Your life, unified and connected.

A mobile AI assistant powered by Google Gemini 3 Flash that connects to your entire digital life through the Model Context Protocol (MCP). One app, one conversation — email, calendar, flights, hotels, web search, memory, and more.

2nd Brain is built entirely around Gemini 3 Flash Preview as its reasoning core. Three features make the app possible:

Thinking Mode — Always enabled (4,000-token budget). Gemini reasons through multi-step problems before acting — decomposing "plan my Tokyo trip" into flight search, hotel search, calendar check, and event creation before executing any tool.

Native Function Calling — 21+ tools from Google Workspace and SerpAPI exposed as Gemini function declarations. The model picks the right tools, calls them with typed arguments, and chains up to 8 iterations per turn across multiple MCP servers — no hardcoded orchestration.

Google Search Grounding — Activates for real-time queries (weather, news, prices). Gemini grounds responses in live Google Search results with citations and source URLs.

Together these features turn Gemini 3 Flash from a text generator into an autonomous agent that thinks, acts, searches, and remembers.

graph TB
    subgraph CLIENT["Mobile App (React Native / Expo)"]
        UI["Voice-first Chat UI"]
        BRAIN["Agent Brain<br/>Skill Detection · Agentic Loop · Memory"]
    end

    subgraph GOOGLE_AI["Google AI"]
        GEMINI["Gemini 3 Flash Preview<br/>Thinking Mode · Function Calling"]
        GSEARCH["Google Search Grounding"]
    end

    subgraph MCP["MCP Servers (Google Cloud Run)"]
        GMCP["Google Workspace MCP<br/>Gmail · Calendar · Drive · Contacts"]
        SMCP["SerpAPI MCP<br/>Web · Flights · Hotels · Local · News"]
        KMCP["Kite MCP<br/>Stocks · Portfolio · Orders"]
    end

    subgraph MEMORY["Memory & Data (Google Cloud)"]
        MEM0["Mem0 Connector<br/>(Cloud Run)"]
        QDRANT["Qdrant — Vector DB<br/>(Compute Engine)"]
        NEO4J["Neo4j — Graph DB<br/>(Compute Engine)"]
        PREFS["Preferences API<br/>(Cloud Run)"]
        MONGO["MongoDB<br/>(Google Cloud)"]
    end

    subgraph AUTH["Auth"]
        CLERK["Clerk + Google Sign-In"]
    end

    UI --> BRAIN
    BRAIN --> GEMINI
    GEMINI --> GSEARCH
    GEMINI -->|"Function Calls"| GMCP
    GEMINI -->|"Function Calls"| SMCP
    GEMINI -->|"Function Calls"| KMCP
    BRAIN -->|"JWT"| MEM0
    BRAIN -->|"JWT"| PREFS
    MEM0 --> QDRANT
    MEM0 --> NEO4J
    PREFS --> MONGO
    UI --> CLERK

    style GEMINI fill:#4285F4,color:#fff
    style GSEARCH fill:#34A853,color:#fff
    style GMCP fill:#4285F4,color:#fff
    style SMCP fill:#EA4335,color:#fff
    style KMCP fill:#FBBC04,color:#000
    style QDRANT fill:#EA4335,color:#fff
    style NEO4J fill:#34A853,color:#fff

All services run on Google Cloud Platform:

graph LR
    subgraph CLOUD_RUN["Google Cloud Run (auto-scaling)"]
        CR1["Google Workspace MCP<br/>europe-west1"]
        CR2["SerpAPI MCP<br/>europe-west1"]
        CR3["Mem0 Connector<br/>us-east1"]
        CR4["Preferences API<br/>europe-west1"]
    end

    subgraph COMPUTE["Google Compute Engine"]
        VM["Mem0 + Qdrant + Neo4j"]
    end

    subgraph DATA["Google Cloud"]
        DB["MongoDB"]
    end

    subgraph EXTERNAL["External APIs"]
        GAPI["Google Workspace APIs"]
        SERP["SerpAPI"]
        KITE["Kite (Zerodha)"]
    end

    CR1 -->|"OAuth"| GAPI
    CR2 -->|"API Key"| SERP
    CR3 --> VM
    CR4 --> DB

    style CR1 fill:#4285F4,color:#fff
    style CR2 fill:#EA4335,color:#fff
    style CR3 fill:#FBBC04,color:#000
    style CR4 fill:#34A853,color:#fff

2nd Brain isn't a stateless chatbot — it has a layered cognitive architecture:

