LLM providers

Every translation pipeline — translate-ui, translate-docs, translate-json, and translate-svg — sends text to an LLM through the same provider-agnostic client. Before any of those commands can run, configure at least one provider in ai-i18n-tools.config.json and set the matching API key in your environment or .env (built-in presets except Ollama). init writes a starter provider / providers block; you still must supply credentials for the active preset.

You configure which API endpoint to call once in config; all translation commands share that setup and the same SQLite cache. Which models that provider tries is a separate step: AI Models.

The CLI resolves the active provider from the top-level provider key (or the sole entry in providers when only one is configured). Each provider block lists an ordered translationModels fallback chain; built-in presets inherit baseUrl and the API-key environment variable automatically (override them per provider when needed).

Built-in providers

Preset provider keys need only translationModels — base URL and API-key env var are filled in automatically:

Provider	Base URL	API-key env var
openrouter	https://openrouter.ai/api/v1	OPENROUTER_API_KEY
openai	https://api.openai.com/v1	OPENAI_API_KEY
anthropic	https://api.anthropic.com/v1	ANTHROPIC_API_KEY
gemini	https://generativelanguage.googleapis.com/v1beta/openai	GOOGLE_API_KEY
deepseek	https://api.deepseek.com	DEEPSEEK_API_KEY
cerebras	https://api.cerebras.ai/v1	CEREBRAS_API_KEY
groq	https://api.groq.com/openai/v1	GROQ_API_KEY
mistral	https://api.mistral.ai/v1	MISTRAL_API_KEY
xai	https://api.x.ai/v1	XAI_API_KEY
nvidia	https://integrate.api.nvidia.com/v1	NVIDIA_API_KEY
alibaba	https://dashscope-intl.aliyuncs.com/compatible-mode/v1	ALIBABA_API_KEY
apifun	https://api.apikey.fun/v1	APIFUN_API_KEY
cheaperinference	https://api.cheaperinference.com/v1	CHEAPERINFERENCE_API_KEY
ollama	http://localhost:11434/v1	(none)

For any non-preset key, set baseUrl and apiKeyEnv explicitly in config.

Set the active provider's API key in your environment or .env file. The CLI auto-loads .env from the working directory without overriding variables already set in the shell. See Environment variables.

Zero data retention (Cheaper Inference)

providers.cheaperinference.zeroDataRetention is optional and defaults to off. Set it when this project must send prompts only through providers that do not keep them.

"cheaperinference": {
  "zeroDataRetention": true,
  "translationModels": ["gemini-2.5-flash", "deepseek-v4-flash"]
}

With the flag on, every chat request includes "zdr": true, and list-models, check-models, and suggest-models (including the dashboard Models tab) read GET /v1/models?zdr=true. That catalog lists the models a zero-data-retention request can use, at the zero-data-retention price. check-models treats an id missing from that catalog as invalid. The flag does not create a route for a model the catalog leaves out. Remove that id from translationModels, uiModels, or localeModels.

Leave the flag off when the Cheaper Inference account does not require zero data retention. If Data Protection is already on for the account, Cheaper Inference still refuses a model with no zero-data-retention route. The error names that model. To keep the account policy, set zeroDataRetention to true and switch to a model from the zero-data-retention catalog. To keep the refused model, turn zero data retention off on that page. The fallback chain then tries the next configured model.

Models

Choosing models, the fallback list, and suggest-models, check-models, list-models, and bench-models are on AI Models.

Multiple providers

When more than one provider is configured, set the top-level provider key to select the default. Switch per run without editing the config:

npx ai-i18n-tools translate-docs -P anthropic
npx ai-i18n-tools bench-models -P deepseek

Each provider block can define its own translationModels, optional uiModels and localeModels, maxTokens, temperature, and requestTimeout (seconds) or requestTimeoutMs. A timeout on the provider overrides the top-level requestTimeout / requestTimeoutMs. A legacy top-level openrouter block is still accepted and auto-migrated to providers.openrouter on load.

Optional pricing and modelPricing set USD per 1,000,000 tokens (inputPerMTokens and outputPerMTokens) when the provider omits usage.cost. pricing is the provider-wide default; a modelPricing entry overrides it for one model id. OpenRouter and Cheaper Inference already return a per-call cost, so leave both unset on those providers. A provider-reported cost is kept as returned. The amount is included in the translation summary, usage, and Usage & costs.

Runnable example with four providers on the same document, including sample rates: examples/multi-provider.

Further reference

• AI Models — ranked shortlist, fallback list, and the model commands.
• Configuration — provider and providers — preset table, custom endpoints, request timeouts, cost rates, Cheaper Inference zero data retention, OpenRouter-specific behaviour.
• Architecture — LLM client — how model fallback, batching, and cost reporting work internally.
• Environment variables — API-key env vars and base-URL overrides.