A keyboard-first email client with an AI copilot that drafts replies and answers questions about your inbox, grounded in your actual prior threads using RAG. Built as a production SaaS with subscription billing, multi-account support, and real-time sync.
Live Demo → (add deployed URL)
Most email clients are passive — they store and display messages but leave all the thinking to you. Inbox Copilot makes your email history queryable and actionable:
- Ask "What did I promise this client last week?" and get a sourced answer from your actual threads
- Draft replies with full context from prior conversations, not hallucinated content
- Navigate and manage your inbox entirely from the keyboard
- AI reply drafting — context-aware drafts grounded in relevant prior threads via RAG
- Inbox Q&A — natural language questions answered from your real email history
- Multi-account support — connect Gmail, Outlook, and other providers through a unified API
- Real-time sync — webhooks + incremental sync tokens keep state current without redundant reprocessing
- Semantic search — find emails by meaning, not just keywords; powered by vector embeddings
- Keyboard-first UX — command palette, shortcuts for composing/replying/navigating
- Subscription SaaS — free vs. pro tiers with feature gating, Stripe billing, and webhook-driven subscription management
Email Providers (Gmail / Outlook)
↓
Aurinko API (OAuth, webhooks, incremental sync)
↓
PostgreSQL via Prisma ──── threads, metadata, user state
↓
Embedding Pipeline (OpenAI)
↓
Pinecone ─────────────────── semantic vector index
↓
RAG Query (LangChain + OpenAI)
↓
Copilot Response (draft reply / answer question)
The core constraint: AI responses are always grounded in emails that exist in your inbox. No hallucinated context.
| Layer | Technology |
|---|---|
| Framework | Next.js 14 (App Router) |
| Language | TypeScript |
| AI / LLM | OpenAI API |
| RAG | LangChain + Pinecone |
| Email Sync | Aurinko API |
| Auth | Clerk |
| Payments | Stripe + webhooks |
| Database | PostgreSQL via Prisma ORM |
| File Storage | AWS S3 |
| Styling | Tailwind CSS |
├── app/ # Next.js App Router — pages and API routes
├── components/ # UI components (email list, compose, copilot panel)
├── lib/ # RAG pipeline, Aurinko sync, embedding logic
├── prisma/ # Schema and migrations
└── public/ # Static assets
Prerequisites: Node.js 18+, PostgreSQL, Pinecone account, OpenAI key, Aurinko account, Clerk app, Stripe account, AWS S3 bucket
git clone https://github.com/VinayakMaharaj/Inbox-Copilot.git
cd Inbox-Copilot
npm installCreate a .env.local file:
# Database
DATABASE_URL=
# Clerk
NEXT_PUBLIC_CLERK_PUBLISHABLE_KEY=
CLERK_SECRET_KEY=
# OpenAI
OPENAI_API_KEY=
# Pinecone
PINECONE_API_KEY=
PINECONE_INDEX_NAME=
# Aurinko (email sync)
AURINKO_CLIENT_ID=
AURINKO_CLIENT_SECRET=
AURINKO_SIGNING_SECRET=
# Stripe
STRIPE_SECRET_KEY=
STRIPE_WEBHOOK_SECRET=
NEXT_PUBLIC_STRIPE_PUBLISHABLE_KEY=
# AWS S3
AWS_ACCESS_KEY_ID=
AWS_SECRET_ACCESS_KEY=
AWS_REGION=
AWS_BUCKET_NAME=npx prisma migrate dev
npm run devOpen http://localhost:3000.
Why RAG over fine-tuning? Fine-tuning would require retraining as the inbox grows and can't reference specific threads at query time. RAG retrieves semantically relevant emails at inference time — keeping responses current and grounded without retraining.
Why Aurinko over direct Gmail/Outlook APIs? Aurinko abstracts OAuth flows, webhook management, and incremental sync across providers. Direct integrations with each provider would have tripled the setup work for the same result.
Why Pinecone over pgvector? For a project already using PostgreSQL for relational data, pgvector would have been simpler. Pinecone's managed infrastructure and filtering capabilities made it the cleaner choice for the vector layer specifically — worth the added dependency.
Webhook-first sync over polling Polling on a fixed interval either over-fetches (wasting API quota) or under-fetches (stale state). Webhooks + incremental delta tokens give low-latency updates with no redundant processing.
- AI response quality depends on embedding quality and retrieval accuracy
- Large inboxes increase indexing and storage costs
- RAG favors correctness over creativity — drafts may be less stylistically varied than a purely generative approach
These tradeoffs were intentional to prioritize reliability and user trust over stylistic flair.
- Designing RAG pipelines over user-generated data with real retrieval constraints
- Handling real-time sync with external APIs using webhooks and delta tokens
- Balancing AI capability against cost and latency at the SaaS layer
- Building and monetizing a full production application end-to-end as a solo developer