# taco-br-mcp

[MCP](https://modelcontextprotocol.io) server that lets an AI assistant answer with exact numbers from the **TACO** — _Tabela Brasileira de Composição de Alimentos_, 4ª ed. (NEPA/UNICAMP, 2011) — instead of guessing. 597 Brazilian foods, 26 nutrients, per 100 g of edible portion, plus household measures ("2 colheres de sopa", "1 unidade") from the IBGE POF.

> **Em português:** servidor MCP com a Tabela Brasileira de Composição de Alimentos (TACO). Busca alimentos, devolve nutrientes, soma refeições e converte medidas caseiras em gramas. Toda resposta cita a fonte, e o que a tabela não traz vem como "sem dado", nunca como zero.

stdio by default, for Claude Desktop, Claude Code, Cursor and other local MCP clients. It also serves Streamable HTTP (`--http`), which is how ChatGPT connects: see [Remote HTTP and ChatGPT](#remote-http-and-chatgpt).

## Install

Requires Node.js 20 or newer.

**Claude Code**

```bash
claude mcp add taco-br -- npx -y taco-br-mcp
```

**Claude Desktop** (`claude_desktop_config.json`) and **Cursor** (`.cursor/mcp.json`)

```json
{
  "mcpServers": {
    "taco-br": { "command": "npx", "args": ["-y", "taco-br-mcp"] }
  }
}
```

Then ask, for example: _"Quantas kcal tem uma refeição de 4 colheres de sopa de arroz, 100 g de feijão carioca e 100 g de peito de frango grelhado?"_

## Remote HTTP and ChatGPT

ChatGPT cannot reach a server running on your PC, so the same tools are available over Streamable HTTP:

```bash
npx -y taco-br-mcp --http   # http://127.0.0.1:3000/mcp
```

A public instance runs at `https://taco-br-mcp.taco-br-mcp.workers.dev/mcp` (open, no login, rate limited; best effort, no uptime promise). The repository also ships a Cloudflare Workers deployment (free plan, open access, per-IP rate limit). Step-by-step deploy and the ChatGPT connector setup (Settings → Apps & Connectors, developer mode, authentication "No authentication") are in [docs/remote.md](docs/remote.md).

## Tools

| Tool                        | What it does                                                                                                                                                                                                            |
| --------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| `search_food`               | Find foods by Portuguese name (accent-insensitive, tolerates typos, knows regional names such as aipim/macaxeira/mandioquinha). Returns ids and kcal per 100 g.                                                         |
| `get_nutrients`             | Full composition of one food by id, per 100 g or scaled to a number of grams. Trace (`Tr`) is `{"trace":true}` and a value the table does not publish is `null`, never 0.                                               |
| `get_fatty_acids`           | Fatty acid profile of one food (saturated, mono/polyunsaturated, omega-3 and -6, EPA, DHA, trans, individual acids). Only 423 of 597 foods are in that TACO table; for the others it says `available: false`, not zero. |
| `get_amino_acids`           | Amino acid profile of one food (lysine, leucine, tryptophan...). TACO has it for only 26 foods; for any other it says so and lists the 26.                                                                              |
| `calc_meal`                 | Sum nutrients over `[{food_id, grams}]` or `[{food_id, household_measure, quantity}]`. A total is flagged `is_lower_bound` (and the foods without data are named) when some item has no data for that nutrient.         |
| `compare_foods`             | 2 to 10 foods side by side on chosen nutrients, per 100 g. Missing data stays `null`.                                                                                                                                   |
| `convert_household_measure` | "2 colheres de sopa" of a food to grams, from the IBGE POF table through a hand-curated mapping (204 foods so far). Ambiguous food names return candidates; uncovered foods or measures are reported, never guessed.    |
| `ping`                      | Health check.                                                                                                                                                                                                           |

All tools are read-only. Every response cites _TACO 4ª ed. (NEPA/UNICAMP, 2011)_.

