Digital Brain MCP Server

🧠 Digital Brain MCP

A Second Brain powered by Model Context Protocol (MCP), Google Gemini Embedding 2, and Supabase pgvector — deployed on Vercel.

Connect any MCP-compatible AI client (Claude, Cursor, OpenCode, Copilot, etc.) and give it persistent long-term memory. Store text, images, PDFs, audio, and video — all embedded in a unified vector space for cross-modal semantic search.

Architecture

AI Client (Claude / Cursor / OpenCode / Copilot)
        │
        ▼  MCP Protocol (Streamable HTTP + SSE)
        │  Authorization: Bearer <api-key>
┌──────────────────────────────────────────┐
│   Vercel (Next.js)                       │
│   /api/mcp/[transport]                   │
│                                          │
│   ┌── Auth Middleware ──┐                │
│   │  Bearer token check │                │
│   └─────────────────────┘                │
│                                          │
│   Tools:                                 │
│    • store_memory      (text)            │
│    • store_file        (base64 upload)   │
│    • store_file_from_url (URL fetch)     │
│    • search_memory     (cross-modal)     │
│    • get_file_url      (signed download) │
│    • list_memories                       │
│    • update_memory                       │
│    • delete_memory                       │
│    • get_stats                           │
│                                          │
│   REST Endpoint:                         │
│    • POST /api/upload  (direct file)     │
└──────────┬─────────────┬─────────────────┘
           │             │
     ┌─────┴─────┐  ┌───┴──────────────┐
     ▼           ▼  ▼                  ▼
┌─────────┐  ┌──────────────┐  ┌───────────┐
│ Gemini  │  │  Supabase    │  │ Supabase  │
│ Embed 2 │  │  PostgreSQL  │  │ Storage   │
│  API    │  │  + pgvector  │  │ (files)   │
│         │  │  vector(768) │  │           │
└─────────┘  └──────────────┘  └───────────┘

Multimodal Embedding

Gemini Embedding 2 maps all modalities into the same 768-dimension vector space. This means:

• A text query like "architecture diagram" can find a stored PNG image
• Searching for "meeting notes" can return an audio recording of a meeting
• A PDF of a research paper and a text summary live side by side in the same search space

Supported File Types

Interleaved Embedding

When you provide a description alongside a file, the system creates an interleaved embedding — a single vector that captures both the visual/audio content AND your text description. This produces significantly richer search results compared to embedding the file alone.

How It Works

1. You say (in Claude/Cursor/etc): "Remember that the EBR system uses Azure Functions for the API layer"
2. MCP client calls your Digital Brain's store_memory tool
3. Gemini Embedding 2 converts the text into a 768-dimension vector
4. Supabase stores the text + vector in PostgreSQL with pgvector
5. Later, you ask: "What tech does the EBR system use?"
6. search_memory embeds your query, runs cosine similarity search, returns the matching memory

For files, the flow is the same — except the file bytes are sent to Gemini for multimodal embedding, and the raw file is stored in Supabase Storage with a signed download URL generated on retrieval.

Security Model

The server uses Bearer token authentication on every request:

• Fail-closed: If no API keys are configured, ALL requests are rejected
• Multi-key support: Set multiple comma-separated keys in DIGITAL_BRAIN_API_KEYS so each client gets its own key (and you can rotate independently)
• Row Level Security (RLS): Enabled on the Supabase memories table — only service_role can access data. The anon key has zero access.
• Service Role Key: Only stored server-side in Vercel env vars, never exposed to clients
• Private Storage: The brain-files bucket is private — files are only accessible via time-limited signed URLs (1 hour expiry)

Generating API Keys

# Generate a strong 256-bit key
openssl rand -hex 32

Tech Stack