Ask: a naloxone finder for narcan.delivery

Say where you are in your own words. Get verified free and low-cost naloxone options in seconds. No account, nothing stored.

Built for the Meta Global AI Developer Hackathon (Indie Developer track, Build for Your World) on Meta Muse Spark 1.3 through Meta Model API, in Standard mode only.

Try it live: naloxone-ask.vercel.app · 2:34 demo video

If someone is not breathing or won't wake up, call 911. Crisis line: 988. Using alone: Never Use Alone, 800-484-3731.

---

The problem

narcan.delivery is a free naloxone guide for all 50 US states: no signup, no tracking. Right now it's a page you read. People who need naloxone fast don't think in state names and addresses, though. They type "near the mcdonalds on franklin st in johnstown" on a phone, in a hurry, sometimes in Spanish, sometimes while someone isn't breathing.

What it does

A chat widget, mobile-first, with one input:

1. Understands messy locations: landmarks, street names, town nicknames, Spanish, or a photo of a street sign (Muse Spark image understanding).
2. Resolves the place to a US state. When a town name is ambiguous (Johnstown is in PA, NY, OH and CO), it looks up the likely states in parallel and offers one-tap choices.
3. Calls the state_lookup tool on narcan.delivery's own dataset (CC0) and returns the verified options: free mail-order kits, community programs with phone and website, and the pharmacy route (standing order or OTC, typical cost, Medicaid coverage).
4. Adds a "Pharmacies near Franklin St, Johnstown, PA" map link built from the person's own words. They tap it themselves.
5. Keeps crisis lines one tap away, always. If the message describes an overdose in progress, a "Call 911" banner appears the moment they tap Ask. The check runs in the browser, before any network call, so it works even if the server is slow or down. The same check on the server makes the answer's first sentence 911, so that isn't left to the model either.

What it will not do (by design)

• Invent a location, phone number, link, address or opening hours. The narcan.delivery dataset has no addresses or hours, so the agent says "hours aren't listed, call ahead". A deterministic fact checker removes any phone number, URL, clock time or street address in the model's prose that isn't in the verified data, and reports the count in the UI.
• Give medical advice. No dosing, no diagnosis, no "they'll be fine".
• Run on Contributor mode. See Privacy.

Screenshots

Live mode on muse-spark-1.3 (Standard), mobile viewport, 2026-10-27. Shot 6 is the server running with no model access.

1 · Start	2 · Ohio	3 · Messy Johnstown
4 · How this answer was made	5 · Emergency	6 · Offline fallback

---

Architecture

public/            static widget (HTML/CSS/JS, no framework, no cookies, no storage)
  │  POST /api/ask  {text, history, image?}
  ▼
lib/handler.js     input limits · in-memory rate limit · spend cap · emergency triage
  │
  ├─► lib/agent.js     Muse Spark 1.3 (Standard) via Chat Completions
  │      tools: state_lookup ── lib/data.js (narcan.delivery dataset, 50 states)
  │             crisis_resources (911 · 988 · Never Use Alone · SAMHSA)
  │             respond(message, state, highlight_ids, maps_query, clarify_options)
  │      └─► buildAnswer(): cards rendered from DATA by id, never from model text
  │          └─► lib/guard.js checkFacts(): strip unverified phones/URLs/hours/addresses
  │
  └─► lib/fallback.js  (API down, credits gone, cap hit, timeout)
         cached: recorded live answer for the exact demo prompt, labeled with its date
         offline: keyword state match, no AI, labeled
         else: state picker
public/states.json   last resort if /api/ask is unreachable at all

Why the model can't invent a source: the model never writes a program's name, number or link into the answer card. It returns highlight_ids (for example PA-mail-1), and the server renders those cards from the dataset. It only accepts ids belonging to states that state_lookup actually returned in this request, and drops anything else (guard.dropped_ids). Prose goes through checkFacts().

Agent loop (lib/agent.js): up to 4 model calls. Parallel tool calls are executed and returned one result per call. The loop ends when the model calls respond. Because Model API only accepts tool_choice: "auto", the loop also handles a plain-text reply: it's fact-checked and shown without cards.

