AI Translation Tool with Human-in-the-Loop: 23 Agent Tools for Translation Quality

AI-Powered Translation Agent in Better i18n

Better i18n goes beyond simple AI translation. The dashboard includes a full conversational AI agent — powered by Google Gemini — that understands your project context and executes translation tasks through natural language. Ask it to translate a namespace, review inconsistencies, or add a new language, and it handles the workflow end-to-end with your approval at every step.

This is not a "generate and hope" approach. The agent operates with 23 specialized tools, human-in-the-loop approval for every write operation, and a 30-second project context cache that keeps it grounded in your actual data.

The AI Agent Architecture

The AI chat system is built as a tool-using agent, not a simple prompt-and-response interface. When you type a request, the agent:

1. Reads your project context — languages, keys, translations, sync status, content models
2. Plans the operation — determines which tools to call and in what order
3. Executes with approval — read operations happen automatically; write operations require your explicit confirmation
4. Streams results progressively — translation tables render in real-time as results arrive

The agent supports up to 50 steps per conversation, with context stripping (slimToolResults) that prevents token overflow on large operations.

23 Specialized Tools

The agent has access to 23 purpose-built tools, divided into reading tools and writing tools.

Reading Tools (Automatic Execution)

These tools run without requiring approval, giving the agent full visibility into your project:

• getTranslations — fetch translations for specific keys, namespaces, or languages
• getKeyDetails — inspect individual key metadata, context notes, and translation status
• getLanguages — list all configured languages and their completion percentages
• getProjectStats — project-wide translation coverage, key counts, and language statistics
• getDoctorReport — diagnostic report identifying missing translations, unused keys, and inconsistencies
• getSyncs — list all configured sync integrations (GitHub, GitLab, etc.)
• getSyncDetails — inspect a specific sync's configuration and recent activity
• getContentModels — list content models and their field structures
• getContentEntries — browse content entries across models
• createPlan — generate an execution plan for complex multi-step operations

Writing Tools (Human-in-the-Loop Approval)

Every writing tool requires your explicit approval before execution. The agent proposes changes, you review them, and only then do they take effect:

• proposeTranslations — generate new translations for missing language pairs
• proposeTranslationEdits — suggest improvements to existing translations
• translateBatch — bulk-translate keys across multiple languages in one operation
• proposeKeys — suggest new translation keys based on your codebase patterns
• proposeDeleteKeys — identify and propose removal of unused or duplicate keys
• proposeLanguages — suggest new languages to add based on your user demographics
• proposeLanguageEdits — modify language configuration (display names, fallbacks)
• publishChanges — push approved translations live to the CDN
• proposeContentEntries — create or update CMS content entries
• proposeContentModel — suggest content model schema changes
• proposePublishEntries — queue content entries for publishing

Human-in-the-Loop: The Key Differentiator

Most AI translation tools operate on a "generate and ship" model. Better i18n takes a fundamentally different approach with human-in-the-loop (HITL) approval.

How HITL works in practice:

1. You ask the agent: "Translate all missing keys in the settings namespace to German and French"
2. The agent reads your project, identifies 47 missing translations, and generates them
3. A structured proposal appears in the chat — showing each source string, the AI translation, and any glossary terms applied
4. You review the proposal: approve all, approve selectively, or request changes
5. Only approved translations are written to your project

This workflow ensures that AI translation quality meets your standards before anything reaches production. The agent never writes data without your consent.

Why this matters:

• No silent errors — every AI-generated translation is reviewed before it goes live
• Glossary enforcement — the AI respects your terminology definitions, and you verify it did
• Audit trail — every approval is logged with who approved it and when
• Reversibility — since you approve each change, there is no batch of unknown quality sitting in production

Progressive Rendering

When the agent generates translations, results stream into the chat interface in real-time. The translation table renders progressively — you see each language column fill in as the AI produces output, rather than waiting for the entire batch to complete.

This is particularly valuable for large operations. When translating 200 keys across 8 languages, you can start reviewing the first results while the agent is still generating the rest.

Chat History and Persistence

The agent maintains conversation history through a dual-storage system:

• IndexedDB (local) — provides instant access to recent conversations with zero latency
• Postgres (server) — maintains a persistent audit trail of all agent interactions

This means your conversations load instantly when you return to the dashboard, while also being available for team audit and compliance purposes.

Practical Examples

Adding a New Language

You: "Add Japanese to the project and translate all existing keys"

The agent calls getLanguages to check current configuration, proposeLanguages to add Japanese, then translateBatch to generate translations for all existing keys. Each step shows you what is happening, and write operations wait for your approval.

Finding Translation Issues

You: "Are there any inconsistencies in the French translations?"

The agent calls getDoctorReport and getTranslations, then analyzes the results for terminology inconsistencies, missing plural forms, or untranslated keys. It presents a summary with specific keys that need attention.

Bulk Translation Review

You: "Show me all untranslated keys in the checkout namespace for Spanish"

The agent calls getTranslations with the appropriate filters and presents a table of missing translations. You can then ask it to generate translations for all of them, review the proposals, and approve in bulk.

Model and Context Management

The agent uses Google Gemini as its underlying model, selected for its strong multilingual capabilities and context window size. Key technical details:

• 50 steps per conversation — sufficient for complex multi-step operations
• Context stripping — the slimToolResults system automatically reduces token usage by trimming large API responses to essential data
• 30-second project cache — project context is cached for 30 seconds, preventing redundant API calls during multi-step operations

Getting Started with the AI Agent

1. Open the Better i18n dashboard and click the AI chat icon
2. Start with a read operation: "Show me the translation status for this project"
3. Try a translation task: "Translate all missing keys in the common namespace to Spanish"
4. Review the agent's proposals and approve or edit as needed
5. Use "Publish changes" when you are ready to push translations live

The AI agent is available on all plans. Create your account to try it with your project.