Releases: Shipi18n/shipi18n
Release list
v2.15.0
npm: @shipi18n/cli@2.15.0. Single file (no npm install): shipi18n.mjs below, with its SHA-256 and a build-provenance attestation. Changes: https://github.com/Shipi18n/shipi18n/blob/v2.15.0/packages/cli/CHANGELOG.md
v2.14.2
npm: @shipi18n/cli@2.14.2. Single file (no npm install): shipi18n.mjs below, with its SHA-256 and a build-provenance attestation. Changes: https://github.com/Shipi18n/shipi18n/blob/v2.14.2/packages/cli/CHANGELOG.md
v2.14.1
npm: @shipi18n/cli@2.14.1. Single file (no npm install): shipi18n.mjs below, with its SHA-256 and a build-provenance attestation. Changes: https://github.com/Shipi18n/shipi18n/blob/v2.14.1/packages/cli/CHANGELOG.md
v2.14.0
npm: @shipi18n/cli@2.14.0. Single file (no npm install): shipi18n.mjs below, with its SHA-256 and a build-provenance attestation. Changes: https://github.com/Shipi18n/shipi18n/blob/v2.14.0/packages/cli/CHANGELOG.md
v2.13.1
npm: @shipi18n/cli@2.13.1. Single file (no npm install): shipi18n.mjs below, with its SHA-256 and a build-provenance attestation. Changes: https://github.com/Shipi18n/shipi18n/blob/v2.13.1/packages/cli/CHANGELOG.md
v2.6.1 — OpenAI adapter fix + MCP desktop bundle
core 2.6.1: the OpenAI adapter now sends max_completion_tokens to api.openai.com (gpt-5.x rejected max_tokens) with a one-shot fallback; OpenAI-compatible endpoints unchanged. Found by the UIStringBench leaderboard runs. Attached: shipi18n.mcpb — the MCP server as a one-click desktop extension bundle (validators need no API key).
v2.5.0 — local models via --base-url, semantic QA hardening
Releases 2.1.0 → 2.5.0 in one catch-up (npm always had them; GitHub releases lagged).
Highlights
--base-urloncheckandtranslate(2.5.0) — run against any OpenAI-compatible endpoint with-p openai: Ollama locally with no API key at all, Gemini's compatibility endpoint, Groq, LM Studio, vLLM. The judge's published accuracy was measured onclaude-haiku-4-5; runevals/semantic/run.mjsbefore trusting a different judge.check --semantic(2.2.0) — the LLM-as-judge pass: advisory by default, structural-first, verdicts cached so unchanged strings cost zero model calls,--glossaryenforced deterministically.shipi18n lock(2.3.0) — protect hand-edited translations; clobber/stale reported as warnings only.- Better failure modes (2.4.0) —
judged 0explains itself; the missing-SDK error names the exact install that works undernpx.
Full details per package: cli · core
Published on npm with OIDC trusted publishing and SLSA provenance.
v2.0.0 — bring your own LLM
The v2 line drops the hosted Shipi18n API entirely. Every package now calls your own LLM with your own key — no account, no per-word fees, no strings leaving your machine except to the provider you choose.
Packages
| Package | Version |
|---|---|
@shipi18n/core |
2.0.0 — new |
@shipi18n/cli |
2.0.0 |
@shipi18n/mcp |
2.0.1 — new |
vite-plugin-shipi18n |
2.0.0 |
Quickstart
export ANTHROPIC_API_KEY=sk-ant-... # or OPENAI_API_KEY
npx @shipi18n/cli translate locales/en.json -t es,fr,deHighlights
@shipi18n/core— the engine, extracted and open-sourced. Provider-agnostic adapters (Anthropic, OpenAI, or any{ complete }object), structure-preserving JSON translation, placeholder validation and incremental mode.@shipi18n/mcp— translate from Claude Desktop, Cursor or any MCP client. Works with no API key at all via MCP sampling, where the client's own model does the translation.- Placeholder-safe —
{{name}},{count},%s,%d,%1$s,$t(...),%{name}and HTML tags are preserved and validated after every translation. - All packages are Apache-2.0 and live in this monorepo.
Breaking changes
- The Shipi18n
apiKey/apiUrloptions are gone. Passproviderplus your own LLM key (or setANTHROPIC_API_KEY/OPENAI_API_KEY). shipi18n keysandshipi18n configare removed — there is no account to configure.- The GitHub Action drops hosted LLM verification, self-correction and dashboard sync; local placeholder/key/length verification is kept.
Note
@shipi18n/mcp@2.0.0 was deprecated immediately: it exited silently when launched through its bin. Use 2.0.1.