KAELIX Local Data MCP — zero-dependency AI tools

A working, local-only MCP server for deterministic CSV/JSON quality checks and SHA-256. Run it with Python 3.10+ in Claude Desktop, Cursor, or any MCP stdio client. This is a free developer utility, not a paid API or investment product.

Built by KAELIX Studio — custom Python/API automation, CSV validation and code QA services.

What it does

CSV/JSON tools return aggregated information only, not raw row/field values. No wallet keys, account authentication, credentials, file access, tracking, external HTTP calls, shell commands, or network listeners. The MCP client itself may log requests; use synthetic examples when in doubt.

Install in 3 steps

1. Download this repository (git clone https://github.com/MaxGab/kaelix-data-mcp.git).
2. Confirm Python >= 3.10: python3 --version.
3. Point your MCP client at the downloaded script.

Example Claude Desktop / Cursor MCP configuration (replace the absolute path with your local path):

{
  "mcpServers": {
    "kaelix-local-data": {
      "command": "python3",
      "args": ["/absolute/path/to/kaelix-data-mcp/kaelix_mcp.py"]
    }
  }
}

No pip install or API tokens necessary. This is an stdio server, not a public HTTP endpoint. The process prints only JSON-RPC responses to stdout.

Test without an MCP client

python3 -m unittest -v test_kaelix_mcp.py
python3 kaelix_mcp.py

Then enter one complete JSON line:

{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"demo","version":"1.0"}}}

On another line, request tools/list or tools/call. The included unit tests run a real subprocess and exercise the full stdio handshake and tool invocation.

MCP CSV example

{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"kaelix_audit_csv","arguments":{"csv_text":"id,total\n1, 20\n1,20\n2,\n","required_columns":["id","total"]}}}

The response contains only quality metrics, not any data rows.

Boundaries and transparency

The server caps inbound requests at 128 KB, tool text at 64 KB, and CSV processing at 4,000 rows × 200 columns. It never opens arbitrary paths. Duplicate-row detection uses transient SHA-256 fingerprints, and nothing is stored between calls.