Lunchslot

Find a meeting time that works for everyone, without accounts or calendar access.

1. Create: pick a length (30 min by default) and, optionally, a title. Recent meetings you've created are offered as presets.
2. Share: copy the link and send it to everyone.
3. Your availability: everyone, including you, writes their availability however they like ("Tue and Thu after 2pm, avoid lunch"), pastes a scheduling link (Calendly, cal.com, zcal…), or drops a calendar screenshot. Google Calendar booking pages can't be read from the server (see below), so for those a screenshot works best. The time zone is prefilled from the browser and can be changed.
4. Best times: the page shows up to 3 best times for the group and updates as people reply.

Everyone with the link is trusted: anyone can add someone else, and see, edit or remove anyone's entry.

Free to use and MIT licensed.

How it works

• The LLM only parses. Each submission makes one call through Vercel AI Gateway, which turns the input into structured data: free ranges, busy ranges, booking-page slot starts, and soft preferences, all in local time (src/lib/ingest/llm.ts, schema in src/lib/availability/schema.ts).
• Matching is plain code. src/lib/match.ts turns each person's data into UTC intervals, timezone- and DST-aware via @date-fns/tz. It intersects them on a 15-minute grid within the range the viewer picks (next 7 days by default, up to 2 months) and scores each candidate on who can come, everyone's preferences, humane local hours and how soon it is. Three picks weigh those differently: the most people possible, even at an awkward time; a balance of everything; and the earliest decent time at least half the group can make. A time that wins several picks is shown once with all its badges. The viewer can also mark people who must attend; then only times they can all make are considered. Page views never cost an LLM call.
• Scheduling links are read without provider-specific code (src/lib/ingest/page-reader.ts). The page is rendered in headless Chromium in the participant's time zone. The reader keeps the visible text plus any JSON the page loads that contains timestamps (ISO, or Unix seconds/milliseconds near today), which is where booking tools keep their slots. Month-view pages only load the month on screen, so if the page has a "Next month" button, the reader clicks it once to load the rest of the horizon. In code, it converts timestamps to local time, drops empty values, and compresses runs of evenly spaced start times ("starts every 30 min from 2026-10-07 09:00 through 11:30"). A month of Calendly or cal.com slots then costs about 1k tokens, and the model only copies runs. Only public http(s) hosts are allowed, and private IPs are blocked. Use pnpm tsx scripts/probe-link.ts <url> to see exactly what the LLM would receive (no LLM call). Google Calendar booking pages work locally, but Google's slot API refuses requests from Vercel's servers (403). If a relay is configured (RELAY_URL, RELAY_SECRET), the reader sends just that API's requests through a small Google Cloud Function (relay/) and the page loads normally. Without one, or if the relay fails, the reader recognizes Google's error page and asks for a screenshot instead.
• Screenshots are downscaled in the browser, sent once, and never stored.

Cost and abuse controls

• Global daily LLM budget (DAILY_LLM_BUDGET_USD, default $2). Each call's real cost comes from the gateway response metadata and is added to the llm_spend table. Once the budget is used up, parsing pauses until the next UTC day.
• Per-IP rate limits (salted hash, never the raw IP) for parses and meeting creation, stored in Postgres. Rate limits and the budget are checked before a link is rendered or the LLM is called.
• Input caps: 4,000 characters, 2 images, 1 link (rendered with a 15 s limit, digest capped at 40k characters), 20 people per meeting, 1,500 output tokens (3,000 with a link), and a 20 s LLM timeout.
• Recommended backstop: set a spend limit on AI Gateway in the Vercel dashboard as well.

Model choice

pnpm eval runs scripts/eval-models.ts, which checks models against fixtures with known answers: English and Italian text, relative dates, "next week", busy exceptions, a stated timezone, a Google Calendar week screenshot, a booking-page screenshot, saved Calendly, cal.com, zcal and Google Calendar booking page digests, and a prompt injection. A full run with the default model costs under $0.01. Use ONLY=link,busy to run a subset.

Latest results for the default, google/gemini-2.5-flash-lite: 50/52 checks on the full run. The failing case passed when re-run, so the model is slightly nondeterministic even at temperature 0. Average time is about 2 s, and cost is about $0.0003–0.001 per parse. Earlier comparisons:

Model	Accuracy	Avg latency	Cost / 1k parses
openai/gpt-4.1-mini (fallback)	38/38	~5.5 s	~$0.97
google/gemini-2.5-flash	34/34*	~4 s	~$2.00
openai/gpt-5-nano (minimal reasoning)	27/38	~3.2 s	~$0.16
mistral/mistral-small	flaky structured output	~1.4 s	~$0.19

* measured on an earlier, smaller fixture set; link fixtures were added later.

Development

pnpm install
vercel link && vercel env pull .env.local   # DATABASE_URL + VERCEL_OIDC_TOKEN for AI Gateway
pnpm db:migrate
pnpm dev

• pnpm test: matching engine unit tests
• pnpm eval [model...]: model comparison (append @none / @low to set the reasoning level)
• pnpm db:generate: create a migration after changing src/db/schema.ts

Google Calendar booking links (relay)

A stopgap until people can connect their calendar directly. relay/ is a dependency-free Cloud Function that forwards requests to an allowlist of hosts (only Google's appointment API) for callers with the shared secret. It relies on Google's internal API and on Google Cloud's network not being refused too, so it can break without notice.

SECRET=$(openssl rand -base64 32)
gcloud functions deploy lunchslot-relay --gen2 --runtime=nodejs22 --region=us-east4 \
  --source=relay --entry-point=relay --trigger-http --allow-unauthenticated \
  --memory=256Mi --max-instances=3 --set-env-vars=RELAY_SECRET="$SECRET"
# Then set RELAY_URL (the function's URL) and RELAY_SECRET on Vercel.

CI and deployment

• GitHub Actions (.github/workflows/ci.yml) runs lint, type checks and unit tests on every push and PR. It makes no LLM calls and needs no secrets.
• Vercel's Git integration builds a preview for every PR and deploys main to production.
• Model evals spend AI Gateway credits, so they are never run automatically. Run pnpm eval by hand (under $0.01 per full run).

Stack: Next.js (App Router) on Vercel, AI SDK + AI Gateway, Neon Postgres via Drizzle, Tailwind CSS.