mise-en-place 1.0.0

Extract structured recipes from YouTube cooking videos with Claude, stored in local SQLite.

Mise En Place (mep)

A personal CLI that turns cooking content into a searchable recipe database. Point it at a YouTube video, a recipe web page, a text file, or pasted text; it extracts a structured recipe (with Claude, or directly from a page's embedded recipe data when available) and stores it in local SQLite. Everything stays on your machine.

Install

git clone https://github.com/AveryClapp/MiseEnPlace.git mep && cd mep
python -m venv .venv && source .venv/bin/activate
pip install -e .

Or simply

pip install mise-en-place

Setup

mep init

This creates ~/.mep/, prompts for the keys, and builds the database at ~/.mep/mep.db.

• An LLM key (required for extraction): set ANTHROPIC_API_KEY (https://console.anthropic.com/ → API Keys) or OPENAI_API_KEY. For OpenAI, also install the extra: pip install 'mise-en-place[openai]'.
• YouTube Data API v3 key (only needed for --channel ingestion): see below.

Provider selection: if you set only one of the two LLM keys, that provider is used automatically. If you set both, it defaults to Anthropic; set LLM_PROVIDER=openai (or choose it at the mep init prompt) to pick OpenAI. An explicit LLM_PROVIDER always wins.

Keys are stored in ~/.mep/config.json. You can also set ANTHROPIC_API_KEY, OPENAI_API_KEY, LLM_PROVIDER, or YOUTUBE_API_KEY as environment variables, which override the config file. EXTRACTION_MODEL overrides the default model for the chosen provider.

Getting a YouTube Data API v3 key

1. Go to the Google Cloud Console.
2. Create a project (top bar → project dropdown → New Project).
3. In the search bar, open YouTube Data API v3 and click Enable.
4. Go to APIs & Services → Credentials → Create Credentials → API key.
5. Copy the key. (Optional: Edit API key → Restrict key → YouTube Data API v3.)

Single-video adds use YouTube's public oEmbed endpoint and do not need this key. It is only required to walk a channel's uploads.

Usage

mep add https://www.youtube.com/watch?v=VIDEO_ID    # a YouTube video
mep add https://www.seriouseats.com/some-recipe      # a recipe web page
mep add recipe.txt                                   # a local text file
mep add --text "2 cups flour, 1 egg, ..."           # pasted recipe text
mep add --channel @JKenjiLopezAlt --limit 10        # latest 10 from a channel
mep add --channel @JKenjiLopezAlt                    # whole channel

mep search "garlic confit"                           # full-text search
mep list                                             # browse, newest first
mep list --tag italian --limit 20                    # filter by tag

mep discover                                         # pick a random recipe
mep discover --type dinner --healthy                 # a random healthy dinner
mep discover --indulgent                             # something to pig out on
mep discover -i chicken -i garlic -n 3               # 3 that use both
mep classify                                         # backfill meal type + health
mep show 42                                           # full recipe
mep show 42 --servings 8                              # scale ingredient amounts
mep show 42 --macros                                 # estimated nutrition breakdown
mep show 42 --check                                  # flag likely missing steps/gaps
mep set-servings 42 4                                 # record how many it makes
mep export 42                                         # print as Markdown (or -o file.md)
mep delete 42                                          # remove a recipe (asks first; -f to skip)
mep shopping-list 42 7 13                              # one combined grocery list

mep plan 42                                          # AI cooking timeline (experimental)
mep plan 42 --servings 8                             # ...scaled to 8 servings
mep cook 42                                          # step-by-step walkthrough (experimental)

mep show 42 --parts                                  # what each ingredient is for
mep adapt 42                                         # rewrite around what you have (interactive)
mep adapt 42 --have pita --sub "yogurt=sour cream"   # ...or state it directly
mep cook 42 --have pita                              # adapt just for this cook

plan makes one Claude call to reorder a recipe's steps into an efficient timeline and caches it. Each step is tagged hands-on or hands-off, with the ingredients and equipment it uses, a named timer for waits, and a "prep this during the wait" hint. The summary shows realistic wall-clock time (hands-off waits run in the background, not added end to end). Re-run plan --regenerate to rebuild.

cook walks that timeline live: it opens with a mise en place gather + equipment list, then one step at a time. On a hands-off step, pressing Enter starts a named background timer that keeps counting while you move on to the next step (like a real kitchen timer); it rings when done. It also nudges you to preheat the oven a couple steps ahead. Ctrl-C stops cleanly and reports any timers still running.

--servings N (on show, plan, cook) scales ingredient amounts to N servings. It is best-effort and display-only: only leading quantities are scaled, vague amounts like "a handful" pass through untouched, and nothing is saved. A recipe with no recorded serving count is treated as a single serving (the batch as written), so --servings 3 simply makes 3× the recipe. Use mep set-servings <id> <count> to record the real serving count when you know it, so scaling maps to people instead.

