lingo

Make forms easier, LLM tools safer.

People type fast and loose: 180cm, 5ft 11, 1.5 cups, 90 min,
next friday, 1,5 kg, typos included. Models emit fluent, over-formatted
strings: "5'11\"", "1½ cups", "twenty-five kg", "3pm EST",
"1,234 kg". Your database wants canonical values: one number in one unit,
one ISO date. The library is the layer in between: natural-language
quantities, units, dates, and ranges parsed into validated values, then
humanized back. Zero dependencies; English-first parsing with locale-aware
formatting (locale packs are on the roadmap, plan 013).

For the web. Use one text field instead of a value box plus a unit dropdown:

• Accepts how people actually type: 180cm, 5ft 11, 2 lb 3 oz,
  an hour and a half, typos with did-you-mean (5 meterz), and fuzzy words
  (it's hot) where a field opts in.
• Validation UX ships as data: fields are never yelled at mid-typing (2 f is
  incomplete, not invalid), and every issue carries a stable code, a human
  message you can override, a span into the original input, and did-you-mean
  suggestions.
• No silent guesses: ambiguous input (1,234: 1.234 or 1234?) returns a
  deterministic best reading plus ranked alternatives and a warning; confirm
  strictness turns each assumption into a one-click confirmation.
• Two-way and tested: whatever format()/humanize*() emits re-parses to the
  same value. 1.9999 m formats as 6′7″, never 5′12″.
• Headless DOM controller, <lingo-input> element, and React hook. No styles
  shipped, no framework lock-in, and parsing is microsecond-scale, fast enough
  to run on every keystroke.

For LLMs. Models are better at emitting "5'11\"" than 1.8034; the
/ai fields make the string the reliable path:

• Standard Schema fields (quantityField, dateField, lingoObject) drop
  into AI SDK tools, OpenAI/Anthropic/Gemini strict modes, LangChain, and MCP
  (lingoTool()). One field definition covers the provider wiring documented
  in the recipes.
• Tool-boundary defaults reject risky values: ambiguous numbers, ignored timezones,
  and clock-drifting relative dates fail loudly with [CODE] ... Did you mean ...?
  messages a model can self-correct from in one round trip.
• Measured on a 160-fixture recorded canonicalization corpus (not an LLM
  benchmark): naive Number()/new Date() coercion accepts 17.5% of it and
  is silently wrong on 6.3%; lingo accepts 96.9% with zero silent-wrongs.
• Deterministic and replayable: pass an explicit now and results are exactly
  reproducible. The /ai date field requires it for relative dates, so a
  queued or retried tool call never drifts across midnight.
• repairToolCallWith/canonicalizeValues fix malformed payloads client-side,
  with no extra model call.

One parser powers both sides. The same lingoObject can validate a model tool
call and a human form through Standard-Schema form resolvers (React Hook Form,
TanStack Form, ...).

This is the monorepo. The library README with install, API tour, and recipes lives
with the package: packages/lingo.

Layout

Quick start

bun install
bun run check     # typecheck + tests + build + size budgets + corpus gate + zero-deps gate

Develop (library + site together):

bun dev           # tsup --watch on the library + site dev (port 3000, auto-increments)
bun kill          # stop all lingo dev processes (ports 3000-3003 + this repo's watchers)
bun restart       # kill + clear .next*/.turbo caches + bun dev

The site consumes the library as a live workspace link, so watch rebuilds show
up in the running site. (cd apps/site && bun dev still works for site-only
work; bun run site:sync refreshes served data files like llms.txt.)

Contributing

See CONTRIBUTING.md. The short version: bun run check must
be green, zero runtime dependencies is non-negotiable, specs in plans/ are
living documents, and notable changes land in the CHANGELOG under [Unreleased].

License

MIT