stillshort

An MCP server that follows medicines in the FDA drug shortage register and answers the only question anybody asks twice, has anything changed since I last looked.

python demo.py --offline     the frozen records, repeatable
python demo.py               the live register
python -m stillshort         the server on stdio, for a client that starts it
python -m unittest discover -s tests -t .

What it does

Three tools, because a voice interface cannot hold more than three verbs.

watch      follow a substance, and hear where it stands now
watching   read the list back, forget one, or forget all of it
changed    what moved since each one was last read

The register keys a shortage on generic_name, and that field carries the dosage form inside the same string, Clonazepam Tablet rather than Clonazepam. Nobody says that out loud, so a spoken name is matched against the leading substance and the forms that came back are reported separately.

A substance can hold two hundred lines, one per package. Reading them is not an answer. What comes back is one reading per status, how many packages, how many suppliers, and the most useful note the register carries.

Related MCP server: Recall Radar MCP Server

What it refuses to do

It does not suggest a substitute, it does not say whether to worry, and it does not translate a status into advice. Every answer that carries a shortage carries the register's own words back with it, verbatim,

Do not rely on openFDA to make decisions regarding medical care. While we make every effort to ensure that data is accurate, you should assume all results are unvalidated.

When the register cannot be reached, it says so and adds nothing to the list. A silent server is worse than one that admits it could not look, because the caller reads silence as good news.

Where the data live

The watchlist is plain JSON under ~/.stillshort, or wherever STILLSHORT_HOME points. Nothing leaves the machine, nothing is encrypted, and watching with forget all empties it in one call. An encrypted file the owner cannot read is worse than a readable one they can delete.

Heard, on a simulated Alexa+ display

python talk.py
python sim/film.py

talk.py holds a real spoken conversation with the server. A neural voice says each line, Whisper transcribes it, a language model client that can reach nothing but the three stillshort tools decides what to call over MCP on stdio, and a second voice reads the answer. The last take is in take/transcript.json, tool calls and register answers included. Whisper wrote clonazapam, and the client still called watch with clonazepam.

sim/film.py plays that take on sim/device.html, a smart display with a light bar, a caption of what it heard, the spoken reply and a card for the result, with every tools/call and its outcome beside it. The caption, the reply, the card's contents and the tool calls are read from the take, not written for the video. The only edit is the model's thinking time, cut to a fixed pause. One take cannot show the register moving, so its last answer is Nothing has changed, which is true and is what the tool is for.

Two things the register does that will bite the next person

related_info is one free text field carrying two unrelated things. Estimated supply: August 2026 is an answer. Marketed by Advagen Pharma Ltd. is not, and read out loud it sounds like one. A supply note wins whenever there is one.

update_date is written MM/DD/YYYY, which does not sort as text. 12/01/2025 sorts above 01/05/2026 character by character, so comparing these as strings picks the wrong latest entry, and only across a year boundary, which is when nobody is looking. There is a test for exactly that.

Fixtures

tests/fixtures holds real answers from the register, saved 2026-09-05, so the shapes under test are the shapes the service really sends. The bench never touches the network.