Both plan and cook are experimental: the timings are AI estimates.

show --macros shows an estimated nutrition breakdown (calories, protein, carbs, fat) for the whole recipe and per serving. It's computed lazily on first request (one model call, then cached, so it's free afterward and costs nothing if you never ask) and is a rough estimate from the ingredients, not exact.

show --parts breaks a recipe into its components (marinade, pita, sauce…) so you can see what each ingredient is for. adapt rewrites the recipe around what you already have: pick the parts you bought or made ahead and it drops the steps and ingredients needed only to make those (keeping the steps that use them), and applies any ingredient swaps. It then offers to save the result as a new copy, overwrite the original, or discard it. cook --have/--sub does the same rewrite in memory for a single cook without saving anything. These are experimental and use Claude; the rewrite is intentionally light (it shifts and trims the recipe, it doesn't reinvent it).

discover picks a random recipe from your collection, optionally filtered. Use --type (breakfast, lunch, dinner, snack, dessert), --healthy (health score

= 7) or --indulgent (<= 4), or --min-health/--max-health for an exact range, plus -i/--ingredient (repeatable) to require ingredients you want to use. -n/--count returns more than one; with no filters it is completely random. A single pick prints the full recipe; several print as a list.

To support that, every recipe is given a meal type and a 1-10 health score (10 = lean and vegetable-forward, 1 = rich and indulgent) by a small model call at add time. They show up in mep show and drive discover's filters. Recipes added before this feature won't have them yet, so run mep classify once to backfill (or mep classify --all to redo every recipe); discover reminds you when a type/health filter could be hiding unclassified recipes.

show --check makes one model call to flag likely holes in a recipe: a step that uses an ingredient never listed, a cooking step missing an obvious time or temperature, an apparent skipped step. It only points at gaps; it never invents or fills them (anything not in the source text, like a detail shown on screen in a video, can't be recovered). The result is cached, and an empty result ("no obvious gaps") is remembered too, so a re-check is free.

Most videos are one recipe, but a video that clearly teaches several independent dishes ("3 weeknight dinners") is split into separate recipes on add, each with its own id. Sub-preparations that belong to one finished dish (a sauce, dough, or marinade) stay part of that single recipe and show up under show --parts, not as their own entries.

export prints a recipe as a portable Markdown card to stdout, or writes it to a file with -o. delete removes a recipe and everything stored with it (ingredients, steps, tags, and any cached plan, components, macros, gaps, and classification); it asks first unless you pass -f. shopping-list takes one or more recipe ids and makes a single model call to merge their ingredients into one grocery list, summing compatible amounts and grouping by aisle. The combined amounts are estimates shown only on screen; nothing in the database is normalized or changed.

Channel ingestion is idempotent: videos already stored are skipped, so you can re-run it to pick up only what's new. Non-recipe videos and videos without transcripts are stored as empty entries (not errors) so they aren't re-fetched.

Sources

mep add takes more than YouTube. Pass it any of:

• A YouTube URL: transcript to a recipe (as above).
• A recipe web page URL: most recipe sites embed their recipe as schema.org data, which mep reads directly: accurate, and usually with no extraction call at all. Pages without it fall back to extracting from the page text.
• A local text file (mep add recipe.txt) or pasted text (mep add --text "..."), for recipes from anywhere else.

Each is de-duplicated by a stable id (the video id, the normalized URL, or a hash of the text), so re-adding the same source is a no-op. mep show notes where a recipe came from. Adding a web page or text that isn't a recipe is a clean error, not a stored stub (only channel syncs keep stubs, to avoid re-fetching duds).

Cost

Everything runs on your own API key, so you pay the provider directly. The only cost is the model calls; storage and search are local and free.

• Adding a recipe is the main cost: one extraction call (empirically around $0.05 with the default Anthropic model, more for very long videos) plus a small classification call per recipe for meal type and health score. A --channel walk is just this times the number of videos. Web pages that embed schema.org recipe data skip the extraction call entirely (only the small classification call remains); text and JSON-LD-less pages cost like a video.
• On-demand features (plan, show --parts, show --macros, show --check, shopping-list, adapt, and cook with --have/--sub) each make one additional call when first used, on the same order as an extraction or less.
• Caching keeps it one-and-done. Plans, components, macros, and gap checks are stored after the first request and reused for free; search, list, show, and a cached cook never call a model at all. Features you never touch cost nothing.

Numbers are rough and depend on the provider, model (EXTRACTION_MODEL), and recipe length. OpenAI (gpt-4o) lands in a similar range.

Channels to try

Recipe-forward channels that work well (most videos are real walkthroughs with transcripts). Single-video adds need no key; the --channel walk needs a YouTube Data API key (see above).

Channel	Handle
Babish Culinary Universe	@babishculinaryuniverse
J. Kenji López-Alt	@JKenjiLopezAlt
Joshua Weissman	@joshuaweissman
Adam Ragusea	@aragusea
Ethan Chlebowski	@EthanChlebowski
Brian Lagerstrom	@brianlagerstrom
Food Wishes (Chef John)	@foodwishes

mep add https://www.youtube.com/watch?v=iErqWGwso7o   # a single Babish video, no key needed
mep add --channel @aragusea --limit 5                 # latest 5 from Adam Ragusea

How it works

source → text (YouTube transcript, web page, or pasted) → an LLM (Claude or OpenAI), or a web page's embedded schema.org recipe → JSON → SQLite. Search uses SQLite FTS5 over dish name, ingredients, and channel. Vague quantities like "a handful" are stored verbatim; nothing is normalized.

Develop

pip install -e '.[dev]'
pytest

The test suite is fully offline (no network, no API keys).