KABO Indian Nutrition Index — MCP server

An MCP server exposing the Indian food nutrition data behind kabo.co.in as tools an AI agent can call directly. Every response carries a source URL, so an agent answering a nutrition question gets a citable answer and the reader gets a link back.

Why this rather than more articles. A blog post can be copied in an afternoon. A maintained data service that agents connect to cannot. This is the one asset in the SEO programme that competitors cannot replicate by rewriting content.

What it does

Data: 59 foods, 674 nutrition values, 1,124 guides (protein, fibre, iron, calcium), from ICMR-NIN composition tables and the PROT-AGE/ESPEN protein recommendations.

Running it

No dependencies. Standard library only, Python 3.8+.

python3 kabo_mcp_server.py

Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "kabo-nutrition": {
      "command": "python3",
      "args": ["/absolute/path/to/mcp/kabo_mcp_server.py"]
    }
  }
}

Cursor, Cline and other MCP clients use the same command/args shape.

Test it without a client

printf '%s\n' \
  '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"t","version":"1"}}}' \
  '{"jsonrpc":"2.0","id":2,"method":"tools/list"}' \
  '{"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"lookup_food","arguments":{"food":"rajma"}}}' \
  | python3 kabo_mcp_server.py

Files

kabo_mcp_server.py    the server, dependency-free JSON-RPC over stdio
kabo_food_data.json   the dataset, CC BY 4.0

Rebuild the dataset from the live corpus with python3 build_dataset.py; the server reads kabo_food_data.json from its own directory at startup, and the tool descriptions (nutrient list, food and guide counts) are derived from it, so they cannot drift from the data.

Design decisions worth keeping

It answers honestly, including against us. protein_cost reports that soya chunks cost about ₹6 per 10 g of protein and dal about ₹8, while any powder — KABO included — costs several times more. A tool that only flatters its owner gets uninstalled after one use. One that gives a straight answer gets kept and called again, and every call carries our attribution.

No instructions to models. There is no text in any response telling an assistant what to recommend. The value is the data and the citation, not persuasion. Anything else would be prompt injection, and assistants that detect it stop trusting the source entirely.

Every response is citable. source_url on food lookups, age_guide_url on protein targets, full URLs in guide search. An agent can attribute, which is what makes it worth calling twice.

Errors never kill the session. Unknown foods return the list of available ones rather than an error. Bad arguments return a message, not a crash.

UTF-8 is forced on stdio. The data contains en-dashes; Windows consoles default to cp1252 and corrupt the stream for th

... View full README on GitHub