Kin by PersonaliAI

An open-source AI personal assistant — chat, memory, real voice calls, and real actions across your calendar, email, and social accounts.

Self-host it with Docker, or use the hosted version. Same code either way — no vendor lock-in.

Documentation · Quick Start · Features · Architecture · Contributing

---

Why Kin

Most "AI assistant" products are a chat window bolted onto an LLM API — closed source, one model, your data on someone else's server. Kin is the actual thing: a personal assistant that remembers context across conversations, takes real actions (books meetings, sends email, posts to your socials, checks your calendar), and talks to you over a real phone-call-style voice channel — not just text.

	Closed-source assistant SaaS	Kin
Your data	Lives on their servers, always	Your own Supabase project
LLM	Locked to one vendor	Any provider — Gemini, OpenAI, Anthropic, OpenRouter, via LiteLLM, with automatic fallback
Cost visibility	A monthly bill, no breakdown	Per-request token + dollar cost tracking, logged to your own database
Voice	Usually absent or a separate pricier tier	Included — real-time voice agent via LiveKit, same memory as chat
Integrations	Whatever they decided to build	Google/Microsoft calendar & email, 15+ social platforms, MCP servers, custom webhooks
Source code	Closed	MIT licensed — fork it, audit it, self-host it

✨ Features

• 💬 Tool-calling chat agent — multi-round reasoning loop that actually does things, not just talks about them
• 🎙️ Real-time voice agent — phone-call-style conversations via LiveKit, BYOK across OpenAI/Anthropic/Google/Groq/xAI for LLM, plus Deepgram/ElevenLabs/Cartesia/Azure and more for STT/TTS
• 🧠 Persistent memory + RAG — remembers facts across conversations, grounded answers over your own uploaded documents (PDF/DOCX/XLSX/PPTX via markitdown)
• 🔌 Real integrations — Google Calendar/Gmail/Drive/Contacts/Tasks, Microsoft 365 equivalents, 15+ social platforms (LinkedIn, X, Instagram, TikTok, Reddit, Discord, Bluesky, Farcaster, and more), MCP server support
• 💰 Multi-provider LLM via LiteLLM — Gemini by default with automatic model fallback, or bring your own OpenAI/Anthropic/OpenRouter key, with per-request token and dollar cost tracking built in
• 🗓️ Agentic scheduling — books real meetings, sends follow-ups, runs on autopilot within limits you set
• 📊 Dashboard — chat, memory browser, integrations, social scheduling, voice-agent config, billing, usage
• 🐳 One-command self-host — docker compose up, point it at a free Supabase project, done

🏗️ Architecture

flowchart LR
    user(("You")) -->|chat / dashboard| frontend["frontend\n(Next.js)"]
    caller(("Phone-call-style\nvoice conversation")) -->|WebRTC| voice["voice-worker\n(LiveKit Agents)"]

    frontend --> backend["backend\n(FastAPI)"]
    voice --> backend
    voice --> livekit["LiveKit Cloud\n(WebRTC transport)"]

    backend --> db[("Supabase\nPostgres + pgvector + RLS")]
    backend --> llm["LiteLLM\n(Gemini / OpenAI / Anthropic / OpenRouter)"]
    backend --> integrations["Google · Microsoft · Social · MCP"]

    style frontend fill:#6366f1,color:#fff,stroke:none
    style backend fill:#009688,color:#fff,stroke:none
    style voice fill:#ff6600,color:#fff,stroke:none
    style db fill:#3ecf8e,color:#000,stroke:none
    style llm fill:#6366f1,color:#fff,stroke:none

Loading

kin/
├── frontend/       Next.js — dashboard and chat UI
├── backend/        FastAPI — chat/tool-calling agent, memory/RAG, integrations, billing
│   └── app/core/llm.py   unified LiteLLM layer: fallback, streaming, cost tracking
├── voice-worker/   LiveKit Agents — phone/voice-call handling (BYOK per provider)
└── docker-compose.yml

All services share a single Supabase Postgres database — schema + row-level security policies, no separate ORM. Supabase's free tier is enough to get started.

Note on voice-worker: it talks to LLM/STT/TTS providers directly through LiveKit's own plugins with per-user BYOK keys fetched from backend at call time — it does not go through the LiteLLM layer. That's intentional: LiveKit's agent runtime doesn't have a generic LiteLLM plugin, so this is a deliberate architectural split, not an oversight.

🚀 Quick Start (Docker Compose)

1. Create a Supabase project at supabase.com — free tier is fine. Grab these from Project Settings → API and → Database: Project URL, anon key, service_role key, database host/password.

2. Apply the database schema:

cd backend
# apply every file in supabase/migrations/ in order, via the Supabase CLI
# (supabase db push) or the SQL editor in your Supabase dashboard

3. Configure environment variables:

cp backend/.env.example backend/.env              # Supabase + GEMINI_API_KEY at minimum
cp frontend/.env.example frontend/.env.local       # Supabase URL + anon key + backend URL
cp voice-worker/.env.example voice-worker/.env     # only needed for the voice profile

At minimum you need: SUPABASE_URL, SUPABASE_SERVICE_ROLE_KEY (backend), GEMINI_API_KEY (free at aistudio.google.com/apikey), FUNCTION_SECRET and BYOK_ENCRYPTION_KEY (generate your own random values). Everything else is optional — each one unlocks a single feature (voice, a specific social platform, billing, etc.) and degrades gracefully when left blank. Every variable is documented inline in each service's .env.example.

4. Run it:

docker compose up --build backend frontend

Dashboard: http://localhost:3000 · Backend: http://localhost:8080

To also run the voice agent (needs a free LiveKit Cloud project):

docker compose --profile voice up --build

💻 Local Development (without Docker)

Backend:

cd backend
pip install -r requirements.txt
uvicorn main:app --reload --port 8080

Frontend:

cd frontend
npm install
npm run dev

Voice worker (optional):

cd voice-worker
pip install -r requirements.txt
python worker.py dev

⚙️ Configuration Reference

Every environment variable is documented inline in backend/.env.example, frontend/.env.example, and voice-worker/.env.example — what it's for, and what happens if you leave it blank.

🗺️ Roadmap

• Test coverage for frontend/ and voice-worker/ (currently backend-only)
• One-click deploy buttons (Railway / Render / Fly.io)
• Complete the LiteLLM migration — tool-calling loop is done behind a KIN_USE_LITELLM flag pending live-traffic verification, google-genai removal is the last step
• Streaming responses over the LiteLLM path

Have an idea? Open an issue.

🤝 Contributing

Issues and PRs welcome — see CONTRIBUTING.md for dev setup per service. Good first areas: voice-worker test coverage, additional STT/TTS provider plugins, docs.

📄 License

MIT — see LICENSE. Use it, fork it, ship it commercially — attribution appreciated but not required.

---

If Kin is useful to you, star the repo ⭐ — it helps other people find it.