# FAQ

## Is the catalog current?

The snapshot is refreshed when someone runs `alhorno-mcp sync`. Recipes are immutable once
published; only tags and relations change, and new recipes appear weekly. Run
`alhorno-mcp sync --check` to learn whether upstream moved, then `sync` and restart.

## Does the server use tokens?

No. It makes no model calls and holds no API key. Your MCP host spends tokens writing the answer
around the data, which is why the tools return compact summaries and a single recipe per detail
call.

## Why does `papas fritas` return nothing?

Because the website returns nothing either. A fuzzy query requires **every** significant token to
match, and no recipe matches both `papas` and `fritas`. This is not a bug in the port; try `papas`
alone, or `papas al horno`.

## Why is the tag filter case-sensitive?

The website compares tags with `===`, so `pollo` and `Pollo` are different values and the latter is
the one that exists. Call `list_tags` once and pass the names verbatim. When the tag is unknown the
server raises an error with close matches instead of pretending there are no recipes.

## Why can I not search for "2 huevos"?

Search runs over names, tags and normalized ingredient names (`huevo`, `harina`), not over quantity
lines. For a pantry query use `find_recipes_by_ingredients` with ingredient names only.

## Can I use it from ChatGPT?

Yes, through a connector. ChatGPT requires a public HTTPS endpoint, so use the hosted
`https://mcp.alhornoconpapa.com.ar/mcp` or deploy your own. See
[Client setup](../getting-started/clients.md#chatgpt).

## Does it need authentication?

No. The catalog is public and the server is read-only, so the deployment is open. If you expose
your own instance, keep the host allowlist narrow and consider a rate limit at the proxy; see
[Deploy to a VPS](../operations/deploy.md#abuse-and-rate-limiting).

## Does it store anything about callers?

No state, no cookies, no identifiers. In stateless HTTP mode each request is independent, and the
log keeps one line per tool call with a name, a duration and a result count.

## Why is `id` returned as a string?

Instagram identifiers exceed JavaScript's safe integer range. Returning a string keeps every host
language exact; the original number stays available in `raw.id`.

## Why is this Python and not TypeScript?

The runtime stays a single small Python process that deploys anywhere with Docker. Node is used
only as a development tool: `tools/goldens.mjs` imports the website's own search code to record the
parity fixtures. Node is not part of the runtime, the image or CI.

## What happens when upstream publishes a broken file?

`sync` validates before writing: duplicate ids or slugs, missing fields, invalid types or a payload
over 8 MiB abort with exit code `1`, leaving the previous snapshot untouched. A server started with
an invalid `ALHORNO_DATA_DIR` exits instead of serving partial data.

## Does it include images?

Only their URLs. `image_url` points at the website, so images keep being served (and attributed)
by the site instead of being copied into this repository.

## Can I filter by several tags?

One tag per call, exactly like the website. To combine tags, use several searches and intersect the
`slug` values, or ask for a broader query and filter the summaries.

## Why are some recipes missing?

Recipes marked `hidden: true` in the upstream document are drafts and are never served, indexed or
counted. Relations pointing at hidden recipes are skipped, which is why a recipe can return fewer
than five related entries.

## Which license applies to the data?

Code is MIT. Recipe text and images belong to
[@al.horno.con.papa](https://instagram.com/al.horno.con.papa); this project only indexes the
publicly published dataset and links back to it.
