Make AI translation match your brand voice and terminology standards.

Without structured context, AI translates "Releases" as "Veröffentlichungen" in German — but it's a product name that should stay in English. This starter stores glossaries, style guides, and locale rules as structured Sanity documents, assembles them into prompts at translation time, and includes an eval framework that proves the quality impact.

• Translation metadata as content — Glossaries, style guides, and locale rules are Sanity documents that content teams manage in the Studio
• Automated stale detection — Serverless functions mark translations stale on publish, run AI analysis on what changed, and pre-translate affected fields
• Translation pane and inspector — Studio UI for reviewing translation status, applying pre-translations, and triggering new translations
• Translations dashboard — Real-time overview of coverage, gaps, and stale documents across all locales (Sanity App SDK)
• Localized frontend — Next.js app with path-based i18n routing, locale switcher, and fallback content
• Quality evals — Translate with and without context, measure the delta with deterministic checks and an LLM judge

Prerequisites: Node.js (>=20.19 or >=22.12), pnpm, a Sanity account

pnpm create sanity@latest --template sanity-labs/starters/agentic-localization --package-manager pnpm

This prompts you to select (or create) a Sanity project and dataset, then writes your SANITY_STUDIO_PROJECT_ID and SANITY_STUDIO_DATASET to .env. All workspaces in the monorepo read from this single file — prefix mappings at the bottom resolve the values to each workspace's expected env var names.

Cloning manually? Copy .env.example to .env and fill in your project ID and dataset from sanity.io/manage.

cd your-project
pnpm bootstrap

pnpm bootstrap consolidates env files, resolves the organization ID, deploys the blueprint (CORS, dataset, robot token, serverless functions), deploys the schema, generates types, seeds locale documents, and imports sample content.

# 1. Consolidate env — copy .env.example to .env if it doesn't exist, fill in project ID and dataset
cp .env.example .env

# 2. Resolve organization ID — find it at sanity.io/manage → project settings
# Uncomment SANITY_STUDIO_ORGANIZATION_ID in .env and set it to your org ID

# 3. Deploy blueprint (CORS, dataset, robot token, serverless functions)
pnpm --filter @starter/functions build
pnpm exec sanity blueprints init --stack-name production --project-id <your-project-id>
pnpm exec sanity blueprints deploy

# 4. Deploy schema
pnpm --filter studio exec sanity schema deploy

# 5. Generate types
pnpm --filter studio exec sanity schema extract
pnpm --filter studio exec sanity typegen generate

# 6. Seed locale documents
pnpm --filter studio exec sanity migration run seed-locales --no-dry-run --no-confirm

# 7. Import sample content
pnpm --filter studio exec sanity dataset import sample-data.ndjson <your-dataset> --replace

pnpm dev

Opens the Studio at localhost:3333, the translations dashboard at localhost:3334, and the Next.js frontend at localhost:3000.

Frontend env: The frontend inherits SANITY_STUDIO_PROJECT_ID and SANITY_STUDIO_DATASET from the root .env automatically. For server-side data fetching with a private dataset, add a SANITY_API_READ_TOKEN (create one at sanity.io/manage → API → Tokens).

1. Define metadata in the Studio: locales, glossaries (approved translations, do-not-translate terms, forbidden terms), and style guides (formality, tone, audience instructions per locale)
2. Assemble prompt: assembleStyleGuide() converts structured metadata into an instruction string; filterGlossaryByContent() prunes to relevant terms
3. Translate: buildTranslateParams() packages everything for the Agent Actions Translate API, including protected phrases extracted from do-not-translate entries
4. Evaluate: translate with and without context, score with deterministic checks and an LLM judge, compare the quality delta

Only glossary entries whose terms appear in the source document are injected (content-aware filtering), so prompts stay focused.

sanity.blueprint.ts              Infrastructure-as-code: dataset, CORS, robot token, functions
functions/                       Serverless automation (Sanity Functions)
  mark-translations-stale.ts       Detects source changes, flags affected translations
  analyze-stale-translations.ts    AI-analyzes what changed, pre-translates affected fields
studio/                          Sanity Studio workspace
  schemaTypes/                     Article, person, topic, tag schemas
  migrations/                      Deterministic locale seeding
packages/l10n/                   Core plugin: schemas, prompt assembly, UI, evals
  src/schemas/                     Locale, glossary, style guide, entry types
  src/core/                        Pure utilities for serverless (zero React)
  src/translations/                Translation pane, inspector, hooks
  evals/                           Quality evals: with-context vs without-context scoring
apps/translations-dashboard/     Real-time translation overview (Sanity App SDK)
apps/frontend/                   Next.js frontend with path-based i18n routing

Deploy the Studio:

pnpm --filter studio exec sanity deploy

To deploy the dashboard app (pnpm bootstrap already wrote your organization ID to .env):

pnpm --filter @starter/translations-dashboard exec sanity deploy

If you want "Open in Studio" links in the deployed dashboard to point to your production Studio, also add:

echo 'SANITY_STUDIO_URL=https://your-studio.sanity.studio' >> .env

pnpm test                    # Unit tests (schema, prompt assembly, locale utils)
pnpm --filter l10n eval      # Model evals — requires sanity login, consumes AI credits

• Sanity Agent Actions — Translate API reference
• @sanity/document-internationalization — Document-level i18n plugin
• docs/I18N_RESEARCH.md — Gap analysis between enterprise TMSes and AI translation agents