• Identity (types.ts) — TypeScript interfaces defining the agent's shape: AgentSoul, AgentSkill, AgentContext
• Soul (soul.ts) — Persistent personality with 5 traits, 5 values, 11 skills, 7 quirks. Evolves with user preferences via evolveSoul()
• Brain (brain.ts) — Orchestration engine: skill detection, parallel memory retrieval, system prompt assembly, agentic loop (up to 8 Gemini calls per turn), post-processing
• Memory (memory.ts) — Long-term recall via Mem0 + Qdrant (vector) + Neo4j (graph). Four memory types: soul, conversation, preference, identity

The Google Workspace MCP is the primary integration, built with FastMCP v3 and Python 3.11+. It acts as an OAuth 2.0 proxy giving the agent full access to Google services:

• Gmail: list, read, send, unread count
• Google Drive: list, search, create files
• Google Calendar: list events, get today's schedule, create events
• Google Contacts: list, search

Security:

• PKCE (Proof Key for Code Exchange) — prevents authorization code interception
• Token Encryption — Fernet (AES-128-CBC + HMAC-SHA256) for stored tokens
• JWT Signing — HS256 with secure key derivation
• Consent Pages — protection against confused deputy attacks

┌──────────────────┐     ┌──────────────────┐     ┌──────────────┐
│  2nd Brain App   │────▶│  FastMCP v3       │────▶│  Google APIs  │
│  (MCP Client)    │     │  OAuth Proxy      │     │  Gmail/Cal/   │
│                  │◀────│  PKCE + JWT + AES │◀────│  Drive/People │
└──────────────────┘     └──────────────────┘     └──────────────┘

Works with any MCP client (Claude Desktop, Cursor, etc.):

{
  "mcpServers": {
    "google-workspace": {
      "url": "http://localhost:3000/mcp"
    }
  }
}

A self-hosted MCP wrapping SerpAPI's full suite of Google search engines:

• search — Web search (compatible with SerpApi MCP spec)
• search_flights / get_cheapest_flights — Google Flights
• search_hotels / get_hotel_details / find_budget_hotels — Google Hotels
• get_cache_stats / clear_cache — Cache management

Also covers local results, social profiles, news, shopping, and images through the search tool. 5-minute in-memory cache reduces API costs by 50-70%.

• Voice-first chat UI (tap to record, toggle to keyboard)
• MCP Marketplace — discover and connect servers
• Profile — preferences that shape the agent's soul
• Onboarding — two-step wizard (profile + goals)
• Clerk + Google Sign-In authentication
• Cross-platform: iOS, Android, Web

npm install
cp .env.example .env
# Edit .env with your keys (see below)
npx expo start

cd google-mcp-python
uv sync
cp .env.example .env
# Set GOOGLE_CLIENT_ID, GOOGLE_CLIENT_SECRET
uv run server.py
# Runs at http://localhost:3000/mcp

See google-mcp-python/README.md for OAuth setup.

cd travel-mcp-python
uv sync
cp .env.example .env
# Set SERPAPI_KEY (and optionally API_KEY for auth)
uv run server.py
# Runs at http://localhost:3001/mcp

See travel-mcp-python/README.md for tool details.

# Google Workspace MCP
cd google-mcp-python && docker-compose up

# SerpAPI MCP
cd travel-mcp-python && docker-compose up

Both MCP servers include Dockerfiles optimized for Cloud Run:

# SerpAPI MCP
cd travel-mcp-python
gcloud run deploy travel-mcp-server --source . --region europe-west1

# Google Workspace MCP
cd google-mcp-python
gcloud run deploy google-mcp --source . --region europe-west1

curl http://localhost:3000/health   # Google Workspace MCP
curl http://localhost:3001/health   # SerpAPI MCP

# Auth
EXPO_PUBLIC_CLERK_PUBLISHABLE_KEY=

# AI
EXPO_PUBLIC_GEMINIKEY=

# MCP Servers
EXPO_PUBLIC_MCP_SERVER_URL=         # Google Workspace MCP
EXPO_PUBLIC_MCP_TRAVEL_URL=         # SerpAPI MCP
EXPO_PUBLIC_MCP_KITE_URL=           # Kite MCP (optional)
EXPO_PUBLIC_MCP_API_KEY=            # API key for SerpAPI MCP (optional)

# Backend
EXPO_PUBLIC_MEMORY_API_URL=         # Mem0 connector
EXPO_PUBLIC_PREFERENCES_API_URL=    # Preferences API

Built at the Gemini 3 Hackathon with:

• Google Gemini — AI model + Search grounding
• Google Cloud Run — MCP server hosting
• Google Workspace APIs — Gmail, Calendar, Drive, Contacts
• Model Context Protocol — Open standard for AI tool use
• FastMCP — Python MCP framework
• SerpAPI — Universal Google search API
• Mem0 — Long-term memory framework
• Clerk — Authentication
• Expo — React Native framework

MIT License