An open-source, self-hosted SDK that adds a voice AI assistant to any documentation site — powered by Agora Conversational AI.
Users click a button, speak naturally, and get spoken answers directly from your documentation. No typing, no searching.
User speaks → Agora ASR → GPT-4o-mini → MiniMax TTS → User hears the answer
↕ MCP
search_docs(query)
→ relevant doc chunks
- Server SDK — a Next.js app you deploy once (Vercel, Railway, anywhere)
- Client SDK —
doctalk.js, a zero-dependency script you drop into any website with 2 lines - Your infrastructure 836C — your Agora account, your docs, your data, your costs
- Agora account — free tier works
- Node.js 18+ and pnpm
git clone https://github.com/anandwana001/doc-talk.git
cd doc-talk
pnpm installcp .env.local.example .env.localFill in your Agora credentials and point to your docs. See Configuration.
pnpm dev
# http://localhost:3000Open http://localhost:3000/embed to test the voice UI in your browser.
<script src="https://YOUR-DOCTALK-URL/doctalk.js"></script>
<script>
DocTalk.init({ apiBase: 'https://YOUR-DOCTALK-URL' });
</script>See Integration guide for all options.
| Guide | What's inside |
|---|---|
| How it works | Architecture, Agora voice pipeline, MiniSearch RAG, transcript links |
| Configuration | All environment variables, docs source options |
| Integration | Adding to your website, widget options, iframe details |
| Deployment | Vercel deploy, production checklist, API auth |
| Roadmap | Known limitations and planned improvements |
| Layer | Technology |
|---|---|
| Voice pipeline | Agora Conversational AI Engine |
| Real-time transport | Agora SD-RTN |
| Transcripts | Agora RTM |
| Framework | Next.js 16 (App Router) |
| Doc retrieval | MCP + MiniSearch (BM25, pre-built index, per-turn) |
| Client widget | Vanilla JS — zero dependencies |
| Styling | Tailwind CSS |
MIT — free to use, modify, and self-host commercially.