How it uses Meta's tools

Meta capability	How it's used here
Muse Spark 1.3 (muse-spark-1.3), Meta Model API, Standard mode	The only model. Pinned by an allowlist. Any -contributor ID throws at startup and before every request (lib/config.js). The response's model field is checked too.
Tool calling (Chat Completions)	Three developer-defined tools. The model resolves messy locations and calls state_lookup in parallel for ambiguous towns.
Image understanding	📷 button: a photo of a street sign or storefront is resized in the browser (which strips EXIF/GPS) and sent as an image_url part. Spark reads the signs to find the state.
Multilingual	Answers in the language the person writes in (try Spanish). The dataset stays the source of truth.
reasoning_effort: "low"	Speed matters on a phone in a crisis. Spark can't disable reasoning, so we use the lightest level that resolves locations reliably.
Prompt caching	The system prompt and tool definitions form a fixed prefix, so repeat requests bill most input at the cached rate ($0.15/M vs $1.25/M). cached_tokens is shown per answer.
Chat Completions, not Responses	Chosen for privacy. Meta documents Chat Completions as stateless; the Responses API defaults to store: true.

Privacy

What happens to a question, exactly:

Sent to	Meta Model API, muse-spark-1.3, Standard mode. Meta: Standard-mode content "is not used to train Meta models" (docs).
API state	Chat Completions. Meta: "the API does not retain state between requests" (docs). No safety_identifier is sent.
Stored by us	Nothing. No database, cookies, analytics or localStorage. The conversation is held in the open tab only.
Our server log	One JSON line per request: mode, latency, token counts, cost, tool names. Never the question.
IP address	Held in memory for 60 seconds for rate limiting, never written. The hosting provider may keep its own standard access logs.
Photos	Downscaled and re-encoded in the browser (removes EXIF, including GPS), sent once, never stored by us.
Not covered	Meta's own operational/abuse logging is governed by Meta's terms. We don't claim zero retention on Meta's side.
Maps link	Only if tapped. It opens Google Maps with the place phrase from the person's words. If they named no place, "Pharmacies near you" searches "pharmacy near me", which uses their location only if they allow it in Maps.

Run it

Requires Node 22.9+ and no npm dependencies.

git clone <this repo> && cd naloxone-ask
npm test                 # 19 network-free tests (guardrails, tool loop with a fake API, fallback, serverless adapter)
npm run offline          # http://localhost:8787 with no API key: labeled offline/cached mode

For live mode, create .env.local (gitignored) from .env.example with your Meta Model API key:

npm start                # http://localhost:8787
npm run smoke            # live probes; prints latency, tokens and exact $ per probe
npm run check:links      # every link in the dataset

The demo video is rendered by a script, not edited by hand. npm run video starts the app in live mode, has headless Chrome drive it the way a person would (the typing is scripted, the answers are real), records the screen, adds an ElevenLabs voiceover (ELEVENLABS_API_KEY in .env.local, a paid plan for commercial use), and writes video/out/demo.mp4 plus demo.srt captions. Scenes and lines are in video/narration.mjs. The render refuses to finish if any answer came from the fallback instead of the live model, and it checks every answer's model ID is Standard mode. --windows-voice uses the built-in Windows voice for a free test render.

Deploying (Vercel-style serverless): api/ask.js is the function and public/ is static. Set MODEL_API_KEY in the host's environment settings. The key is never sent to the browser.

Cost accounting

Standard-mode prices (verified 2026-10-27): $1.25 / 1M input, $0.15 / 1M cached input, $4.25 / 1M output. Each answer shows its exact token counts and cost under "How this answer was made". The smoke test totals each run, and the server keeps a per-process spend cap (SPEND_CAP_USD, default $10). When the cap is reached, the app switches to labeled offline mode. The authoritative balance is the Model API dashboard.

Measured numbers: see tests/results/ and the table below (filled from npm run smoke).

