Languages

The kit's own texts (status bar, spinner, panels, shortcuts, tool labels, turn summaries, authentication prompts) come in English, Spanish, French and German, and the agent is asked to reply in the same language.

How the language is chosen

Once per process, in this order:

1. --language=<code> on the command line (or --language <code>);
2. the agent's option: config.language, or the language option of a UI entry point;
3. the system's language: the Intl locale, then LC_ALL, LC_MESSAGES, LANG;
4. English.

Codes are en, es, fr, de; region variants work (es-ES, fr_FR.UTF-8). An unsupported code shows a one-line warning and falls through to the next source.

npx tsx agent.ts --language=fr
npm start -- --language=de     # with npm, arguments go after --

const config: BaseSessionConfig = { mode: "guided", projectDir, language: "es" };

buildSessionOptions() chooses the language (and appends the reply line); a chat, progress view or wizard started afterwards without its own language keeps it. getLanguage() returns it, setLanguage(code) sets it directly (a desktop host with its own language setting), and resolveLanguage() / detectLanguage() resolve without setting anything.

Switching on the fly

The language is still one per process, but a host can change it while the chat runs: switchLanguage(code) sets it and keeps it over --language and config.language from then on (setLanguage() alone would lose to them the next time options are built). The chat controller's setLanguage(code) does it for you: it switches, opens the session again in the new language keeping the conversation, and tells the model with the person's next message.

The agent's own texts follow only if they're chosen when the session opens, not once at startup: build the spec (prompt, names, subagents) in the session opener, from getLanguage(). Captain Whiskers does this, and keeps the person's choice in his home folder:

const chat = await createChatController(async (run) => buildSessionOptions(config, run.dir, makeSpec(), { run }), { runsDir, panels: "state" });
await chat.setLanguage("fr", { note: "Your name is now Capitaine Moustaches." });

Even so, the earlier conversation pulls: the replies switch at once, but the model keeps a name it already used in another language unless told. The notice asks it to use the names its instructions give now, and note says which: without it, the captain answered in French as "Capitaine Whiskers"; with it, as "Capitaine Moustaches".

The reply language

buildSessionOptions() appends one line to the system prompt and to every subagent's prompt:

Language: use French both to reply and for everything you keep (knowledge base pages, preferences, notes, the log), whatever language these instructions, your skills, your sources or the human's name are in. Switch both to another language only when the human's message is clearly written in it; a greeting, a slash command or a single word doesn't count.

So the agent answers, and writes its knowledge base, in the chosen language, and follows the person when they clearly write in another one. Turn it off with replyInLanguage: false in the spec when your prompt decides the language itself.

Limits

The line is a strong hint, not a guarantee. Checked by hand:

• The person's name pulls what the agent writes towards its language: the CLI always puts the account's name in the context, and a tool asking for something "in the person's language" got Spanish pages from an English chat with a Spanish name. The kit's tools say "in the language you reply in"; say the same in yours.
• Text in another language pulls the replies towards it: the agent's prompt, its skills, its commands, its subagents' names. Write everything the model reads in English (or in the language you want) if the agent should follow the chosen one.
• The runner's identity: the CLI puts the logged-in Claude Code account's name and e-mail in the context, and the model may take the runner's language from them, most for questions about the agent itself. The kit already drops the git context (the runner's git user name) for this reason; the account can't be removed.
• English messages don't always switch the language: the model tends to take English as neutral.

An agent's own texts

Texts the agent passes to the kit (header, welcome message, prompt label, approval texts, tool labels of its own) are its own: the kit doesn't translate them. To have them follow the language too, pick them with detectLanguage(), as Captain Whiskers does:

import { detectLanguage, messagesFor, type Language } from "@falkenslab/agent-kit";

const { language } = detectLanguage(); // --language, else the system's

const TEXTS: Record<Language, { name: string; you: string; welcome: string }> = {
  en: { name: "Captain Whiskers", you: "you>", welcome: "Captain Whiskers is aboard." },
  es: { name: "Capitán Bigotes", you: "tú>", welcome: "El Capitán Bigotes ha subido a bordo." },
  fr: { name: "Capitaine Moustaches", you: "toi>", welcome: "Le Capitaine Moustaches est à bord." },
  de: { name: "Käpt'n Schnurrbart", you: "du>", welcome: "Käpt'n Schnurrbart ist an Bord." },
};
const text = TEXTS[language];

await runChatInk(opener, {
  runsDir,
  header: { title: text.name, fields: { mode: messagesFor(language).mode(config.mode) } },
  promptLabel: `${ui.user(text.you)} `,
  welcomeMessage: text.welcome,
});

messagesFor(language) gives the kit's own texts in a language (the mode names above), with English for anything missing.

What stays in English

Text for the model: tool descriptions, hook denial reasons, the knowledge base section, the reply line itself (but for the language's name). Several are tuned to SDK behavior confirmed by hand, and English is what the model follows most reliably.

Adding a language

The catalogs live in src/core/messages/ of the kit, one file per language, typed against the English one; a test checks every catalog has every key. Contributions are welcome: add the catalog and its code to SUPPORTED_LANGUAGES, and check the busiest status bar still fits 80 columns.