Build a Multilingual AI Chatbot (40+ Languages)

Most AI chatbot tutorials assume English. But if your users are in Brazil, Japan, or the Netherlands, that's a significant gap. This post covers how to build an OpenClaw agent that handles 40+ languages gracefully — including which models to use, how to detect language, and how to test non-English conversations.

The Core Problem: Models Aren't Equal Across Languages

Not all LLMs perform equally across languages. GPT-4o and Claude 3.5 Sonnet handle most major languages well. But if you're routing to a smaller or cheaper model for cost reasons, performance can degrade significantly for non-Latin scripts or lower-resource languages.

Here's a rough quality comparison for common non-English languages:

For languages where quality matters, stick with GPT-4o or Claude 3.5 Sonnet. For Spanish, French, German, or other major European languages, Llama 3.3 70B via Groq is a cost-effective option.

Step 1: Enable Language Detection in OpenClaw

OpenClaw can automatically detect the user's language and adjust responses. In your agent config:

agent:
  name: "SupportBot"
  language:
    auto_detect: true
    fallback: "en"
    respond_in_detected_language: true

With respond_in_detected_language: true, the agent will reply in whatever language the user writes in — no prompt engineering needed for the basics.

Step 2: System Prompt for Multilingual Agents

Your SOUL.md needs a language instruction. Be explicit:

# Language Handling

Always respond in the same language the user writes in.

If the user writes in Spanish, respond in Spanish.
If the user writes in Japanese, respond in Japanese.
If the user switches languages mid-conversation, switch with them.

Do not translate the user's message to English before answering.
Respond naturally in their language, not in translated English.

## Supported Languages
You support all languages. If you receive a language you're less
confident in, still try — but acknowledge if you're uncertain.

The last part is important. Telling the model to acknowledge uncertainty in low-resource languages is better than having it confidently produce bad output.

Step 3: Language-Specific System Prompts (Optional)

For high-traffic languages, you can define language-specific instructions:

agent:
  language_configs:
    es:
      system_prompt_append: |
        Usa un tono informal (tutear). Adapta las respuestas
        al contexto latinoamericano o español según el usuario.
    ja:
      system_prompt_append: |
        Use appropriate keigo (敬語) based on context.
        Default to polite form (丁寧語) unless the user is casual.
    ar:
      system_prompt_append: |
        Use Modern Standard Arabic unless the user writes in dialect.
        Right-to-left text will be handled automatically.

Step 4: Model Routing by Language

If you want to save costs by using cheaper models for high-resource languages but better models for others:

agent:
  model_routing:
    default: "gpt-4o-mini"
    language_overrides:
      ja: "gpt-4o"
      ar: "gpt-4o"
      ko: "gpt-4o"
      zh: "gpt-4o"
      hi: "claude-3-5-sonnet"

This lets you save money on Spanish and French (where cheaper models perform well) while using premium models for languages that need it.

Step 5: Handling RTL Languages

Arabic, Hebrew, Farsi, and Urdu are right-to-left. Most chat interfaces handle this automatically, but if you're building a custom UI, ensure your CSS accounts for it:

.message-bubble {
  direction: auto;  /* Browser detects RTL/LTR automatically */
  unicode-bidi: isolate;
}

OpenClaw's default web widget handles RTL automatically. If you're using a custom integration (Discord, Telegram, Slack), those platforms also render RTL correctly.

Step 6: Testing Multilingual Conversations

Don't just test English. Here's a testing checklist:

Basic functionality

• User writes in Spanish → agent responds in Spanish
• User writes in Japanese → agent responds in Japanese
• User switches from English to French mid-conversation → agent follows
• User writes in an unsupported/obscure language → agent handles gracefully

Quality checks

• Native speaker review for top 3 languages in your user base
• Check for unnatural phrasing (a common issue with translated English)
• Verify that product names, brand names, and technical terms aren't translated when they shouldn't be
• Test date/number formatting (12/03/2026 means different things in the US vs Europe)

Edge cases

• Mixed-language message ("Can you help me with mi cuenta?") → agent should handle naturally
• Code-switching in multilingual communities (Spanglish, Hinglish) → shouldn't break

Step 7: Localized Knowledge Base

If your agent has a knowledge base or FAQ, you need it in the right languages. Options:

1. Machine translate + human review — cheapest, acceptable for most use cases
2. Write native content per language — best quality, expensive
3. Translate on-demand — agent translates the English content when needed (slowest, less accurate)

For most SaaS products, option 1 (machine translate + review for your top 3 languages) is the right balance.

knowledge_base:
  sources:
    - path: "docs/en/"
      language: "en"
    - path: "docs/es/"
      language: "es"
    - path: "docs/fr/"
      language: "fr"
  fallback_language: "en"

When the agent detects Spanish but can't find a Spanish-language document, it falls back to English and translates the response.

Real Cost Implications

Running a multilingual agent costs more than a single-language one. Japanese and Arabic queries tend to be longer in token count than equivalent English queries (different character encoding, more verbose phrasing). Expect roughly:

• Japanese: ~1.3x token overhead vs English
• Arabic: ~1.4x token overhead
• Chinese: ~1.2x token overhead
• European languages: ~1.1x overhead

Budget for this when estimating your monthly API costs.

Self-Hosting Complexity for Multilingual Agents

Self-hosting a multilingual OpenClaw agent adds complexity beyond the base setup:

• Language detection libraries need to be kept updated
• Model routing logic requires maintenance as model capabilities evolve
• Testing across 10+ languages requires native speaker involvement

Launching on ClawPort

ClawPort handles the infrastructure side. The language detection, model routing config, and environment management are all handled through the dashboard — you configure once and it runs. No server management needed.

The $10/month plan includes unlimited language support. You bring your OpenAI/Anthropic keys, and ClawPort handles routing.

Start the 7-day free trial at clawport.io — no credit card required.