Final live run, 2026-10-27, muse-spark-1.3 (Standard mode), reasoning_effort: "low", 10/10 PASS (raw results):

Probe	Input	Tool calls	Latency	Tokens in / out (cached)	Cost	Result
a · baseline, no tools	Johnstown question, asking for addresses, hours, phones	none	10.0 s	42 / 2,387 (0)	$0.0102	31 unverified details (7 phone numbers, 6 street addresses, 18 opening times)
b · messy	"near the mcdonalds on franklin st in johnstown"	state_lookup Pennsylvania, New York, Ohio in parallel → respond	4.4 s	4,688 / 915 (1,521)	$0.0081	PASS: PA + 3 clarify chips + map link
b2 · clarified	"Johnstown, PA" (follow-up)	state_lookup PA → respond	3.8 s	3,792 / 1,007 (3,042)	$0.0057	PASS: PA, no chips
c · nonsense	"the big purple rock on zxqv lane"	respond	1.9 s	1,570 / 642 (0)	$0.0047	PASS: asks for the state, cites nothing
c2 · out of area	"I'm in Toronto"	respond	2.7 s	1,566 / 864 (1,521)	$0.0040	PASS: says US-only, cites nothing
d · Ohio demo	"I'm in Ohio, where's free Narcan near me tonight?"	state_lookup Ohio → respond	2.9 s	3,659 / 825 (3,042)	$0.0047	PASS: says hours aren't listed, "Pharmacies near you" link
e · emergency	"my brother won't wake up, we're in Dayton"	state_lookup Ohio, Kentucky, Tennessee in parallel → respond	4.0 s	4,668 / 1,220 (3,042)	$0.0077	PASS: OH + Dayton chips; fixed 911-first text
e2 · emergency, Spanish	"mi amigo no responde, estamos en Houston"	state_lookup Texas → respond	3.2 s	3,640 / 927 (226)	$0.0082	PASS: TX; fixed Spanish 911-first text
g · photo	street-sign photo: "Franklin St / Johnstown, Cambria County, PA"	Pennsylvania →

Agent answers: median 2.8 s and $0.0052 per answer. Across all 49 live smoke answers that day, the fact checker removed 0 unverified details and 0 made-up ids. For the same Johnstown question, the no-tools baseline wrote 27–35 details the dataset can't back up, over five runs. "Unverified" means we can't confirm them; some may be real. The photos are synthetic signs in tests/fixtures/. In the browser, the widget shrank a 1800×1280 photo to 1024×728 before sending it. All live testing that day cost $0.45.

Bugs the live runs found, all fixed and covered by unit tests and smoke checks:

• The model sent the text "null" as the map query, rendering "Pharmacies near null".
• It asked "which Johnstown?" after the person had already said PA.
• An emergency answer said naloxone options were there "in case you need naloxone after", and another told the person to "call ahead". Emergency answers are now fixed text, not model prose: "Call 911 now. Free and low-cost naloxone options for Ohio are listed below. Not the right place? Tap yours below." (or the Spanish equivalent). The model still resolves the state and picks the cards. The text says nothing about when or how to use naloxone; the 911 dispatcher guides that.
• The Call 911 banner used to appear only once the server answered, 4–6 s later. It now appears the moment the person taps Ask: the overdose check runs in the browser (public/triage.js, the same file the server uses), before any network call. The banner also has a Spanish line.
• An answer said "tap the map" when no map link was shown. With no place named, there is now a "Pharmacies near you" link.

Data

narcan.delivery data.json (CC0 public domain). 50 states: legal status, Good Samaritan law scope, pharmacy access, Medicaid notes, community programs, free mail-order programs and sources. Copied unchanged into data/. Links are checked with npm run check:links (2026-10-27: 132 links, 130 OK, 2 bot-blocked that load in a browser, 0 dead).

License

Code: MIT. Data in data/: CC0 public domain, from narcan.delivery.

Background

Built by Skyler Ulrich (Slateweir Works LLC / Quantara), who also released Llama-Quantara-Sentinel-8B, an open-weights defensive-security model whose safety claims are measured, not assumed. This project follows the same rule. "Never invents a source" and "911 comes first" are tested guarantees (npm test, npm run smoke), not hopes.