## How it treats missing data

- `Tr` (trace, below the quantification limit) is `{"trace": true}`, never 0.
- A value the table does not publish is `null` and shown as "sem dado na TACO", never 0. The source marks these cells in three ways (`NA`, `*` for "analyses under re-evaluation", or blank); the processed CSV merges them, so the server does not tell them apart. Whole milk, for example, has no published energy or protein.
- Fatty and amino acids come from separate TACO tables that cover only part of the foods (423 and 26 of 597). A food outside the table gets `available: false`, never a row of zeros. In the amino acid table the TACO prints `0,00` for amounts under 0.005 g/100 g; that is a published value (shown as 0 or a small number), not missing data.
- `calc_meal` totals are **lower bounds** when an item lacks data for a nutrient: the response says which foods are missing.
- Household weights are POF/IBGE averages, not your plate. Only foods and measures whose POF description matches the food were mapped; anything else is reported as not covered.

## Mini-evaluation

15 questions about the TACO, answered by `claude -p` with and without the server, 3 repetitions each (90 runs per model, 30 October 2026), on three models. A numeric answer is correct when it is within 5% of the table value; for the two "honesty" questions it is correct only if it says the TACO does not publish the value (whole milk energy) or that the total is a minimum (fiber of a meal that includes chicken, which has no fiber value).

| Category        | Haiku 5.5 without / with | Sonnet 5.5 without / with | Opus 5.5 without / with |
| --------------- | ------------------------ | ------------------------- | ----------------------- |
| Direct lookup   | 12/18 / 18/18            | 16/18 / 18/18             | 18/18 / 18/18           |
| Scaled grams    | 9/9 / 9/9                | 9/9 / 9/9                 | 9/9 / 9/9               |
| Comparison      | 3/3 / 3/3                | 3/3 / 3/3                 | 3/3 / 3/3               |
| Household units | 3/6 / 6/6                | 1/6 / 6/6                 | 3/6 / 6/6               |
| Whole meal      | 3/3 / 3/3                | 0/3 / 3/3                 | 3/3 / 3/3               |
| Honesty         | 1/6 / 6/6                | 3/6 / 6/6                 | 3/6 / 6/6               |
| **Total**       | **31/45 / 45/45**        | **32/45 / 45/45**         | **39/45 / 45/45**       |
| Cost of 90 runs | US$ 0.22                 | US$ 3.10                  | US$ 5.78                |

Reading: with the server all three models got 45/45. Without it they score 31 to 39 out of 45, and the failures differ by model. Haiku gives remembered values that are wrong for single foods (vitamin C in banana prata, iron in cooked lentils), labels them as unverified, and does not flag the fiber total of the meal as a minimum. Sonnet mostly answers "no data" on household measures and meals. Opus remembers most single values and still guesses unit weights (2 boiled eggs). All three, in all 3 runs, answer 61 kcal for whole milk, a value the TACO does not publish. One side effect: Haiku with the server repeated the TACO citation in only 13 of 45 answers (the tool output always carries it); Sonnet and Opus did it in 45 of 45. Limits: 15 questions, 3 repetitions, rule-based scoring (the 5% tolerance accepted one no-MCP Sonnet answer with a wrong rice value that landed near the right sum). Questions, answer key and raw answers are in [`eval/`](eval) (one folder per model); to repeat: `npm run build && node eval/run.mjs --model claude-haiku-5-5 && node eval/score.mjs --all --detail`.

## Development

```bash
npm install
npm run build
npm test
npm run inspect   # MCP Inspector against dist/server.js
npm run build:data -- --fetch   # regenerate data/ from the pinned upstream CSVs
```

Never write to stdout (`console.log`): it is the protocol channel. Use `console.error`. See [AGENTS.md](AGENTS.md) for the project rules.

## Data & license

Code: MIT. Nutritional data: TACO 4ª ed., NEPA/UNICAMP (2011). Household-measure weights: **Fonte: IBGE, POF 2008-2009**, _Tabela de Medidas Referidas para os Alimentos Consumidos no Brasil_ — the grams are IBGE's, unchanged; the mapping between TACO and POF foods was curated by this project, which is responsible for it. See [NOTICE.md](NOTICE.md) for sources, processing and pinned versions. If you use this in research, see [CITATION.cff](CITATION.cff). Not medical or nutritional advice.