Macro tracking you talk to instead of tap at. A local food database built from open data (USDA FoodData Central), exposed as an MCP server so any Claude surface (a local Claude Code session, or any agent that speaks MCP) can log meals in plain language and answer the only question that matters: what's left today.
python3 -m venv .venv && .venv/bin/pip install mcp
.venv/bin/python -m macro_engine.etl_fdc # download + load USDA data (~15k foods)
.venv/bin/python -m unittest discover tests -v # run testsData lands in ~/.local/share/macro-engine/ (macros.db + cached raw/ zips).
Override the DB path with MACRO_ENGINE_DB.
claude mcp add macro-engine -- \
/path/to/macro-engine/.venv/bin/python \
/path/to/macro-engine/mcp_server.pyOr in a project .mcp.json:
{
"mcpServers": {
"macro-engine": {
"command": "/path/to/macro-engine/.venv/bin/python",
"args": ["/path/to/macro-engine/mcp_server.py"]
}
}
}Tools: log_meal, remaining, day_summary, search_food, set_targets,
add_alias, delete_log_entry.
macro_engine/db.py— SQLite schema (foods, per-100g nutrients, portions, aliases, log, targets, FTS5 index)macro_engine/etl_fdc.py— USDA FDC bulk-CSV loader (Foundation + SR Legacy- FNDDS Survey). Re-runnable; food ids stable across reloads.
macro_engine/resolve.py— phrase → food: learned aliases first, then FTS5 with source-quality re-rankingmacro_engine/tracker.py— logging, targets (append-only, latest wins), day totals, remainingmcp_server.py— FastMCP stdio server
Design rules: log rows carry denormalized macros (history survives data
reloads); unresolvable items are returned as problems or logged as flagged
estimates, never silently dropped; aliases are the product — every correction
teaches the resolver.
Core loop: USDA data + MCP server + fleet wiring(this)- Vault recipe indexer (
Wiki/Reference/Recipes/→ per-serving macros) - Open Food Facts import (Canadian packaged goods), miss-queue → recon sourcing, chain-restaurant adapters
- Optional UI (standalone PWA or Training Engine feature)
MIT. See LICENSE.