_run=true to build and return the URL without spending credits.

• `search` — Power search. Cost: up to limit credits. Filters: segment_codes, segment_standard, legal_forms, address, max_distance_km, countries, status, revenue_min/max, earnings_min/max, limit, pos, approve_high_cost.
• `company` — Full company profile (1 credit). Identify by register (e.g. HRB 123456/Muenchen) or by name + optional city. Toggles: owners, financials, representatives, extras, sheets, events.
• `person` — Person lookup (1 credit). Returns birth date and known roles. Takes first_name, last_name, optional city.
• `publications` — Publications lookup (1 credit). Useful for shareholder lists via source="Hrb".

Credit guard

NorthData bills per returned company. A single search call with limit=100 costs up to 100 credits. This server enforces two tiers:

Every billed tool call is appended to the local credit log (JSON-lines). The local_credit_log tool surfaces it. The log is informational; billing remains the source of truth.

Use dry_run=true on any billed tool to inspect the request URL before spending anything.

Security considerations

Connecting a credit-consuming data source to an LLM carries real costs and real risks. A malicious or malformed prompt can cause an agent to burn credits, leak structured company data into an untrusted chat, or combine NorthData data with other tools in unintended ways.

Recommendations

• Keep the credit guard strict. The defaults (approval 25, absolute max 100) exist to prevent a runaway agent from draining your credits in a single call. Only raise them deliberately.
• Prefer free tools during exploration. suggest, reference_*, and billing are free. LLMs should call these before moving to company / person / publications.
• Use `dry_run=true` when experimenting with new search parameters or register identifiers.
• Review tool calls before approval. Most MCP clients (Claude Desktop, Cursor) ask you to approve each tool call. Leave that setting on.
• Limit scope per session. Launch the server with a trial API key, not your main production key, when giving it to an agent for open-ended work.
• Prompt injection. NorthData responses contain user-generated text (company descriptions, publications) that an adversary could stuff with instructions. Treat tool output as untrusted data, not as trusted instructions to the LLM.

Development

cd northdata-mcp

python -m venv .venv && source .venv/bin/activate pip install -e "../northdata-cli" # local editable CLI pip install -e ".[dev]" pytest

Tests use FastMCP's in-memory transport — no subprocess, no network. The underlying NorthDataClient is injected via a mock factory backed by httpx.MockTransport.

Resources

• Model Context Protocol: https://modelcontextprotocol.io
• FastMCP v3 docs: https://gofastmcp.com
• NorthData API: https://github.com/northdata/api
• Sibling package: northdata-cli — same tools, plain command-line

License

MIT. See LICENSE.