Fluent

Give your AI agents fluent multilingual support.

Fluent is an open-source control plane that takes your English-built AI agent and makes it work beautifully in every language. It uses Lingo.dev to generate locale-specific runtime guidance — glossaries, FAQ overrides, escalation rules, and system instructions — then measures the actual improvement with automated evaluations.

Your agent in English → Fluent + Lingo.dev → Your agent in every language

How It Works

1. Create an agent — Define your agent's instructions, model, and target locales.
2. Run a baseline — Fluent evaluates your agent across locales and finds where it breaks.
3. Generate fixes — Lingo.dev analyzes the failures and generates locale-specific improvements (glossary, FAQ overrides, tone, escalation rules).
4. Replay — Re-run the same evaluation with fixes applied and measure the uplift.
5. Publish — Ship the improved agent as an immutable version your app can consume.

What Lingo.dev Does

Lingo.dev is the engine behind Fluent's locale-specific fixes. When you click "Generate localized fixes", Fluent sends your agent's context to Lingo.dev, which returns:

• Localized system instructions — Tone, formality, and phrasing adapted for the target locale
• Glossary locks — Branded terms and operational language that must stay consistent
• FAQ overrides — Locale-specific answers for common questions
• Escalation rules — When and how to hand off to humans in each locale
• Forbidden terms — Words and phrases to avoid in each language

These aren't generic translations. Lingo.dev understands your agent's domain and generates guidance that actually improves answer quality.

Quick Start

# Install
bun install

# Set up environment
cp .env.example .env.local

Add your keys to .env.local:

CONVEX_DEPLOYMENT=your-deployment
NEXT_PUBLIC_CONVEX_URL=https://your-deployment.convex.cloud
LINGO_API_KEY=your-lingo-dev-key        # Powers locale fix generation
OPENAI_API_KEY=your-openai-key          # For agent invocation

# Start
bun run dev

Use in Your React App

Install the SDK:

npm install @fluent/sdk

Drop in the chat component:

import { FluentAgentChat } from "@fluent/sdk";

<FluentAgentChat
  agentId="your-agent-id"
  apiBaseUrl="http://localhost:3000"
  locale="ja"
/>

Or use the headless client:

import { createFluentClient } from "@fluent/sdk";

const fluent = createFluentClient({ apiBaseUrl: "http://localhost:3000" });

const result = await fluent.invokePublishedAgent("your-agent-id", {
  locale: "ja",
  messages: [{ role: "user", content: "How do I reset my password?" }],
});

console.log(result.text);                    // Localized response
console.log(result.interventionSummary);     // What Lingo.dev fixes were applied

Or bring your own LLM — extract Fluent's improvements as a system prompt:

import { createFluentClient, composeSystemPrompt } from "@fluent/sdk";

const fluent = createFluentClient({ apiBaseUrl: "..." });
const agent = await fluent.getPublishedAgent("your-agent-id");
const pack = agent.version.packs.find(p => p.locale === "ja");

// Use with any LLM provider
const systemPrompt = composeSystemPrompt(pack, {
  baseInstructions: agent.agent.baseInstructions,
  locale: "ja",
});

Repo Layout

apps/web/       Next.js dashboard + API routes
packages/sdk/   SDK with typed client, React component, and prompt composition
convex/         Schema, queries, mutations, and backend orchestration

Supported Locales

English (source), Spanish, Japanese, Hindi — with more coming via Lingo.dev.

Verification

bun run typecheck    # Type check
bun run build        # Production build

Built With

• Lingo.dev — Locale-specific fix generation
• Next.js 15 — Dashboard and API
• Convex — Real-time backend
• OpenAI / Ollama — LLM providers

License

MIT